Contact fatigue or dynamic selection? Decomposing the declining conversion hazard across repeated telemarketing contacts and designing client-specific stopping rules
Paper
Contact Fatigue or Dynamic Selection? Decomposing the Declining Conversion Hazard in Repeated Telemarketing and Designing Client-Specific Stopping Rules
Abstract
Managers read the falling per-contact conversion rate in repeated telemarketing as contact fatigue; it may instead reflect dynamic selection, as receptive clients convert early, or the firm's choice of whom to keep calling. We model contacts to subscription as a grouped-time proportional hazard with campaign-month effects, client covariates and gamma or split-population heterogeneity, decompose the decline exactly into four parts, and characterize the profit-maximizing stopping rule. In a Portuguese bank's 2008–2010 campaigns (41,188 clients, 105,754 contacts), the hazard falls from 0.0558 at the first contact to 0.0110 at the tenth. With month effects, composition explains 77.7% (95% interval 64.0–99.2%) of the decline to the fourth contact; a single 2008/2009–2010 effect, which fits far worse (log-likelihood −15,943.12 against −15,236.87) and omits within-regime month variation, gives a conservative bound of 51.6% (41.8–67.4%). If continuation decisions ignore unobserved receptiveness, the second and third contacts are not detectably less productive than the first. Later contacts show a structural decline in every specification except a 2008 gamma model, whose implied equal attenuation of month and client contrasts is rejected (p = 0.009), though not its overall pattern (p = 0.194); the client contrasts are too imprecise to show or exclude attenuation. At a cost-to-margin ratio of 2%, rules evaluated exactly on logged calls raise profit over the bank's practice by 7.8% (6.3–9.6%) with client covariates only, feasible in real time, and by up to 10.4% (8.6–12.6%) adding same-month outcomes. Rules built under different heterogeneity assumptions differ by at most 0.22 conversion-equivalents per 1,000 clients.
§Ask about this paper
Grounded in this runQuestions are answered from this run's record — the paper, the executed analysis, the data provenance and the referee report. Nothing else is consulted.
This run is finished, so the recorded paper cannot be rewritten. Questions are still answered from its record.
Pipeline
1.Literature review
Literature review complete (live:agent:claude-opus-5-5): 10 key papers, 69 grounded sources, 5 open gaps.
Summary
An aggregate per-contact conversion rate that falls with the number of contacts admits three readings, and the literature has so far separated only two of them. The first is true duration dependence: each client's receptiveness changes with repeated contact, through wear-in at first and wear-out or fatigue later, as the advertising-repetition literature documents for exposures and as recent work documents for email solicitations to panels and for repeated randomized marketing interventions. The second is dynamic selection: receptive clients convert early and leave, so the pool still being called is increasingly unreceptive even if no individual changes. The econometric literature on true versus spurious duration dependence (Lancaster; Heckman and Borjas; Elbers and Ridder; Heckman and Singer; Vaupel and Yashin) shows that the two are separable only through structure, typically the mixed proportional hazard (MPH) assumption that covariates shift the hazard proportionally; that the heterogeneity needed to do so is weakly identified in single-spell data (Ridder and Woutersen; Baker and Melino); and that estimated duration dependence is sensitive to the heterogeneity distribution assumed. Grouped-time proportional hazards with gamma heterogeneity (Prentice and Gloeckler; Meyer; Han and Hausman) give a closed-form likelihood suited to contact counts, and split-population models capture clients who would never convert. Marketing has the same sorting logic in retention (Fader and Hardie) and in inferring customer activity from silence (Schmittlein, Morrison and Colombo). The third reading, largely absent from the dichotomy, is policy composition: the firm decides whom to call again, so the late-contact pool reflects the firm's own continuation rule and, in campaign data, the calendar months in which it kept calling. On the decision side, contact policy is a classic Management Science problem, solved as a dynamic program for catalog mailing (Bitran and Mondschein; Gonul and Shi; Simester, Sun and Tsitsiklis), for online panel invitations (Neslin et al.) and for sales-call allocation (Lodish); optimal stopping theory gives the one-step look-ahead rule in monotone problems, and recent work learns stopping policies from data, including within a sales call. Targeting research stresses profit-based evaluation on held-out data. In telemarketing specifically, scholarly evidence is thin: the Portuguese bank data used here have been analysed almost only as a classification benchmark (Moro, Cortez and Rita), and no study has decomposed the contact-number gradient, bounded the role of unobserved heterogeneity with explicit uncertainty, or asked when the decomposition changes the stopping rule a manager should use.
Key papers (10)
- A Method for Minimizing the Impact of Distributional Assumptions in Econometric Models for Duration Data. Heckman JJ, Singer B; Econometrica 52(2):271-320, 1984.Shows that estimated structural duration dependence depends on the assumed heterogeneity distribution and proposes the discrete mass-point NPMLE; the canonical statement of the fatigue-versus-selection identification problem.
- True and Spurious Duration Dependence: The Identifiability of the Proportional Hazard Model. Elbers C, Ridder G; Review of Economic Studies 49(3):403-409, 1982.Proves that the mixed proportional hazard model is identified from covariate variation, the structural assumption under which fatigue and selection can be separated in single-spell contact data.
- Heterogeneity's Ruses: Some Surprising Effects of Selection on Population Dynamics. Vaupel JW, Yashin AI; The American Statistician 39(3):176-185, 1985.Shows how selection makes population hazards fall even when individual hazards are constant or rising; the mechanism behind dynamic selection across contacts.
- Unemployment Insurance and Unemployment Spells. Meyer BD; Econometrica 58(4):757-782, 1990.Estimates a grouped-time proportional hazard model with a nonparametric baseline and gamma heterogeneity in closed form, the estimator appropriate for discrete contact counts.
- The Singularity of the Information Matrix of the Mixed Proportional Hazard Model. Ridder G, Woutersen TM; Econometrica 71(5):1579-1589, 2003.Explains why unobserved heterogeneity is weakly identified in single-spell data, motivating profile-likelihood inference and explicit identified ranges rather than a point estimate of the fatigue share.
- How to Project Customer Retention. Fader PS, Hardie BGS; Journal of Interactive Marketing 21(1):76-90, 2007.Shows that rising retention with tenure can be pure sorting under heterogeneity (shifted-beta-geometric), the marketing mirror image of falling conversion with contact number.
- An Optimal Contact Model for Maximizing Online Panel Response Rates. Neslin SA, Novak TP, Baker KR, Hoffman DL; Management Science 55(5):727-737, 2009.A Management Science contact-policy optimisation in which recent contacts depress response; the closest decision model to a client-specific stopping rule.
- Dynamic Catalog Mailing Policies. Simester DI, Sun P, Tsitsiklis JN; Management Science 52(5):683-696, 2006.Learns dynamic contact policies from historical data and benchmarks them against the firm's practice, the template for evaluating stopping rules on logged campaign data.
- Silently Killing Your Panelists One Email at a Time: The True Cost of Email Solicitations. Ferecatu A, De Bruyn A, Mukherjee P; Journal of the Academy of Marketing Science 52(4):1216-1239, 2024.Recent evidence that repeated solicitations cause wear-out and attrition, with an optimised solicitation schedule; direct evidence that contact fatigue exists in a related channel.
- A Data-Driven Approach to Predict the Success of Bank Telemarketing. Moro S, Cortez P, Rita P; Decision Support Systems 62:22-31, 2014.Source paper for the Portuguese retail bank telemarketing data; documents the campaigns (outbound list calls plus inbound offers), the period and the predictive benchmark.
Methods in use
- Discrete-time (grouped) hazard models estimated on person-period data with logit or complementary log-log links; the cloglog link is the grouped-time proportional hazards model (Prentice and Gloeckler; Allison; Jenkins).
- Mixed proportional hazard models with gamma heterogeneity in closed form (Meyer; Han and Hausman), discrete mass-point NPMLE heterogeneity (Heckman and Singer; Laird), and split-population (never-convert) models (Schmidt and Witte; Sinha and Chandrashekaran).
- Likelihood-ratio tests of no heterogeneity on the parameter boundary with chi-bar-squared limits (Self and Liang), and profile-likelihood inference with nuisance parameters profiled out (Murphy and van der Vaart).
- Index-number (within versus composition) decompositions of aggregate change; simulation from the estimated sampling distribution (Krinsky and Robb) or bootstrap for intervals on nonlinear functions such as shares.
- Dynamic programming and optimal stopping for contact policies (catalog mailing, panel invitations, sales-call allocation); the one-step look-ahead rule in monotone stopping problems (Chow and Robbins).
- Profit-based evaluation of targeting and contact policies against the firm's practice and uniform rules on held-out data with cross-fitting; reporting of percentage profit improvement, contacts saved and sales retained, with sensitivity to the cost ratio.
Open gaps
- No study decomposes the decline in per-contact conversion across repeated telemarketing contacts into structural contact effects (wear-in and fatigue), dynamic selection on observed and unobserved client characteristics, and the composition created by the firm's own continuation policy and calendar.
- Unobserved heterogeneity is weakly identified in single-spell data, and estimates that rest on a few mass points are fragile; a profile-likelihood interval for the heterogeneity variance, compared across heterogeneity families, would state how much of the within-client decline selection can explain.
- Whether second and third telephone contacts are more productive than the first for the same client (wear-in) is undocumented; because selection alone can never raise a conditional hazard, a rising within-cell hazard is direct evidence of a structural contact effect.
- It is not formalised when the fatigue/selection split matters for a within-campaign stopping rule; if the rule depends only on the identified conditional hazard path, the split matters only for counterfactuals outside the observed policy.
- Stopping rules are rarely evaluated on logged campaign data without a response model; rules that stop no later than the firm did can be evaluated exactly by truncating the observed contact sequences, yielding profit comparisons against the firm's practice and a uniform cap with intervals.
Grounded search sources (69) · mode: live:agent:claude-opus-5-5
- A Method for Minimizing the Impact of Distributional Assumptions in Econometric Models for Duration Data (Heckman and Singer 1984, Econometrica 52(2):271-320)
- True and Spurious Duration Dependence: The Identifiability of the Proportional Hazard Model (Elbers and Ridder 1982, Review of Economic Studies 49(3):403-409)
- Econometric Methods for the Duration of Unemployment (Lancaster 1979, Econometrica 47(4):939-956)
- Does Unemployment Cause Future Unemployment? Definitions, Questions and Answers from a Continuous Time Model of Heterogeneity and State Dependence (Heckman and Borjas 1980, Economica 47(187):247-283)
- The Impact of Heterogeneity in Individual Frailty on the Dynamics of Mortality (Vaupel, Manton and Stallard 1979, Demography 16(3):439-454)
- Heterogeneity's Ruses: Some Surprising Effects of Selection on Population Dynamics (Vaupel and Yashin 1985, The American Statistician 39(3):176-185)
- Duration Models: Specification, Identification and Multiple Durations (Van den Berg 2001, Handbook of Econometrics vol. 5, 3381-3460)
- Unemployment Insurance and Unemployment Spells (Meyer 1990, Econometrica 58(4):757-782)
- Flexible Parametric Estimation of Duration and Competing Risk Models (Han and Hausman 1990, Journal of Applied Econometrics 5(1):1-28)
- Regression Analysis of Grouped Survival Data with Application to Breast Cancer Data (Prentice and Gloeckler 1978, Biometrics 34(1):57-67)
- Discrete-Time Methods for the Analysis of Event Histories (Allison 1982, Sociological Methodology 13:61-98)
- Easy Estimation Methods for Discrete-Time Duration Models (Jenkins 1995, Oxford Bulletin of Economics and Statistics 57(1):129-136)
- Predicting Criminal Recidivism Using 'Split Population' Survival Time Models (Schmidt and Witte 1989, Journal of Econometrics 40(1):141-159)
- Asymptotic Properties of Maximum Likelihood Estimators and Likelihood Ratio Tests under Nonstandard Conditions (Self and Liang 1987, JASA 82(398):605-610)
- On Approximating the Statistical Properties of Elasticities (Krinsky and Robb 1986, Review of Economics and Statistics 68(4):715-719)
- The Singularity of the Information Matrix of the Mixed Proportional Hazard Model (Ridder and Woutersen 2003, Econometrica 71(5):1579-1589)
- Duration Dependence and Nonparametric Heterogeneity: A Monte Carlo Study (Baker and Melino 2000, Journal of Econometrics 96(2):357-393)
- Estimating a Semi-Parametric Duration Model without Specifying Heterogeneity (Hausman and Woutersen 2014, Journal of Econometrics 178:114-131)
- On Profile Likelihood (Murphy and van der Vaart 2000, JASA 95(450):449-465)
- Nonparametric Maximum Likelihood Estimation of a Mixing Distribution (Laird 1978, JASA 73(364):805-811)
- Jobs, Strikes, and Wars: Probability Models for Duration (Morrison and Schmittlein 1980, Organizational Behavior and Human Performance 25(2):224-251)
- Analyzing Duration Times in Marketing: Evidence for the Effectiveness of Hazard Rate Models (Helsen and Schmittlein 1993, Marketing Science 12(4):395-414)
- A Split Hazard Model for Analyzing the Diffusion of Innovations (Sinha and Chandrashekaran 1992, Journal of Marketing Research 29(1):116-127)
- How to Project Customer Retention (Fader and Hardie 2007, Journal of Interactive Marketing 21(1):76-90)
- Counting Your Customers: Who Are They and What Will They Do Next? (Schmittlein, Morrison and Colombo 1987, Management Science 33(1):1-24)
- State Dependence and Alternative Explanations for Consumer Inertia (Dube, Hitsch and Rossi 2010, RAND Journal of Economics 41(3):417-445)
- Investigating Household Purchase Timing Decisions: A Conditional Hazard Function Approach (Jain and Vilcassim 1991, Marketing Science 10(1):1-23)
- The Proportional Hazard Model for Purchase Timing: A Comparison of Alternative Specifications (Seetharaman and Chintagunta 2003, Journal of Business & Economic Statistics 21(3):368-382)
- Television Commercial Wearout: An Information Processing View (Calder and Sternthal 1980, Journal of Marketing Research 17(2):173-186)
- Advertising Repetition: A Critical Review of Wearin and Wearout (Pechmann and Stewart 1988, Current Issues and Research in Advertising 11(1-2):285-329)
- Planning Media Schedules in the Presence of Dynamic Advertising Quality (Naik, Mantrala and Sawyer 1998, Marketing Science 17(3):214-235)
- Online Display Advertising: Modeling the Effects of Multiple Creatives and Individual Impression Histories (Braun and Moe 2013, Marketing Science 32(5):753-767)
- Wearout or Weariness? Measuring Potential Negative Consequences of Online Ad Volume and Placement on Website Visits (Chae, Bruno and Feinberg 2019, Journal of Marketing Research 56(1):57-75)
- Ad Wearout Wearout: How Time Can Reverse the Negative Effect of Frequent Advertising Repetition on Brand Preference (Kronrod and Huber 2019, International Journal of Research in Marketing 36(2):306-324)
- Dynamic and Competitive Effects of Direct Mailings: A Charitable Giving Application (Van Diepen, Donkers and Franses 2009, Journal of Marketing Research 46(1):120-133)
- Does Irritation Induced by Charitable Direct Mailings Reduce Donations? (Van Diepen, Donkers and Franses 2009, International Journal of Research in Marketing 26(3):180-188)
- Modeling Customer Opt-In and Opt-Out in a Permission-Based Marketing Context (Kumar, Zhang and Luo 2014, Journal of Marketing Research 51(4):403-419)
- Silently Killing Your Panelists One Email at a Time: The True Cost of Email Solicitations (Ferecatu, De Bruyn and Mukherjee 2024, Journal of the Academy of Marketing Science 52(4):1216-1239)
- An Optimal Contact Model for Maximizing Online Panel Response Rates (Neslin, Novak, Baker and Hoffman 2009, Management Science 55(5):727-737)
- Learning from Many Experiments: A Hierarchical Bayesian Framework for Decomposing Marketing Treatment Heterogeneity (Ebbes, Ascarza and Netzer 2026, SSRN working paper)
- Privacy and Marketing Externalities: Evidence from Do Not Call (Goh, Hui and Png 2015, Management Science 61(12):2982-3000)
- Measuring Multichannel Advertising Response (Zantedeschi, Feit and Bradlow 2017, Management Science 63(8):2706-2728)
- Are Physicians 'Easy Marks'? Quantifying the Effects of Detailing and Sampling on New Prescriptions (Mizik and Jacobson 2004, Management Science 50(12):1704-1715)
- CALLPLAN: An Interactive Salesman's Call Planning System (Lodish 1971, Management Science 18(4-part-II):P25-P40)
- Mailing Decisions in the Catalog Sales Industry (Bitran and Mondschein 1996, Management Science 42(9):1364-1381)
- Optimal Mailing of Catalogs: A New Methodology Using Estimable Structural Dynamic Programming Models (Gonul and Shi 1998, Management Science 44(9):1249-1262)
- Dynamic Catalog Mailing Policies (Simester, Sun and Tsitsiklis 2006, Management Science 52(5):683-696)
- Optimizing Rhenania's Direct Marketing Business Through Dynamic Multilevel Modeling (DMLM) in a Multicatalog-Brand Environment (Elsner, Krafft and Huchzermeier 2004, Marketing Science 23(2):192-206)
- Optimizing the Marketing Interventions Mix in Intermediate-Term CRM (Rust and Verhoef 2005, Marketing Science 24(3):477-489)
- Valuing Customer Portfolios with Endogenous Mass and Direct Marketing Interventions Using a Stochastic Dynamic Programming Decomposition (Esteban-Bravo, Vidal-Sanz and Yildirim 2014, Marketing Science 33(5):621-640)
- Optimal Selection for Direct Mail (Bult and Wansbeek 1995, Marketing Science 14(4):378-394)
- Targeting Prospective Customers: Robustness of Machine-Learning Methods to Typical Data Challenges (Simester, Timoshenko and Zoumpoulis 2020, Management Science 66(6):2495-2522)
- Efficiently Evaluating Targeting Policies: Improving on Champion vs. Challenger Experiments (Simester, Timoshenko and Zoumpoulis 2020, Management Science 66(8):3412-3424)
- Heterogeneous Treatment Effects and Optimal Targeting Policy Evaluation (Hitsch, Misra and Zhang 2024, Quantitative Marketing and Economics 22(2):115-168)
- Policy Learning with Observational Data (Athey and Wager 2021, Econometrica 89(1):133-161)
- Retention Futility: Targeting High-Risk Customers Might Be Ineffective (Ascarza 2018, Journal of Marketing Research 55(1):80-98)
- Double/Debiased Machine Learning for Treatment and Structural Parameters (Chernozhukov, Chetverikov, Demirer, Duflo, Hansen, Newey and Robins 2018, Econometrics Journal 21(1):C1-C68)
- Interpretable Optimal Stopping (Ciocan and Misic 2022, Management Science 68(3):1616-1638)
- Learning When to Quit in Sales Conversations (Manzoor, Ascarza and Netzer 2025, arXiv:2511.01181)
- A Martingale System Theorem and Applications (Chow and Robbins; 1961 Berkeley Symposium paper reprinted in Herbert Robbins Selected Papers, Springer 1985, 413-424)
- Learning About the Arrival of Sales (Mason and Valimaki 2011, Journal of Economic Theory 146(4):1699-1711)
- Telephone Call Centers: Tutorial, Review, and Research Prospects (Gans, Koole and Mandelbaum 2003, Manufacturing & Service Operations Management 5(2):79-141)
- Predictive Dialing for Outbound Telephone Call Centers (Samuelson 1999, Interfaces 29(5):66-81)
- Subsampling Callbacks to Improve Survey Efficiency (Elliott, Little and Lewitzky 2000, JASA 95(451):730-738)
- Responsive Design for Household Surveys: Tools for Actively Controlling Survey Errors and Costs (Groves and Heeringa 2006, JRSS Series A 169(3):439-457)
- A Data-Driven Approach to Predict the Success of Bank Telemarketing (Moro, Cortez and Rita 2014, Decision Support Systems 62:22-31)
- Bank Marketing [Dataset] (Moro, Rita and Cortez 2014, UCI Machine Learning Repository)
- Using Customer Lifetime Value and Neural Networks to Improve the Prediction of Bank Deposit Subscription in Telemarketing Campaigns (Moro, Cortez and Rita 2015, Neural Computing and Applications 26(1):131-139)
- A Divide-and-Conquer Strategy Using Feature Relevance and Expert Knowledge for Enhancing a Data Mining Approach to Bank Telemarketing (Moro, Cortez and Rita 2018, Expert Systems 35(3):e12253)
2.Ideation
3 candidate ideas generated; selected: “Contact fatigue, dynamic selection or the firm's own continuation policy? An exact decomposition of the declining conversion hazard across repeated telemarketing contacts, and client-specific stopping rules evaluated on the logged campaigns”.
- Idea 1Selected
Contact fatigue, dynamic selection or the firm's own continuation policy? An exact decomposition of the declining conversion hazard across repeated telemarketing contacts, and client-specific stopping rules evaluated on the logged campaigns
Hypothesis. Primary hypothesis H1 (composition, not fatigue): most of the fall in the aggregate per-contact conversion hazard between the first and the fourth contact is composition - dynamic selection on observed client and calendar characteristics plus the bank's own choice of whom to keep calling - rather than a structural contact effect on the individual client. Operationally, at contact 4 (the first contact the bank makes to only a minority of clients) the share of the log decline log(h4/h1) that is not attributable to the within-client-and-month contact effect exceeds one half. H1 is supported if the lower end of the 95% interval for that composition share exceeds 0.5, refuted if the upper end is below 0.5, and inconclusive otherwise; it is evaluated in the model without unobserved heterogeneity, which by construction assigns the largest possible share to fatigue, so the test is conservative. Secondary hypotheses: H2 (wear-in) the within-cell hazard does not fall between contacts 1 and 3, so the second and third contacts are at least as productive as the first for a given client (refuted if the within-cell contact effects at k=2 or k=3 are significantly negative); H3 (bounded unobserved selection) the profile-likelihood 95% interval for the gamma heterogeneity variance excludes the values needed for unobserved selection to account for the within-cell decline by contact 7; H4 (decision relevance) stopping rules derived from the conditional hazard raise campaign profit over the bank's practice and over the best uniform contact cap, and the rule computed assuming no unobserved heterogeneity earns the same profit, within sampling error, as the rule computed at the upper confidence limit of the heterogeneity variance.
Dataset. Portuguese bank direct-marketing telephone campaigns bank_marketing_calls
Test plan, novelty and risk
Test plan. Data: all 41,188 client-campaign records of bank-additional-full (May 2008 to November 2010), no rows dropped. Duration variable T = campaign (number of contacts in the campaign, including the last); event d = 1 if y = 'yes', taken to occur at the last recorded contact; clients with y = 'no' are right-censored at T by the bank's decision to stop. Contacts 1-10 get their own baseline parameter and contacts 11+ share one. Calendar period = the 26 (month, consumer price index) cells, which the documentation's date ordering identifies with the 26 campaign months. Client covariates: age and age squared, job, marital status, education, default (not 'no'), housing loan (yes/unknown), personal loan, contact channel, previous-campaign outcome and an indicator for two or more previous contacts. Excluded: duration (known only after the call) and day_of_week (an attribute of the last call only). Estimators: (i) life-table hazards h_k with Wilson 95% intervals; (ii) grouped-time proportional hazards (complementary log-log) by maximum likelihood with period fixed effects and client covariates, gamma heterogeneity with variance theta in closed form (Meyer 1990; Han and Hausman 1990), and the theta = 0 special case; (iii) profile likelihood over theta with the likelihood-ratio test of theta = 0 on the boundary (chi-bar-squared, Self and Liang) and the 95% profile interval; (iv) a split-population (never-convert) heterogeneity family as an alternative, also profiled; (v) the theta profile re-estimated on the 2008 campaigns only, to check that the bound on heterogeneity does not come only from high-conversion months. Decomposition: for each k = 2..10, log(h_k/h_1) is split exactly into (a) the bank's continuation policy (model-implied hazard in the observed at-risk pool versus the pool if every client had been called until conversion), (b) conversion-driven selection on observables, (c) conversion-driven selection on unobservables (zero at theta = 0), and (d) the structural contact effect (hazard of the contact-1 population with its covariates and heterogeneity frozen). 95% intervals by simulation from the cluster-robust (client-level) sandwich distribution of the estimates (Krinsky-Robb, 400 draws). A sequential decomposition (no controls, period effects, period plus client covariates) separates the calendar from client composition. Numerical study: 5-fold cross-fitting. On training folds fit the hazard models; on the held-out fold compute each client's conditional hazard path and the dynamic-programming stopping rule for a cost-to-margin ratio rho in {0.01, 0.02, 0.03} (the first contact is given; the rule decides contacts 2 onward). Policies are evaluated exactly by truncating the logged contact sequences: a client keeps her conversion only if it occurred at or before the rule's stopping point, and contacts after it are saved; rules are never allowed to call more than the bank did, so no counterfactual response model is needed. Benchmarks: the bank's practice, the best uniform cap and the best month-specific cap (both chosen on training folds), the one-step look-ahead rule, a client-only model without calendar information, and the rule computed at the upper limit of theta. Report profit in conversion-equivalents per 1,000 clients and as a percentage of the bank's profit, contacts saved and conversions retained, with 95% intervals from a Poisson bootstrap over held-out clients. Refutation: H1 as stated; H4 is refuted if the client-specific rule does not beat the best uniform cap at any rho, or if the theta = 0 and upper-theta rules differ in profit by more than their sampling error.
Novelty. The fatigue-versus-selection dichotomy omits the firm's own continuation policy and the calendar, which in campaign data jointly determine who is still being called. The decomposition here is exact and separates four channels; it replaces the fragile point estimate of a 'fatigue share' based on a few mass points with a profile-likelihood identified range that exploits the tenfold variation in conversion across months, where sorting would be fast if heterogeneity were large. It formalises when the decomposition matters for a stopping rule (only through the conditional hazard path), and evaluates rules on logged data by exact truncation rather than by simulation from a fitted response model. No published analysis of these data addresses contact-number dynamics; they are used as a classification benchmark.
Expected outcome. A falling raw hazard driven mainly by composition at early contacts (much of it calendar: high-conversion months' clients convert quickly and leave the pool), little or no structural decline through the third contact, a structural decline at later contacts, a tightly bounded heterogeneity variance, and stopping-rule gains over the bank that grow with the cost ratio, most of which come from adapting to the campaign month, with little difference between rules built under different heterogeneity assumptions.
Main risk. The bank's continuation decisions may depend on what was said in earlier calls (information about the client's receptiveness that the data do not record), which violates conditionally independent censoring and contaminates the policy and selection components; covariates are recorded only at the last contact; the public file does not record call direction, so some counted contacts may be inbound; subscription is assumed to occur at the last contact; separating fatigue from unobserved selection rests on the proportional-hazards functional form; and month effects in the stopping rules are estimated from other clients in the same month, an upper bound on what real-time monitoring could learn.
- Idea 2
Regime-adaptive campaign intensity: how much of the value of a contact cap comes from the macroeconomic regime, and what does a pooled cap cost?
Hypothesis. A contact cap that adapts to the Euribor regime (high versus low three-month rate) earns at least 5% more campaign profit than the best pooled cap at a cost-to-margin ratio of 2%, because the per-contact conversion hazard differs by an order of magnitude across regimes.
Dataset. Portuguese bank direct-marketing telephone campaigns bank_marketing_calls
Test plan, novelty and risk
Test plan. Classify the 26 campaign months by the Euribor rate at the median; estimate grouped-time hazards by regime with client covariates; choose pooled and regime-specific caps by cross-validation; evaluate profit by truncation of the logged contact sequences with bootstrap intervals; refuted if the regime-adaptive cap's gain over the pooled cap has a 95% interval including zero.
Novelty. Prior analyses of these data use macro indicators as predictors of a single call's success, not as inputs to the contact-intensity decision.
Expected outcome. A sizeable gain at moderate cost ratios, concentrated in the low-conversion 2008 regime.
Main risk. Euribor is confounded with the campaign's product terms and timing, so the regime proxies for many changes; few months fall in each regime.
- Idea 3
The cost of a linear temperature response in peak-load planning for Victoria
Hypothesis. A piecewise-linear (heating/cooling) temperature model with daily and weekly seasonality reduces the error of day-ahead daily-peak demand forecasts by at least 10% relative to a linear temperature model, and the linear model under-provisions capacity on the hottest 5% of days.
Dataset. Half-hourly electricity demand for Victoria, Australia vic_electricity_demand
Test plan, novelty and risk
Test plan. Aggregate half-hourly demand to daily peaks in local time; fit linear and piecewise temperature models with calendar effects on 2012-2013 and evaluate on 2014 with rolling origin; report MAPE, the 95th-percentile error and the reserve needed for a 99% service level, with block-bootstrap intervals; refuted if the MAPE improvement interval includes zero.
Novelty. Quantifies the operational (reserve-margin) cost of a common misspecification rather than only forecast accuracy.
Expected outcome. Clear gains on hot days, smaller on average days.
Main risk. Three years give few extreme heat days, so tail estimates are imprecise.
Ranking rationale
Idea 0 answers the posed question directly with a model, structural results and a numerical study of decision rules, and the bank data identify it unusually well because conversion varies tenfold across months. Idea 1 is a narrower policy question on the same data. Idea 2 is sound but off-topic.
3.Experimentation
Experiment succeeded in 1 round (88.875s).
Data provenance
Portuguese bank direct-marketing telephone campaigns
- Source
- UCI Machine Learning Repository (dataset 222), donated by S. Moro, P. Cortez and P. Rita
- Citation
- Moro, S., Cortez, P. and Rita, P. (2014) A data-driven approach to predict the success of bank telemarketing. Decision Support Systems 62, 22-31. DOI: 10.1016/j.dss.2014.03.001. Data: Moro, S., Rita, P. and Cortez, P. (2014) Bank Marketing [Dataset]. UCI Machine Learning Repository. DOI: 10.24432/C5K306. Note that this 41,188 x 20 'bank-additional-full' variant, with the socio-economic predictors, corresponds to the 2014 paper; the older 17-input 'bank-full' file belongs to Moro, Laureano and Cortez (2011) and should not be cited for it.
- Licence
- Creative Commons Attribution 4.0 International (CC BY 4.0): the UCI dataset page states "This dataset is licensed under a Creative Commons Attribution 4.0 International (CC BY 4.0) license. This allows for the sharing and adaptation of the datasets for any purpose, provided that the appropriate credit is given."
- File
- https://archive.ics.uci.edu/static/public/222/bank+marketing.zip
- Observations
- 41,188
- Variables (21)
- age, job, marital, education, default, housing, loan, contact, month, day_of_week, duration, campaign, pdays, previous, poutcome, emp_var_rate, cons_price_idx, cons_conf_idx, euribor3m, nr_employed, y
- Units
- y is the outcome, 'yes' or 'no' to subscribing a term deposit. duration is call length in seconds, campaign the number of contacts in this campaign, pdays days since the client was last contacted (999 means never) and previous the number of prior contacts. emp_var_rate, cons_price_idx, cons_conf_idx, euribor3m and nr_employed are quarterly or monthly Portuguese macro indicators merged onto each call. Categorical fields use the literal level 'unknown' for missing rather than an empty cell.
- Retrieved
- 24 Oct 2026, 03:31
- Checksum
- sha256 e0bf5f5de5b846e2…
Real, publicly available data — downloaded and checksum-verified before the analysis ran, never simulated.
Result (RESULT_JSON)
{
"n_clients": 41188,
"n_contacts": 105754,
"n_subscriptions": 4640,
"n_campaign_months": 26,
"share_2008_at_risk_k1_k4_k7_k10": [
0.672,
0.759,
0.851,
0.931
],
"share_campaign_le3": 0.8146,
"n_client_covariates": 31,
"loglik_theta0": -15236.87,
"loglik_month_only": -15532.45,
"max_gap_model_vs_raw_k2_9": 0.0069,
"life_table": [
[
1,
41188,
2300,
0.0558,
0.0537,
0.0581
],
[
2,
23546,
1211,
0.0514,
0.0487,
0.0543
],
[
3,
12976,
574,
0.0442,
0.0408,
0.0479
],
[
4,
7635,
249,
0.0326,
0.0289,
0.0368
],
[
5,
4984,
120,
0.0241,
0.0202,
0.0287
],
[
6,
3385,
75,
0.0222,
0.0177,
0.0277
],
[
7,
2406,
38,
0.0158,
0.0115,
0.0216
],
[
8,
1777,
17,
0.0096,
0.006,
0.0153
],
[
9,
1377,
17,
0.0123,
0.0077,
0.0197
],
[
10,
1094,
12,
0.011,
0.0063,
0.0191
],
[
11,
5386,
27,
0.005,
0.0034,
0.0073
]
],
"contact_effects": {
"month_client": {
"b": [
0.068,
0.07,
-0.134,
-0.351,
-0.324,
-0.549,
-0.944,
-0.623,
-0.721,
-1.405
],
"se": [
0.036,
0.047,
0.068,
0.095,
0.118,
0.164,
0.244,
0.246,
0.29,
0.201
],
"p": [
0.0578,
0.1405,
0.048,
0.0002,
0.0058,
0.0008,
0.0001,
0.0112,
0.0129,
0
]
},
"month_only": {
"b": [
0.065,
0.04,
-0.178,
-0.4,
-0.379,
-0.604,
-1.012,
-0.697,
-0.776,
-1.448
],
"se": [
0.036,
0.047,
0.067,
0.094,
0.118,
0.163,
0.245,
0.247,
0.289,
0.2
],
"p": [
0.0682,
0.3936,
0.008,
0,
0.0013,
0.0002,
0,
0.0047,
0.0072,
0
]
}
},
"decomposition": {
"raw": [
[
-0.082,
-0.15,
-0.015
],
[
-0.233,
-0.322,
-0.144
],
[
-0.538,
-0.666,
-0.409
],
[
-0.841,
-1.022,
-0.66
],
[
-0.924,
-1.152,
-0.697
],
[
-1.263,
-1.581,
-0.945
],
[
-1.764,
-2.239,
-1.289
],
[
-1.509,
-1.983,
-1.035
],
[
-1.627,
-2.192,
-1.063
]
],
"model": [
[
-0.078,
-0.136,
-0.015
],
[
-0.229,
-0.315,
-0.148
],
[
-0.538,
-0.655,
-0.414
],
[
-0.842,
-1.026,
-0.665
],
[
-0.917,
-1.136,
-0.684
],
[
-1.258,
-1.576,
-0.956
],
[
-1.764,
-2.235,
-1.292
],
[
-1.514,
-1.971,
-1.051
],
[
-1.623,
-2.168,
-1.089
]
],
"policy": [
[
0.013,
0.006,
0.024
],
[
-0.012,
-0.021,
0.007
],
[
-0.027,
-0.039,
-0.007
],
[
-0.052,
-0.066,
-0.028
],
[
-0.104,
-0.12,
-0.076
],
[
-0.18,
-0.202,
-0.15
],
[
-0.266,
-0.289,
-0.23
],
[
-0.31,
-0.336,
-0.272
],
[
-0.297,
-0.326,
-0.253
]
],
"sel_obs": [
[
-0.151,
-0.165,
-0.143
],
[
-0.279,
-0.301,
-0.266
],
[
-0.391,
-0.418,
-0.373
],
[
-0.471,
-0.503,
-0.449
],
[
-0.52,
-0.557,
-0.494
],
[
-0.576,
-0.616,
-0.545
],
[
-0.623,
-0.669,
-0.588
],
[
-0.632,
-0.677,
-0.593
],
[
-0.663,
-0.714,
-0.621
]
],
"sel_unobs": [
[
0,
0,
0
],
[
0,
0,
0
],
[
0,
0,
0
],
[
0,
0,
0
],
[
0,
0,
0
],
[
0,
0,
0
],
[
0,
0,
0
],
[
0,
0,
0
],
[
0,
0,
0
]
],
"within": [
[
0.061,
0.001,
0.123
],
[
0.062,
-0.023,
0.139
],
[
-0.12,
-0.237,
-0.003
],
[
-0.318,
-0.491,
-0.146
],
[
-0.294,
-0.508,
-0.076
],
[
-0.502,
-0.804,
-0.204
],
[
-0.875,
-1.336,
-0.426
],
[
-0.571,
-1.015,
-0.142
],
[
-0.663,
-1.194,
-0.152
]
],
"comp_share": [
[
1.779,
0.985,
7.009
],
[
1.27,
0.929,
1.939
],
[
0.777,
0.64,
0.992
],
[
0.622,
0.514,
0.78
],
[
0.68,
0.556,
0.892
],
[
0.601,
0.485,
0.784
],
[
0.504,
0.4,
0.672
],
[
0.623,
0.483,
0.866
],
[
0.592,
0.45,
0.855
]
],
"calendar_only_share": [
[
1.731,
0.908,
7.851
],
[
1.157,
0.851,
1.794
],
[
0.697,
0.572,
0.891
],
[
0.562,
0.468,
0.725
],
[
0.621,
0.505,
0.805
],
[
0.553,
0.449,
0.73
],
[
0.462,
0.371,
0.623
],
[
0.573,
0.443,
0.802
],
[
0.553,
0.417,
0.808
]
]
},
"gamma_pooled": {
"hat": 0.074,
"LR": 1.182,
"p": 0.1384,
"lo": 0,
"up": 0.221,
"rows": [
[
0,
0,
0.068,
-0.134,
-0.549,
-0.538,
0,
-0.12,
-1.258,
0,
-0.502
],
[
0.074,
1.182,
0.084,
-0.096,
-0.504,
-0.535,
-0.043,
-0.085,
-1.254,
-0.081,
-0.456
],
[
0.1,
1.048,
0.09,
-0.084,
-0.489,
-0.534,
-0.058,
-0.074,
-1.253,
-0.107,
-0.441
],
[
0.221,
-2.637,
0.115,
-0.028,
-0.426,
-0.532,
-0.123,
-0.025,
-1.251,
-0.218,
-0.377
],
[
0.5,
-24.99,
0.169,
0.083,
-0.301,
-0.533,
-0.249,
0.069,
-1.249,
-0.424,
-0.257
]
]
},
"split_pooled": {
"hat": 0.018,
"LR": 4.575,
"p": 0.0162,
"lo": 0.001,
"up": 0.045,
"rows": [
[
0.018,
4.575,
0.075,
-0.108,
-0.499,
-0.535,
-0.035,
-0.097,
-1.249,
-0.115,
-0.453
],
[
0.045,
0.792,
0.083,
-0.082,
-0.461,
-0.538,
-0.083,
-0.073,
-1.248,
-0.218,
-0.417
]
]
},
"regimes": {
"2008": {
"n": 27690,
"conversions": 1339,
"effects_theta0": {
"b": [
0.15,
0.124,
-0.098,
-0.288,
-0.145,
-0.617,
-0.802,
-0.29,
-0.772,
-1.307
],
"se": [
0.07,
0.085,
0.114,
0.147,
0.163,
0.234,
0.292,
0.254,
0.357,
0.214
],
"p": [
0.0317,
0.1457,
0.3877,
0.0507,
0.373,
0.0083,
0.006,
0.2536,
0.0305,
0
]
},
"gamma": {
"hat": 10.11,
"LR": 7.124,
"p": 0.0038,
"lo": 1.89,
"up": 24.972
},
"gamma_rows": [
[
10.11,
7.124,
0.378,
0.535,
0.43,
-0.214,
-0.628,
0.433,
-0.747,
-1.079,
0.35
],
[
1.89,
3.281,
0.193,
0.024,
-0.413,
-0.208,
-0.186,
0.022,
-0.744,
-0.313,
-0.378
],
[
24.972,
3.283,
0.696,
1.411,
1.895,
-0.216,
-1.094,
0.886,
-0.745,
-1.873,
1.131
]
],
"split_best": [
0.4,
1.833,
0.166,
-0.05,
-0.534,
-0.203,
-0.129,
-0.046,
-0.739,
-0.211,
-0.483
]
},
"2009_10": {
"n": 13498,
"conversions": 3301,
"effects_theta0": {
"b": [
0.041,
0.052,
-0.141,
-0.377,
-0.482,
-0.439,
-1.153,
-2.236,
-0.487,
-1.67
],
"se": [
0.042,
0.057,
0.085,
0.124,
0.172,
0.233,
0.45,
1.005,
0.501,
0.6
],
"p": [
0.3269,
0.364,
0.0944,
0.0024,
0.005,
0.0596,
0.0104,
0.0261,
0.3304,
0.0054
]
},
"gamma": {
"hat": 0.05,
"LR": 0.412,
"p": 0.2606,
"lo": 0,
"up": 0.222
},
"gamma_rows": [
[
0.05,
0.412,
0.055,
-0.105,
-0.383,
-0.421,
-0.036,
-0.091,
-0.896,
-0.065,
-0.338
],
[
0,
0,
0.041,
-0.141,
-0.439,
-0.423,
0,
-0.124,
-0.901,
0,
-0.391
],
[
0.222,
-3.332,
0.101,
0.012,
-0.216,
-0.418,
-0.15,
0.01,
-0.888,
-0.26,
-0.184
]
],
"split_best": [
0.1,
-25.379,
0.081,
0.004,
-0.129,
-0.451,
-0.203,
0.003,
-0.891,
-0.457,
-0.11
]
}
},
"pooled_decomp_regime_heterogeneity": {
"k4": [
-0.538,
-0.022,
-0.373,
-0.203,
0.06,
-0.087
],
"k7": [
-1.26,
-0.209,
-0.58,
-0.343,
-0.128,
-0.388
],
"k10": [
-1.621,
-0.408,
-0.662,
-0.438,
-0.113,
-0.443
]
},
"policy": {
"rho_0.01": {
"bank": [
86.972,
0,
0,
0,
-0.521,
-0.806,
-0.234,
100,
100
],
"cap": [
87.493,
0.599,
0.272,
0.957,
0,
0,
0,
94.757,
99.267
],
"month_cap": [
87.372,
0.46,
-0.031,
0.941,
-0.121,
-0.457,
0.194,
91.634,
98.448
],
"one_step": [
87.812,
0.965,
0.44,
1.504,
0.319,
-0.098,
0.69,
88.407,
98.103
],
"dp": [
87.85,
1.01,
0.506,
1.501,
0.357,
-0.049,
0.702,
89.298,
98.341
],
"dp_theta_up": [
87.854,
1.013,
0.536,
1.466,
0.361,
0.003,
0.662,
90.136,
98.534
],
"dp_regime": [
87.74,
0.883,
0.427,
1.351,
0.247,
-0.033,
0.492,
91.051,
98.642
],
"dp_client_only": [
87.855,
1.015,
0.463,
1.478,
0.362,
-0.081,
0.74,
89.28,
98.341
],
"dp_minus_dp_theta_up": [
-0.003,
-0.165,
0.128
],
"dp_minus_dp_regime": [
0.11,
-0.186,
0.379
],
"dp_minus_month_cap": [
0.478,
0.103,
0.861
],
"dp_minus_one_step": [
0.038,
-0.106,
0.194
],
"dp_vs_cap_by_month": [
0.302,
-0.215,
5.78,
8
],
"caps_chosen": [
7,
11
]
},
"rho_0.02": {
"bank": [
61.301,
0,
0,
0,
-3.151,
-3.785,
-2.55,
100,
100
],
"cap": [
64.452,
5.14,
4.061,
6.23,
0,
0,
0,
88.615,
97.608
],
"month_cap": [
66.608,
8.657,
6.904,
10.831,
2.156,
1.363,
3.168,
66.922,
89.634
],
"one_step": [
67.706,
10.449,
8.609,
12.672,
3.254,
2.388,
4.348,
64.31,
89.418
],
"dp": [
67.65,
10.358,
8.554,
12.552,
3.198,
2.362,
4.277,
64.56,
89.483
],
"dp_theta_up": [
67.482,
10.083,
8.294,
12.238,
3.03,
2.228,
4.098,
65.456,
89.741
],
"dp_regime": [
67.427,
9.994,
8.142,
12.35,
2.975,
2.126,
3.997,
64.238,
89.138
],
"dp_client_only": [
66.084,
7.802,
6.267,
9.614,
1.632,
0.932,
2.376,
76.453,
93.513
],
"dp_minus_dp_theta_up": [
0.168,
0.016,
0.329
],
"dp_minus_dp_regime": [
0.223,
-0.268,
0.706
],
"dp_minus_month_cap": [
1.043,
0.475,
1.589
],
"dp_minus_one_step": [
-0.056,
-0.126,
0.039
],
"dp_vs_cap_by_month": [
2.801,
-1.373,
16.512,
6
],
"caps_chosen": [
6
]
},
"rho_0.03": {
"bank": [
35.616,
0,
0,
0,
-7.443,
-8.475,
-6.484,
100,
100
],
"cap": [
43.059,
20.897,
17.492,
25.006,
0,
0,
0,
80.701,
93.405
],
"month_cap": [
52.823,
48.311,
42.018,
56.594,
9.764,
8.429,
11.04,
51.292,
81.961
],
"one_step": [
53.226,
49.442,
43.238,
57.35,
10.167,
8.929,
11.466,
52.124,
82.888
],
"dp": [
53.219,
49.424,
43.217,
57.213,
10.161,
8.911,
11.445,
52.258,
82.974
],
"dp_theta_up": [
53.188,
49.338,
43.056,
57.129,
10.13,
8.879,
11.401,
52.739,
83.276
],
"dp_regime": [
53.043,
48.929,
42.77,
56.871,
9.984,
8.763,
11.215,
52.267,
82.823
],
"dp_client_only": [
49.739,
39.654,
34.632,
45.961,
6.68,
5.688,
7.833,
64.184,
88.039
],
"dp_minus_dp_theta_up": [
0.031,
-0.141,
0.196
],
"dp_minus_dp_regime": [
0.177,
-0.174,
0.495
],
"dp_minus_month_cap": [
0.397,
-0.06,
0.843
],
"dp_minus_one_step": [
-0.006,
-0.089,
0.085
],
"dp_vs_cap_by_month": [
12.15,
-1.172,
38.708,
1
],
"caps_chosen": [
4
]
}
},
"H1_comp_share_k4_gt_half": true,
"H1_refuted": false,
"H2_no_decline_k1_k3": true,
"H3_structural_k7_negative_at_theta_up": true,
"H3_holds_within_2008": false,
"H4_decision_invariance": false,
"hypothesis_supported": true,
"revision": {
"client_names": [
"age10",
"age10_sq",
"job_blue-collar",
"job_entrepreneur",
"job_housemaid",
"job_management",
"job_retired",
"job_self-employed",
"job_services",
"job_student",
"job_technician",
"job_unemployed",
"job_unknown",
"marital_married",
"marital_single",
"marital_unknown",
"education_basic.6y",
"education_basic.9y",
"education_high.school",
"education_illiterate",
"education_professional.course",
"education_university.degree",
"education_unknown",
"housing_unknown",
"housing_yes",
"contact_telephone",
"poutcome_nonexistent",
"poutcome_success",
"default_not_no",
"loan_yes",
"previous_2plus"
],
"coef_pooled": {
"alpha": [
[
-4.035,
0.128
],
[
-3.967,
0.13
],
[
-3.966,
0.133
],
[
-4.169,
0.142
],
[
-4.386,
0.158
],
[
-4.359,
0.172
],
[
-4.585,
0.204
],
[
-4.98,
0.27
],
[
-4.659,
0.275
],
[
-4.756,
0.315
],
[
-5.44,
0.233
]
],
"months": [
[
"jun08",
0.131,
0.098
],
[
"jul08",
-0.012,
0.096
],
[
"aug08",
-0.177,
0.107
],
[
"oct08",
4.213,
0.177
],
[
"nov08",
0.221,
0.114
],
[
"dec08",
1.93,
1.02
],
[
"mar09",
2.095,
0.128
],
[
"apr09",
1.465,
0.097
],
[
"may09",
0.595,
0.095
],
[
"jun09",
2.201,
0.104
],
[
"jul09",
2.033,
0.147
],
[
"aug09",
2.024,
0.107
],
[
"sep09",
2.291,
0.13
],
[
"oct09",
2.397,
0.114
],
[
"nov09",
2.278,
0.118
],
[
"dec09",
2.15,
0.14
],
[
"mar10",
2.518,
0.123
],
[
"apr10",
2.348,
0.143
],
[
"may10",
2.442,
0.126
],
[
"jun10",
2.279,
0.131
],
[
"jul10",
2.32,
0.118
],
[
"aug10",
2.294,
0.132
],
[
"sep10",
2.251,
0.123
],
[
"oct10",
2.212,
0.139
],
[
"nov10",
2.265,
0.155
]
],
"client": [
[
-0.034,
0.019
],
[
0.023,
0.008
],
[
-0.094,
0.06
],
[
-0.074,
0.1
],
[
-0.092,
0.106
],
[
-0.019,
0.062
],
[
0.069,
0.082
],
[
-0.058,
0.087
],
[
-0.14,
0.066
],
[
0.065,
0.078
],
[
0.002,
0.053
],
[
-0.007,
0.088
],
[
-0.115,
0.184
],
[
0.054,
0.05
],
[
0.058,
0.058
],
[
0.149,
0.297
],
[
0.106,
0.09
],
[
0.044,
0.071
],
[
0.05,
0.069
],
[
0.572,
0.518
],
[
0.025,
0.077
],
[
0.068,
0.067
],
[
0.075,
0.084
],
[
-0.087,
0.104
],
[
-0.033,
0.03
],
[
-0.57,
0.058
],
[
0.3,
0.048
],
[
1.056,
0.057
],
[
-0.199,
0.054
],
[
-0.055,
0.042
],
[
0.008,
0.06
]
]
},
"coef_2008": {
"alpha": [
[
-4.691,
0.304
],
[
-4.54,
0.31
],
[
-4.567,
0.31
],
[
-4.789,
0.321
],
[
-4.979,
0.339
],
[
-4.836,
0.345
],
[
-5.308,
0.379
],
[
-5.493,
0.41
],
[
-4.981,
0.394
],
[
-5.462,
0.466
],
[
-5.997,
0.361
]
],
"months": [
[
"jun08",
0.128,
0.098
],
[
"jul08",
0.116,
0.135
],
[
"aug08",
0.061,
0.158
],
[
"oct08",
4.35,
0.181
],
[
"nov08",
0.451,
0.164
],
[
"dec08",
2.18,
1.019
]
],
"client": [
[
-0.104,
0.036
],
[
0.025,
0.034
],
[
0.144,
0.104
],
[
0.198,
0.149
],
[
-0.185,
0.211
],
[
0.209,
0.114
],
[
0.407,
0.173
],
[
0.024,
0.156
],
[
0.126,
0.111
],
[
0.479,
0.286
],
[
0.041,
0.099
],
[
-0.197,
0.213
],
[
0.287,
0.285
],
[
-0.022,
0.089
],
[
-0.011,
0.103
],
[
0.639,
0.473
],
[
0.247,
0.149
],
[
0.214,
0.121
],
[
0.176,
0.129
],
[
0.98,
1.034
],
[
0.073,
0.145
],
[
0.173,
0.129
],
[
0.033,
0.178
],
[
-0.36,
0.221
],
[
-0.053,
0.055
],
[
-0.435,
0.127
],
[
0.545,
0.218
],
[
1.023,
0.622
],
[
-0.037,
0.07
],
[
0.005,
0.076
],
[
-13.781,
0.924
]
]
},
"coef_2009_10": {
"alpha": [
[
-1.837,
0.143
],
[
-1.796,
0.145
],
[
-1.786,
0.149
],
[
-1.979,
0.162
],
[
-2.215,
0.185
],
[
-2.319,
0.22
],
[
-2.276,
0.267
],
[
-2.99,
0.466
],
[
-4.074,
1.012
],
[
-2.325,
0.518
],
[
-3.507,
0.613
]
],
"months": [
[
"apr09",
-0.6,
0.106
],
[
"may09",
-1.446,
0.104
],
[
"jun09",
0.097,
0.11
],
[
"jul09",
-0.085,
0.151
],
[
"aug09",
-0.094,
0.113
],
[
"sep09",
0.163,
0.134
],
[
"oct09",
0.28,
0.119
],
[
"nov09",
0.156,
0.124
],
[
"dec09",
0.028,
0.144
],
[
"mar10",
0.406,
0.132
],
[
"apr10",
0.222,
0.144
],
[
"may10",
0.324,
0.131
],
[
"jun10",
0.18,
0.138
],
[
"jul10",
0.203,
0.122
],
[
"aug10",
0.179,
0.139
],
[
"sep10",
0.136,
0.129
],
[
"oct10",
0.098,
0.144
],
[
"nov10",
0.139,
0.158
]
],
"client": [
[
-0.008,
0.023
],
[
0.021,
0.009
],
[
-0.217,
0.077
],
[
-0.234,
0.135
],
[
-0.046,
0.127
],
[
-0.111,
0.074
],
[
-0.038,
0.093
],
[
-0.083,
0.106
],
[
-0.276,
0.082
],
[
0.049,
0.084
],
[
-0.006,
0.063
],
[
0.037,
0.099
],
[
-0.281,
0.225
],
[
0.072,
0.06
],
[
0.075,
0.071
],
[
-0.117,
0.356
],
[
0.04,
0.113
],
[
-0.042,
0.09
],
[
0.007,
0.082
],
[
0.577,
0.594
],
[
0.015,
0.091
],
[
0.033,
0.08
],
[
0.079,
0.097
],
[
-0.006,
0.121
],
[
-0.026,
0.036
],
[
-0.613,
0.066
],
[
0.286,
0.05
],
[
1.043,
0.057
],
[
-0.432,
0.091
],
[
-0.083,
0.05
],
[
0.009,
0.061
]
]
},
"table3_extra": {
"holm_p": [
0.7517,
1,
0.1831,
0.0008,
0.0397,
0.0072,
0.0013,
0.1362,
0.1723,
0,
0.7517,
1,
0.7197,
0.0072,
0.1512,
0.0262,
0.0039,
0.2247,
0.2448,
0,
0.5178,
1,
1,
0.7197,
1,
0.1831,
0.1512,
1,
0.5178,
0,
1,
1,
0.9441,
0.0722,
0.1408,
0.7517,
0.218,
0.4693,
1,
0.1462
],
"ci_month_client": [
[
-0.002,
0.138
],
[
-0.023,
0.162
],
[
-0.266,
-0.001
],
[
-0.536,
-0.165
],
[
-0.555,
-0.094
],
[
-0.871,
-0.228
],
[
-1.424,
-0.465
],
[
-1.105,
-0.141
],
[
-1.289,
-0.153
],
[
-1.798,
-1.012
]
],
"ratio_lower": [
-0.002,
-0.023
]
},
"attenuation": {
"pooled": {
"scale": [
[
0.983,
0.038
],
[
1.008,
0.057
],
[
1.109,
0.094
],
[
0.92,
0.14
]
],
"tie": [
[
1.014,
0.026
],
[
0.991,
0.041
]
],
"LR": {
"scale_vs_none": [
2.635,
0.6206
],
"tie_vs_none": [
0.487,
0.784
],
"scale_vs_tie": [
2.149,
0.3415
]
},
"implied_hat": {
"theta": 0.074,
"scale": [
0.99,
0.973,
0.974,
0.949
],
"tie": [
0.986,
0.967
],
"wald": [
3.056,
0.5485
]
},
"implied_up": {
"theta": 0.221,
"scale": [
0.971,
0.923,
0.927,
0.87
],
"tie": [
0.959,
0.91
],
"wald": [
8.423,
0.0773
]
}
},
"y2008": {
"scale": [
[
0.313,
0.271
],
[
-0.373,
0.658
],
[
1.012,
0.494
],
[
1.331,
1.012
]
],
"tie": [
[
0.695,
0.214
],
[
0.818,
0.301
]
],
"LR": {
"scale_vs_none": [
12.158,
0.0162
],
"tie_vs_none": [
2.831,
0.2428
],
"scale_vs_tie": [
9.327,
0.0094
],
"full_vs_none": [
85.475,
0.1704,
74
]
},
"implied_hat": {
"theta": 10.11,
"scale": [
0.722,
0.433,
0.78,
0.454
],
"tie": [
0.764,
0.449
],
"wald": [
6.069,
0.1941
]
},
"implied_lo": {
"theta": 1.89,
"scale": [
0.934,
0.844,
0.967,
0.866
],
"tie": [
0.959,
0.861
],
"wald": [
11.499,
0.0215
]
}
},
"y2008_no_small": {
"n": 27613,
"scale": [
[
0.318,
0.267
],
[
-0.394,
0.626
],
[
0.971,
0.432
],
[
1.31,
0.894
]
],
"tie": [
[
0.7,
0.208
],
[
0.856,
0.287
]
],
"LR": {
"scale_vs_none": [
12.457,
0.0143
],
"tie_vs_none": [
2.798,
0.2469
],
"scale_vs_tie": [
9.659,
0.008
]
},
"theta_hat": [
9.962,
6.268
]
}
},
"loo_2008": [
[
"may08",
7763,
23.256,
13.176,
0.0001,
-0.661,
-1.827,
1.177
],
[
"jun08",
4374,
8.273,
6.535,
0.0053,
-0.782,
-0.965,
0.202
],
[
"jul08",
6685,
23.451,
17.045,
0,
-0.88,
-1.811,
0.936
],
[
"aug08",
5175,
6.466,
4.31,
0.0189,
-0.988,
-0.812,
-0.14
],
[
"oct08",
67,
10.15,
6.483,
0.0054,
-0.668,
-1.028,
0.376
],
[
"nov08",
3616,
9.118,
7.207,
0.0036,
-0.601,
-0.97,
0.375
],
[
"dec08",
10,
9.911,
6.921,
0.0043,
-0.745,
-1.062,
0.335
]
],
"profiles": {
"pooled_gamma": [
[
1,
-91.698
],
[
2,
-261.811
],
[
5,
-768.173
],
[
10,
-1401.46
],
[
20,
-2147.178
],
[
40,
-2775.88
]
],
"pooled_split": [
[
0.1,
-24.696
],
[
0.2,
-122.846
],
[
0.5,
-964.073
],
[
0.9,
-7638.563
]
],
"split_2008": {
"hat": 0.363,
"LR": 2.03,
"p": 0.0771,
"lo": 0,
"up": 0.503,
"k7": [
-0.742,
-0.198,
-0.498
]
},
"gamma_2008": [
[
0.5,
1.154
],
[
1,
2.048
],
[
2,
3.412
],
[
5,
5.889
],
[
10,
7.124
],
[
20,
5.004
],
[
40,
-1.87
]
],
"split_2009_10": {
"hat": 0.016,
"LR": 4.394,
"p": 0.018,
"lo": 0.001,
"up": 0.043,
"k7": [
-0.881,
-0.115,
-0.296
]
},
"gamma_2009_10": [
[
0.1,
0.068
],
[
0.5,
-20.448
],
[
1,
-68.761
],
[
5,
-511.963
],
[
40,
-1356.781
]
]
},
"upper_2_71": {
"pooled_gamma": 0.196,
"gamma_2009_10": 0.195
},
"regime_effects": {
"loglik": -15943.12,
"k4": {
"total": [
-0.537,
-0.665,
-0.412
],
"policy": [
-0.003,
-0.012,
0.011
],
"sel_obs": [
-0.274,
-0.296,
-0.257
],
"within": [
-0.26,
-0.386,
-0.135
],
"comp_share": [
0.516,
0.418,
0.674
]
},
"k7": {
"total": [
-1.264,
-1.584,
-0.942
],
"policy": [
-0.138,
-0.151,
-0.119
],
"sel_obs": [
-0.398,
-0.428,
-0.374
],
"within": [
-0.728,
-1.04,
-0.424
],
"comp_share": [
0.424,
0.337,
0.556
]
}
},
"tableA1_ci": {
"k4": [
[
-0.666,
-0.411
],
[
-0.033,
-0.002
],
[
-0.413,
-0.34
],
[
-0.236,
-0.182
],
[
-0.06,
0.177
]
],
"k7": [
[
-1.559,
-0.989
],
[
-0.296,
-0.115
],
[
-0.682,
-0.461
],
[
-0.478,
-0.247
],
[
-0.424,
0.153
]
],
"k10": [
[
-2.156,
-1.01
],
[
-0.691,
-0.096
],
[
-0.842,
-0.401
],
[
-0.735,
-0.238
],
[
-0.604,
0.527
]
]
}
},
"runtime_seconds": 88.282
}Experiment script · experiment_v1.py
"""Contact fatigue, dynamic selection or the bank's continuation policy? Grouped-time proportional hazards for the
number of telemarketing contacts to subscription (real UCI bank data), gamma and split-population heterogeneity with
profile likelihoods (pooled and by regime), an exact decomposition of the declining aggregate hazard, and client-specific
stopping rules evaluated by truncating the logged contact sequences. Section 6 (revision) adds the Proposition 2 attenuation
test, 2008 leave-one-month-out profiles, completed profiles, coefficient tables, a regime-effects decomposition, one-sided
bounds and Krinsky-Robb intervals for the regime-heterogeneity decomposition; it uses its own random stream."""
import time, json
import numpy as np, pandas as pd
from scipy.optimize import minimize, minimize_scalar, brentq
from scipy.stats import chi2, norm
from dataset import load
T0 = time.time(); RNG = np.random.default_rng(20261024); CRIT = chi2.ppf(0.95, 1)
df = load()
print('loaded', df.shape, '| missing cells', int(df.isna().sum().sum()), '| dtypes', df.dtypes.astype(str).value_counts().to_dict())
print('unknown levels', {c: int((df[c] == 'unknown').sum()) for c in df.columns if not pd.api.types.is_numeric_dtype(df[c])})
assert df.y.isin(['yes', 'no']).all() and (df.campaign >= 1).all()
N = len(df); t = df.campaign.to_numpy(int); d = (df.y == 'yes').to_numpy(int)
# calendar: 26 (month, CPI) cells numbered in order of first appearance (the file is ordered by date)
cell = df.groupby(['month', 'cons_price_idx'], sort=False).ngroup().to_numpy()
mn = df.month.groupby(cell).first().map({m: i for i, m in enumerate(
['jan', 'feb', 'mar', 'apr', 'may', 'jun', 'jul', 'aug', 'sep', 'oct', 'nov', 'dec'])}).to_numpy()
year = (2008 + np.r_[0, np.cumsum(np.diff(mn) < 0)])[cell]
cl = pd.DataFrame({'age10': (df.age - 40) / 10}); cl['age10_sq'] = cl.age10 ** 2
for v in ['job', 'marital', 'education', 'housing', 'contact', 'poutcome']:
cl = cl.join(pd.get_dummies(df[v], prefix=v, drop_first=True, dtype=float))
cl['default_not_no'] = (df.default != 'no').astype(float); cl['loan_yes'] = (df.loan == 'yes').astype(float)
cl['previous_2plus'] = (df.previous >= 2).astype(float)
C = cl.to_numpy(float); K = 11 # baseline categories: contacts 1..10 and 11+
def cell_dummies(idx): # month fixed effects for the cells present in idx
cs = np.unique(cell[idx]); return (cell[:, None] == cs[None, 1:]).astype(float)
ALL, I08, I09 = np.arange(N), np.where(year == 2008)[0], np.where(year > 2008)[0]
XF, XP = np.hstack([cell_dummies(ALL), C]), cell_dummies(ALL)
print(f'N={N} clients, {t.sum()} contacts, {d.sum()} subscriptions, {cell.max() + 1} campaign months '
f'({len(I08)} clients in 2008), {C.shape[1]} client covariates')
def ncount(k): # n_c(k): contacts <= k in baseline category c
k = np.asarray(k)
return np.stack([(k >= c + 1).astype(float) for c in range(K - 1)] + [np.maximum(k - (K - 1), 0.0)], -1)
def logS(u, fam, h): # log survival and d log S / du for each heterogeneity family
if fam == 'gamma' and h > 0:
return -np.log1p(h * u) / h, -1.0 / (1.0 + h * u)
if fam == 'split' and h > 0:
e = np.exp(-u); s = h + (1 - h) * e
return np.log(s), -(1 - h) * e / s
return -u, -np.ones_like(u)
def data(idx, X):
return dict(t=t[idx], d=d[idx], N1=ncount(t[idx]), N0=ncount(t[idx] - 1), X=X[idx], idx=idx)
def nll(par, D, fam, h, scores=False): # client-level likelihood of the grouped PH model
ea = np.exp(par[:K]); c = np.exp(D['X'] @ par[K:])
u1, u0 = c * (D['N1'] @ ea), c * (D['N0'] @ ea)
(a1, g1), (a0, g0) = logS(u1, fam, h), logS(u0, fam, h)
ev = D['d'] == 1; gap = np.maximum(a0 - a1, 1e-300); q = 1.0 / np.expm1(np.minimum(gap, 700))
ll = np.where(ev, a0 + np.log(-np.expm1(-gap)), a1)
du1 = np.where(ev, -q * g1, g1); du0 = np.where(ev, (1 + q) * g0, 0.0)
ga = (du1 * c)[:, None] * D['N1'] * ea + (du0 * c)[:, None] * D['N0'] * ea; ge = du1 * u1 + du0 * u0
if scores:
return np.hstack([ga, ge[:, None] * D['X']])
return -ll.mean(), -np.r_[ga.sum(0), D['X'].T @ ge] / len(ll)
def fit(D, fam='none', h=0.0, x0=None):
if x0 is None:
hz = [((D['t'] == k) & (D['d'] == 1)).sum() / max((D['t'] >= k).sum(), 1) for k in range(1, K + 1)]
x0 = np.r_[np.log(-np.log(1 - np.clip(hz, 1e-3, 0.5))), np.zeros(D['X'].shape[1])]
r = minimize(nll, x0, args=(D, fam, h), jac=True, method='L-BFGS-B', options=dict(maxiter=5000, maxcor=30, ftol=1e-13, gtol=1e-8))
return r.x, -r.fun * len(D['t'])
def surv(par, X, fam, h, kmax=10): # survival S(0..kmax) and frozen structural hazards per client
ea = np.exp(par[:K]); c = np.exp(X @ par[K:])
S = np.exp(logS(c[:, None] * (ncount(np.arange(kmax + 1)) @ ea)[None, :], fam, h)[0])
return S, 1 - np.exp(logS(c[:, None] * ea[None, :kmax], fam, h)[0])
def agg(S, Hs, tt): # exact split of log(h_k / h_1): total, policy, obs., unobs., structural
hz = 1 - S[:, 1:] / S[:, :-1]
obs = np.array([hz[tt >= k, k - 1].mean() for k in range(1, hz.shape[1] + 1)])
conv = (S[:, :-1] * hz).sum(0) / S[:, :-1].sum(0); W = hz.mean(0); H = Hs.mean(0)
return np.array([np.log(obs / obs[0]), np.log(obs / conv), np.log(conv / W), np.log(W / H), np.log(H / H[0])])
decomp = lambda par, D, fam, h: agg(*surv(par, D['X'], fam, h), D['t'])
def vcov(par, D, fam, h): # client-clustered sandwich covariance
G = nll(par, D, fam, h, scores=True); E = np.eye(len(par)) * 1e-5
Hm = np.array([nll(par + e, D, fam, h)[1] - nll(par - e, D, fam, h)[1] for e in E]) / 2e-5
Hi = np.linalg.inv((Hm + Hm.T) / 2); return Hi @ (G.T @ G) @ Hi / len(D['t']) ** 2
def effects(par, V): # alpha_k - alpha_1, k = 2..10 and 11+
se = np.sqrt(np.diag(V)[1:K] + V[0, 0] - 2 * V[0, 1:K]); b = par[1:K] - par[0]
return dict(b=[r3(x) for x in b], se=[r3(x) for x in se], p=[r4(2 * norm.sf(abs(x))) for x in b / se])
r3 = lambda x: round(float(x), 3); r4 = lambda x: round(float(x), 4)
pct = lambda a: [r3(np.percentile(a, 2.5)), r3(np.percentile(a, 97.5))]
cache = {}
def prof(h, D, key, fam): # refit at fixed heterogeneity h, warm-started from the nearest fit
lst = cache[(key, fam)]; x0 = min(lst, key=lambda e: abs(np.log1p(e[0]) - np.log1p(h)))[1]
x, ll = fit(D, fam, h, x0); lst.append((h, x)); return ll, x
def profile(D, key, fam, x0, ll0, bound, hmax, xt): # MLE, LR test of h = 0 and 95% profile-likelihood interval
cache[(key, fam)] = [(0.0, x0)]
hh = float(minimize_scalar(lambda h: -prof(h, D, key, fam)[0], bounds=(0.0, bound), method='bounded', options={'xatol': xt}).x)
llh = prof(hh, D, key, fam)[0]
if llh < ll0: hh, llh = 0.0, ll0
dev = lambda h: 2 * (llh - prof(h, D, key, fam)[0]) - CRIT
lo = brentq(dev, 0.0, hh, xtol=xt) if hh > 0 and dev(0.0) > 0 else 0.0
up = brentq(dev, hh + xt, hmax, xtol=xt) if dev(hmax) > 0 else None
LR = 2 * (llh - ll0)
return dict(hat=r3(hh), LR=r3(LR), p=r4(0.5 * chi2.sf(LR, 1)), lo=r3(lo), up=None if up is None else r3(up)), hh, lo, up
# ---------- 1. life table (contacts 1-10; 11+ pooled over contact-periods) ----------
life = []
for k in range(1, K + 1):
n_k = int((t >= k).sum()) if k < K else int(np.maximum(t - (K - 1), 0).sum())
e_k = int(((t == k) & (d == 1)).sum()) if k < K else int(((t >= K) & (d == 1)).sum())
p, z = e_k / n_k, 1.959964; m = z * np.sqrt(p * (1 - p) / n_k + z * z / (4 * n_k * n_k)) / (1 + z * z / n_k)
mid = (p + z * z / (2 * n_k)) / (1 + z * z / n_k); life.append([k, n_k, e_k, r4(p), r4(mid - m), r4(mid + m)])
hx = np.array([x[2] / x[1] for x in life[:10]]); ex = np.array([x[2] for x in life[:10]])
raw = np.log(hx / hx[0]); raw_se = np.sqrt((1 - hx) / ex + (1 - hx[0]) / ex[0])
# ---------- 2. pooled models without heterogeneity, Krinsky-Robb intervals for the decomposition ----------
DF, DP = data(ALL, XF), data(ALL, XP)
pF, llF = fit(DF); pP, llP = fit(DP); VF, VP = vcov(pF, DF, 'none', 0.0), vcov(pP, DP, 'none', 0.0)
drF = np.array([decomp(z, DF, 'none', 0.0) for z in RNG.multivariate_normal(pF, VF, size=1000, method='cholesky')])
drP = np.array([decomp(z, DP, 'none', 0.0) for z in RNG.multivariate_normal(pP, VP, size=1000, method='cholesky')])
dF, dP = decomp(pF, DF, 'none', 0.0), decomp(pP, DP, 'none', 0.0)
with np.errstate(divide='ignore', invalid='ignore'):
share = lambda a: 1 - a[..., 4, :] / a[..., 0, :] # composition (non-structural) share of the log decline
shF, shP, shFd, shPd = share(dF), share(dP), share(drF), share(drP)
est = lambda a, dr, j, k: [r3(a[j, k - 1])] + pct(dr[:, j, k - 1])
dec = {'raw': [[r3(raw[k - 1]), r3(raw[k - 1] - 1.96 * raw_se[k - 1]), r3(raw[k - 1] + 1.96 * raw_se[k - 1])] for k in range(2, 11)],
**{nm: [est(dF, drF, j, k) for k in range(2, 11)] for j, nm in enumerate(['model', 'policy', 'sel_obs', 'sel_unobs', 'within'])},
'comp_share': [[r3(shF[k - 1])] + pct(shFd[:, k - 1]) for k in range(2, 11)],
'calendar_only_share': [[r3(shP[k - 1])] + pct(shPd[:, k - 1]) for k in range(2, 11)]}
eff = {'month_client': effects(pF, VF), 'month_only': effects(pP, VP)}
print(f'pooled theta=0 {time.time() - T0:.1f}s | composition share k4 {dec["comp_share"][2]} k7 {dec["comp_share"][5]}')
# ---------- 3. unobserved heterogeneity: pooled gamma and split-population profiles, then by regime ----------
def rows(D, key, fam, hs, ll0):
out = []
for h in hs:
ll, x = prof(h, D, key, fam); a = decomp(x, D, fam, h)
out.append([r3(h), r3(2 * (ll - ll0)), r3(x[1] - x[0]), r3(x[3] - x[0]), r3(x[6] - x[0])] + [r3(a[j, k - 1]) for k in (4, 7) for j in (0, 3, 4)])
return out # [h, LR, a2-a1, a4-a1, a7-a1, (total, unobs, structural) at k=4 and k=7]
gam, th_hat, _, th_up = profile(DF, 'F', 'gamma', pF, llF, 0.3, 1.0, 2e-3)
gam['rows'] = rows(DF, 'F', 'gamma', [0.0, th_hat, 0.1, th_up, 0.5], llF); p_up = prof(th_up, DF, 'F', 'gamma')[1]
spl, pi_hat, _, pi_up = profile(DF, 'F', 'split', pF, llF, 0.2, 0.3, 1e-3)
spl['rows'] = rows(DF, 'F', 'split', [pi_hat, pi_up], llF)
reg = {}
for lab, I in [('2008', I08), ('2009_10', I09)]:
D = data(I, np.hstack([cell_dummies(I), C])); p0, l0 = fit(D); g, hh, lo, up = profile(D, lab, 'gamma', p0, l0, 40.0, 100.0, 0.02)
cache[(lab, 'split')] = [(0.0, p0)]; s = rows(D, lab, 'split', [0.1, 0.2, 0.3, 0.4], l0)
reg[lab] = dict(n=len(I), conversions=int(d[I].sum()), effects_theta0=effects(p0, vcov(p0, D, 'none', 0.0)), gamma=g,
gamma_rows=rows(D, lab, 'gamma', [hh, lo] + ([up] if up else []), l0), split_best=max(s, key=lambda r: r[1]),
D=D, p_hat=prof(hh, D, lab, 'gamma')[1], hh=hh, lo=lo)
def regime_decomp(which): # pooled decomposition with regime-specific heterogeneity
parts = [surv(prof(reg[r][which], reg[r]['D'], r, 'gamma')[1], reg[r]['D']['X'], 'gamma', reg[r][which]) for r in reg]
return agg(np.vstack([p[0] for p in parts]), np.vstack([p[1] for p in parts]), np.r_[t[I08], t[I09]])
dRh, dRlo = regime_decomp('hh'), regime_decomp('lo')
regime_pooled = {f'k{k}': [r3(x) for x in np.r_[dRh[:, k - 1], dRlo[4, k - 1]]] for k in (4, 7, 10)}
print(f'heterogeneity {time.time() - T0:.1f}s | theta pooled {th_hat:.3f} up {th_up:.3f} | pi {pi_hat:.3f} up {pi_up:.3f} | '
f'2008 {reg["2008"]["gamma"]["hat"]} [{reg["2008"]["gamma"]["lo"]}, {reg["2008"]["gamma"]["up"]}] | 2009-10 {reg["2009_10"]["gamma"]["hat"]} up {reg["2009_10"]["gamma"]["up"]}')
# ---------- 4. numerical study: cross-fitted stopping rules evaluated by truncation of the logged sequences ----------
HZ, RHOS = 20, [0.01, 0.02, 0.03]
def paths(par, X, fam, h): # conditional hazard at contacts 1..HZ given survival
ea = np.exp(par[:K]); c = np.exp(X @ par[K:]); A = logS(c[:, None] * (ncount(np.arange(HZ + 1)) @ ea)[None, :], fam, h)[0]
return 1 - np.exp(A[:, 1:] - A[:, :-1])
def stop_at(go): # contacts placed; the first contact is always made
go[:, 0] = True; return np.where((~go).any(1), (~go).argmax(1), HZ)
def dp(hz, rho): # backward induction over the predicted hazard path
V, go = np.zeros(len(hz)), np.zeros(hz.shape, bool)
for k in range(HZ - 1, -1, -1):
Q = hz[:, k] - rho + (1 - hz[:, k]) * V; go[:, k] = Q > 0; V = np.maximum(Q, 0)
return stop_at(go)
def cap_best(idx, rho): # best common cap on the given training clients
return 1 + int(np.argmax([(d[idx] * (t[idx] <= kc)).sum() - rho * np.minimum(t[idx], kc).sum() for kc in range(1, HZ + 1)]))
POL = ['bank', 'cap', 'month_cap', 'one_step', 'dp', 'dp_theta_up', 'dp_regime', 'dp_client_only']
fold = RNG.permutation(N) % 5; tau = {(r, p): np.zeros(N, int) for r in RHOS for p in POL}
for f in range(5):
tr, te = np.where(fold != f)[0], np.where(fold == f)[0]
hz = {'dp': paths(fit(data(tr, XF), x0=pF)[0], XF[te], 'none', 0.0),
'dp_theta_up': paths(fit(data(tr, XF), 'gamma', th_up, p_up)[0], XF[te], 'gamma', th_up),
'dp_client_only': paths(fit(data(tr, C))[0], C[te], 'none', 0.0), 'dp_regime': np.zeros((len(te), HZ))}
for r, I in [('2008', I08), ('2009_10', I09)]:
trr, ter = np.intersect1d(tr, I), np.isin(te, I); Xr = np.hstack([cell_dummies(I), C])
hz['dp_regime'][ter] = paths(fit(data(trr, Xr), 'gamma', reg[r]['hh'], reg[r]['p_hat'])[0], Xr[te[ter]], 'gamma', reg[r]['hh'])
for r in RHOS:
tau[(r, 'bank')][te] = 10 ** 6; tau[(r, 'cap')][te] = cap_best(tr, r); tau[(r, 'one_step')][te] = stop_at(hz['dp'] >= r)
for m in np.unique(cell[te]):
tau[(r, 'month_cap')][te[cell[te] == m]] = cap_best(tr[cell[tr] == m], r) if (cell[tr] == m).any() else cap_best(tr, r)
for p in ['dp', 'dp_theta_up', 'dp_regime', 'dp_client_only']:
tau[(r, p)][te] = dp(hz[p], r)
print(f'cross-fitting {time.time() - T0:.1f}s')
W = RNG.poisson(1.0, size=(500, N)).astype(np.float32); policy = {}
for r in RHOS: # per policy: profit/1000, %vs bank [ci], gain vs cap/1000 [ci], contacts %, conversions %
conv = {p: np.where(tau[(r, p)] >= t, d, 0) for p in POL}; cont = {p: np.minimum(t, tau[(r, p)]) for p in POL}
pr = np.stack([conv[p] - r * cont[p] for p in POL], 1).astype(np.float32); B = W @ pr; nB = W.sum(1); tot = pr.sum(0)
res = {p: [r3(1000 * tot[j] / N), r3(100 * (tot[j] / tot[0] - 1))] + pct(100 * (B[:, j] / B[:, 0] - 1)) + [r3(1000 * (tot[j] - tot[1]) / N)]
+ pct(1000 * (B[:, j] - B[:, 1]) / nB) + [r3(100 * cont[p].sum() / t.sum()), r3(100 * conv[p].sum() / d.sum())] for j, p in enumerate(POL)}
for a, b in [('dp', 'dp_theta_up'), ('dp', 'dp_regime'), ('dp', 'month_cap'), ('dp', 'one_step')]:
ja, jb = POL.index(a), POL.index(b); res[f'{a}_minus_{b}'] = [r3(1000 * (tot[ja] - tot[jb]) / N)] + pct(1000 * (B[:, ja] - B[:, jb]) / nB)
gm = np.array([1000 * (pr[cell == m, 4] - pr[cell == m, 1]).mean() for m in range(cell.max() + 1)])
res['dp_vs_cap_by_month'] = [r3(gm.mean()), r3(gm.min()), r3(gm.max()), int((gm < -1e-9).sum())]
res['caps_chosen'] = sorted({int(x) for x in tau[(r, 'cap')]}); policy[f'rho_{r}'] = res
# ---------- 5. hypotheses and results ----------
k4 = dec['comp_share'][2]; e = eff['month_client']; g8 = reg['2008']
H1 = bool(k4[1] > 0.5); H2 = bool(not any(e['b'][j] < 0 and e['p'][j] < 0.05 for j in (0, 1)))
H3 = bool(gam['rows'][3][10] < 0) # structural log effect at k=7 still negative at theta_up (pooled)
H3_2008 = bool(all(r[10] < 0 for r in g8['gamma_rows']))
H4 = bool(any(policy[f'rho_{r}']['dp'][5] > 0 for r in RHOS) and all(policy[f'rho_{r}'][f'dp_minus_{b}'][1] <= 0 <= policy[f'rho_{r}'][f'dp_minus_{b}'][2]
for r in RHOS for b in ('dp_theta_up', 'dp_regime')))
# ---------- 6. revision analyses (separate random stream: every earlier number is unchanged) ----------
RNG2, T6, rv = np.random.default_rng([20261024, 2]), time.time(), {}
lab_m = [f"{df.month[cell == c].iloc[0]}{str(year[cell == c][0])[2:]}" for c in range(cell.max() + 1)]
def coefs(par, V, cells): # estimates with robust SEs (one likelihood term per client)
se, nm = np.sqrt(np.diag(V)), len(cells) - 1
return dict(alpha=[[r3(a), r3(b)] for a, b in zip(par[:K], se[:K])], months=[[lab_m[c], r3(a), r3(b)] for c, a, b in zip(cells[1:], par[K:K + nm], se[K:K + nm])],
client=[[r3(a), r3(b)] for a, b in zip(par[K + nm:], se[K + nm:])])
def zp(par, V):
se = np.sqrt(np.diag(V)[1:K] + V[0, 0] - 2 * V[0, 1:K]); b = par[1:K] - par[0]; return b, se, 2 * norm.sf(np.abs(b / se))
rv['client_names'] = list(cl.columns); rv['coef_pooled'] = coefs(pF, VF, np.unique(cell)); pvals = [zp(pP, VP)[2], zp(pF, VF)[2]]
for lab, I in [('2008', I08), ('2009_10', I09)]:
D = reg[lab]['D']; p0 = cache[(lab, 'gamma')][0][1]; V0 = vcov(p0, D, 'none', 0.0)
reg[lab].update(p0=p0, l0=-nll(p0, D, 'none', 0.0)[0] * len(I)); rv[f'coef_{lab}'] = coefs(p0, V0, np.unique(cell[I])); pvals.append(zp(p0, V0)[2])
pv = np.concatenate(pvals); o = np.argsort(pv); hp = np.maximum.accumulate(np.minimum(1, (len(pv) - np.arange(len(pv))) * pv[o])); holm = np.empty(len(pv)); holm[o] = hp
b, se, _ = zp(pF, VF)
rv['table3_extra'] = dict(holm_p=[r4(x) for x in holm], ci_month_client=[[r3(x - 1.96 * s), r3(x + 1.96 * s)] for x, s in zip(b, se)],
ratio_lower=[r3(np.exp(x - 1.96 * s) - 1) for x, s in zip(b[:2], se[:2])])
# 6a. Proposition 2: do covariate contrasts attenuate across contact bins (1, 2-3, 4+) as the frailty estimates imply?
BIN = np.array([0, 1, 1] + [2] * (K - 3))
def pp_data(idx, M): # person-period rows (contact by contact) for clients idx
tt = t[idx]; row = np.repeat(np.arange(len(idx)), tt); kk = np.concatenate([np.arange(1, x + 1) for x in tt]); cat = np.minimum(kk, K) - 1
y = np.zeros(len(row)); y[np.cumsum(tt) - 1] = d[idx]
return dict(row=row, k=kk, cat=cat, bin=BIN[cat], y=y, M=M[idx][row], Z=C[idx][row], n=len(idx), start=np.r_[0, np.cumsum(tt)[:-1]])
def pp_eta(par, P, mode): # theta = 0 cloglog with bin-specific scales ('scale', 'tie') or coefficients ('full')
a, b, pM, pZ = par[:K], P['bin'], P['M'].shape[1], P['Z'].shape[1]
if mode == 'full':
return a[P['cat']] + np.einsum('ij,ij->i', P['X'], par[K:].reshape(3, pM + pZ)[b]), None
sM, sZ, g = P['M'] @ par[K:K + pM], P['Z'] @ par[K + pM:K + pM + pZ], par[K + pM + pZ:]
gM = np.r_[1.0, g[:2]]; gZ = np.r_[1.0, g[2:]] if mode == 'scale' else gM
return a[P['cat']] + gM[b] * sM + gZ[b] * sZ, (sM, sZ, gM, gZ)
def pp_parts(par, P, y, mode):
eta, aux = pp_eta(par, P, mode); mu = np.exp(np.minimum(eta, 50))
return y * np.log(-np.expm1(-mu)) - (1 - y) * mu, y * mu / np.expm1(np.minimum(mu, 700)) - (1 - y) * mu, aux
def pp_nll(par, P, y, mode):
ll, w, aux = pp_parts(par, P, y, mode); b = P['bin']; G = [np.bincount(P['cat'], w, K)]
if mode == 'full':
G += [P['X'][b == j].T @ w[b == j] for j in range(3)]
else:
sM, sZ, gM, gZ = aux; G += [P['M'].T @ (w * gM[b]), P['Z'].T @ (w * gZ[b])]
G += [[w[b == j] @ s[b == j] for s in ((sM, sZ) if mode == 'scale' else (sM + sZ,)) for j in (1, 2)]]
return -ll.sum() / P['n'], -np.concatenate([np.ravel(x) for x in G]) / P['n']
def pp_fit(P, y, x0, mode):
r = minimize(pp_nll, x0, args=(P, y, mode), jac=True, method='L-BFGS-B', options=dict(maxiter=5000, maxcor=30, ftol=1e-12, gtol=1e-8))
return r.x, -r.fun * P['n']
def pp_vcov(par, P, mode): # robust sandwich, one score per client
E = np.eye(len(par)) * 1e-5; Hm = np.array([pp_nll(par + e, P, P['y'], mode)[1] - pp_nll(par - e, P, P['y'], mode)[1] for e in E]) / 2e-5
_, w, (sM, sZ, gM, gZ) = pp_parts(par, P, P['y'], mode); b = P['bin']
S = [np.eye(K)[P['cat']] * w[:, None], P['M'] * (w * gM[b])[:, None], P['Z'] * (w * gZ[b])[:, None]]
S += [np.stack([w * s * (b == j) for s in ((sM, sZ) if mode == 'scale' else (sM + sZ,)) for j in (1, 2)], 1)]
G = np.add.reduceat(np.hstack(S), P['start'], axis=0); Hi = np.linalg.inv((Hm + Hm.T) / 2); return Hi @ (G.T @ G) @ Hi / P['n'] ** 2
def attenuation(I, Mfull, D, key, p0, l0, thetas, full):
P = pp_data(I, Mfull); xs, ls = pp_fit(P, P['y'], np.r_[p0, 1, 1, 1, 1], 'scale'); xt, lt = pp_fit(P, P['y'], np.r_[p0, 1, 1], 'tie')
Vs, Vt = pp_vcov(xs, P, 'scale'), pp_vcov(xt, P, 'tie'); LR = lambda a, c, k: [r3(a - c), r4(chi2.sf(a - c, k))]
res = dict(scale=[[r3(g), r3(s)] for g, s in zip(xs[-4:], np.sqrt(np.diag(Vs))[-4:])], tie=[[r3(g), r3(s)] for g, s in zip(xt[-2:], np.sqrt(np.diag(Vt))[-2:])],
LR=dict(scale_vs_none=LR(2 * ls, 2 * l0, 4), tie_vs_none=LR(2 * lt, 2 * l0, 2), scale_vs_tie=LR(2 * ls, 2 * lt, 2)))
for nm, th in thetas: # attenuation the theta = 0 bins model would show if the gamma fit were true
S = surv(prof(th, D, key, 'gamma')[1], D['X'], 'gamma', th, kmax=int(P['k'].max()))[0]; yi = (1 - S[:, 1:] / S[:, :-1])[P['row'], P['k'] - 1]
gi, gti = pp_fit(P, yi, xs, 'scale')[0][-4:], pp_fit(P, yi, xt, 'tie')[0][-2:]; dv = xs[-4:] - gi; W = float(dv @ np.linalg.solve(Vs[-4:, -4:], dv))
res[nm] = dict(theta=r3(th), scale=[r3(x) for x in gi], tie=[r3(x) for x in gti], wald=[r3(W), r4(chi2.sf(W, 4))])
if full:
P['X'] = np.hstack([P['M'], P['Z']]); lf = pp_fit(P, P['y'], np.r_[p0[:K], np.tile(p0[K:], 3)], 'full')[1]
res['LR']['full_vs_none'] = LR(2 * lf, 2 * l0, 2 * P['X'].shape[1]) + [2 * P['X'].shape[1]]
return res
rv['attenuation'] = dict(pooled=attenuation(ALL, XP, DF, 'F', pF, llF, [('implied_hat', th_hat), ('implied_up', th_up)], False),
y2008=attenuation(I08, cell_dummies(I08), reg['2008']['D'], '2008', reg['2008']['p0'], reg['2008']['l0'],
[('implied_hat', reg['2008']['hh']), ('implied_lo', reg['2008']['lo'])], True))
I = I08[~np.isin(cell[I08], [4, 6])]; Ms = cell_dummies(I); Ds = data(I, np.hstack([Ms, C])); ps, lsm = fit(Ds); cache[('small', 'gamma')] = [(0.0, ps)]
rv['attenuation']['y2008_no_small'] = dict(n=len(I), **{k: v for k, v in attenuation(I, Ms, Ds, 'small', ps, lsm, [], False).items() if k in ('scale', 'tie', 'LR')})
hs = float(minimize_scalar(lambda h: -prof(h, Ds, 'small', 'gamma')[0], bounds=(0.0, 40.0), method='bounded', options={'xatol': 0.1}).x)
rv['attenuation']['y2008_no_small']['theta_hat'] = [r3(hs), r3(2 * (prof(hs, Ds, 'small', 'gamma')[0] - lsm))]
print(f'attenuation {time.time() - T6:.1f}s')
# 6b. 2008 leave-one-month-out profiles
rv['loo_2008'] = []
for m in np.unique(cell[I08]):
I = I08[cell[I08] != m]; D = data(I, np.hstack([cell_dummies(I), C])); pm, lm = fit(D); cache[('loo', 'gamma')] = [(0.0, pm)]
hh = float(minimize_scalar(lambda h: -prof(h, D, 'loo', 'gamma')[0], bounds=(0.0, 40.0), method='bounded', options={'xatol': 0.1}).x)
ll, x = prof(hh, D, 'loo', 'gamma')
if ll < lm: hh, ll, x = 0.0, lm, pm
a = decomp(x, D, 'gamma', hh); LRm = max(0.0, 2 * (ll - lm))
rv['loo_2008'].append([lab_m[m], int((cell[I08] == m).sum()), r3(hh), r3(LRm), r4(0.5 * chi2.sf(LRm, 1)), r3(a[0, 6]), r3(a[3, 6]), r3(a[4, 6])])
print(f'leave-one-month-out {time.time() - T6:.1f}s')
# 6c. completed profiles: pooled gamma over 0-40, pooled split grid, continuous split searches by regime, regime gamma grids
rv['profiles'] = dict(pooled_gamma=[[h, r3(2 * (prof(h, DF, 'F', 'gamma')[0] - llF))] for h in [1.0, 2.0, 5.0, 10.0, 20.0, 40.0]],
pooled_split=[[h, r3(2 * (prof(h, DF, 'F', 'split')[0] - llF))] for h in [0.1, 0.2, 0.5, 0.9]])
for lab in reg:
D, l0 = reg[lab]['D'], reg[lab]['l0']
hh = float(minimize_scalar(lambda h: -prof(h, D, lab, 'split')[0], bounds=(0.0, 0.95), method='bounded', options={'xatol': 0.002}).x)
ll, x = prof(hh, D, lab, 'split'); dv = lambda h: 2 * (ll - prof(h, D, lab, 'split')[0]) - CRIT
lo = brentq(dv, 0.0, hh, xtol=0.002) if hh > 0.002 and dv(0.0) > 0 else 0.0; up = brentq(dv, hh + 0.002, 0.99, xtol=0.002) if dv(0.99) > 0 else None
a = decomp(x, D, 'split', hh); LRs = max(0.0, 2 * (ll - l0))
rv['profiles'][f'split_{lab}'] = dict(hat=r3(hh), LR=r3(LRs), p=r4(0.5 * chi2.sf(LRs, 1)), lo=r3(lo), up=None if up is None else r3(up), k7=[r3(a[0, 6]), r3(a[3, 6]), r3(a[4, 6])])
rv['profiles'][f'gamma_{lab}'] = [[h, r3(2 * (prof(h, D, lab, 'gamma')[0] - l0))] for h in ([0.5, 1, 2, 5, 10, 20, 40] if lab == '2008' else [0.1, 0.5, 1, 5, 40])]
C271 = chi2.ppf(0.90, 1) # one-sided 95% bound for a boundary parameter
l_th = prof(th_hat, DF, 'F', 'gamma')[0]; l_09 = prof(reg['2009_10']['hh'], reg['2009_10']['D'], '2009_10', 'gamma')[0]
rv['upper_2_71'] = dict(pooled_gamma=r3(brentq(lambda h: 2 * (l_th - prof(h, DF, 'F', 'gamma')[0]) - C271, th_hat + 1e-3, th_up, xtol=2e-3)),
gamma_2009_10=r3(brentq(lambda h: 2 * (l_09 - prof(h, reg['2009_10']['D'], '2009_10', 'gamma')[0]) - C271,
reg['2009_10']['hh'] + 1e-3, reg['2009_10']['gamma']['up'], xtol=2e-3)))
print(f'profiles {time.time() - T6:.1f}s')
# 6d. decomposition with one regime effect (2008 vs 2009-2010) in place of the 26 month effects
XR = np.hstack([(year > 2008).astype(float)[:, None], C]); DR = data(ALL, XR); pR, llR = fit(DR); VR = vcov(pR, DR, 'none', 0.0)
dR = decomp(pR, DR, 'none', 0.0); drR = np.array([decomp(z, DR, 'none', 0.0) for z in RNG2.multivariate_normal(pR, VR, size=1000, method='cholesky')])
with np.errstate(divide='ignore', invalid='ignore'):
shR, shRd = share(dR), share(drR)
rv['regime_effects'] = dict(loglik=round(llR, 2), **{f'k{k}': dict(total=est(dR, drR, 0, k), policy=est(dR, drR, 1, k), sel_obs=est(dR, drR, 2, k), within=est(dR, drR, 4, k),
comp_share=[r3(shR[k - 1])] + pct(shRd[:, k - 1])) for k in (4, 7)})
# 6e. Krinsky-Robb intervals for Table A1 (regime-specific gamma heterogeneity, conditional on the regime estimates)
drw = {r: RNG2.multivariate_normal(reg[r]['p_hat'], vcov(reg[r]['p_hat'], reg[r]['D'], 'gamma', reg[r]['hh']), size=500, method='cholesky') for r in reg}
drA = []
for j in range(500):
sv = [surv(drw[r][j], reg[r]['D']['X'], 'gamma', reg[r]['hh']) for r in reg]
drA.append(agg(np.vstack([s[0] for s in sv]), np.vstack([s[1] for s in sv]), np.r_[t[I08], t[I09]]))
drA = np.array(drA); rv['tableA1_ci'] = {f'k{k}': [pct(drA[:, j, k - 1]) for j in range(5)] for k in (4, 7, 10)}
print(f'revision analyses {time.time() - T6:.1f}s')
for r in reg.values():
for kk in ('D', 'p_hat', 'hh', 'lo', 'p0', 'l0'): r.pop(kk)
out = dict(n_clients=N, n_contacts=int(t.sum()), n_subscriptions=int(d.sum()), n_campaign_months=int(cell.max() + 1),
share_2008_at_risk_k1_k4_k7_k10=[r3((year[t >= k] == 2008).mean()) for k in (1, 4, 7, 10)], share_campaign_le3=r4((t <= 3).mean()),
n_client_covariates=int(C.shape[1]), loglik_theta0=round(llF, 2), loglik_month_only=round(llP, 2),
max_gap_model_vs_raw_k2_9=r4(np.max(np.abs(dF[0, 1:9] - raw[1:9]))), life_table=life, contact_effects=eff, decomposition=dec,
gamma_pooled=gam, split_pooled=spl, regimes=reg, pooled_decomp_regime_heterogeneity=regime_pooled, policy=policy,
H1_comp_share_k4_gt_half=H1, H1_refuted=bool(k4[2] < 0.5), H2_no_decline_k1_k3=H2, H3_structural_k7_negative_at_theta_up=H3,
H3_holds_within_2008=H3_2008, H4_decision_invariance=H4, hypothesis_supported=H1, revision=rv, runtime_seconds=r3(time.time() - T0))
print('RESULT_JSON: ' + json.dumps(out, separators=(',', ':')))
4.Analysis
Analysis complete: experiment succeeded; hypothesis supported; 24 key numbers.
Results and analysis
The run succeeded (41,188 clients, 105,754 contacts, 4,640 subscriptions; 88.3 s). All earlier results reproduce exactly; the revision analyses use a separate random stream. H1 is supported as pre-specified: with month effects, composition explains 0.777 (0.640-0.992) of the log decline to contact 4. With one 2008/2009-2010 effect in place of the 26 month effects, the share is 0.516 (0.418-0.674), leaving H1 inconclusive under that coding; but that model fits far worse (log-likelihood -15,943.12 against -15,236.87) and omits within-regime month variation, so 0.516 is a conservative bound rather than an equally supported estimate. Within cells, the 95% intervals for contacts 2 and 3 are -0.002 to 0.138 and -0.023 to 0.162: a decline of up to 2.3% is not excluded. After Holm adjustment over Table 3's 40 tests, the 2008 contact-2 rise (p = 0.518) and the pooled contact-4 fall (p = 0.720) are not significant. Proposition 2's test: pooled covariate contrasts do not attenuate (scale factors 0.920-1.109; LR = 2.635, p = 0.621), consistent with theta = 0.074 (Wald p = 0.549). In 2008 they attenuate (LR = 12.158, p = 0.016), and the pattern theta = 10.11 implies is not rejected (Wald p = 0.194). The month contrasts attenuate (0.313 and -0.373), but the client factors (1.012 and 1.331) are too imprecise to show or exclude the attenuation that pattern implies; what is rejected is equal attenuation of the month and client indices, which frailty acting through the whole index implies (LR = 9.327, p = 0.009), also without the two smallest months (p = 0.008). Leaving out each 2008 month keeps theta-hat between 6.466 and 23.451 (LR 4.31-17.045), but the structural effect at contact 7 then runs from -0.14 to +1.177. The completed profiles confirm the pooled gamma maximum at 0.074 over 0 to 40. The 2008 split-population maximum is interior (pi = 0.363, LR = 2.03, p = 0.077) and keeps a structural decline (-0.498). The stopping-rule results are unchanged; the real-time-feasible client-only rule earns 7.802% (6.267-9.614%) over the bank.
Key numbers
| Metric | Value | Interpretation |
|---|---|---|
| Analysed sample and run time | 41,188 clients; 105,754 contacts; 4,640 subscriptions; 26 campaign months; 88.282 s | Same data and filters; earlier results reproduce exactly |
| Composition share at contact 4 (month effects, theta = 0) | 0.777 [0.640, 0.992] | H1 supported as pre-specified |
| Composition share with one regime effect instead of 26 month effects, contacts 4 and 7 | 0.516 [0.418, 0.674]; 0.424 [0.337, 0.556] | H1 inconclusive under this coding, but a conservative bound: the model fits far worse and omits within-regime month variation |
| Log-likelihood: month effects vs regime effect | -15,236.87 vs -15,943.12 | Month effects strongly preferred by the data |
| Regime-effects decomposition at contact 4: policy; selection on x; within cell | -0.003 [-0.012, 0.011]; -0.274 [-0.296, -0.257]; -0.260 [-0.386, -0.135] | Coarser calendar shifts composition into the within-cell term |
| Within-cell contact effects at contacts 2 and 3, 95% CI (month + client) | [-0.002, 0.138]; [-0.023, 0.162] | No decline larger than about 0.2% and 2.3% of the first-contact hazard |
| Holm-adjusted p (40 tests): 2008 contact 2; pooled contact 4; pooled contact 5; pooled contact 8 | 0.5178; 0.7197; 0.0072; 0.0039 | Contact-2 rise and contact-4 fall not significant after adjustment |
| Pooled attenuation scales (month 2-3, month 4+, client 2-3, client 4+) with SE | 0.983 (0.038); 1.008 (0.057); 1.109 (0.094); 0.920 (0.140) | No attenuation pooled |
| Pooled attenuation tests: none vs scales; implied at theta 0.074 and 0.221 | LR 2.635 (p 0.6206); Wald 3.056 (p 0.5485); Wald 8.423 (p 0.0773) | Consistent with the small pooled theta |
| 2008 attenuation scales (month 2-3, month 4+, client 2-3, client 4+) with SE | 0.313 (0.271); -0.373 (0.658); 1.012 (0.494); 1.331 (1.012) | Month contrasts attenuate; client factors too imprecise to show or exclude the frailty-implied attenuation |
| 2008 implied scales at theta 10.11 and 1.89 | 0.722, 0.433, 0.780, 0.454; 0.934, 0.844, 0.967, 0.866 | Frailty implies near-equal attenuation of both indices |
| 2008 tests: none vs scales; common vs none; scales vs common; vs implied 10.11; vs implied 1.89 | LR 12.158 (p 0.0162); 2.831 (p 0.2428); 9.327 (p 0.0094); Wald 6.069 (p 0.1941); Wald 11.499 (p 0.0215) | Attenuation present; equal attenuation of the month and client indices, which frailty implies, rejected; the pattern implied at theta 10.11 not rejected |
| 2008 full covariate-by-bin interaction vs none | LR 85.475, 74 df, p 0.1704 | General non-proportionality not detected at this power |
| 2008 without the two smallest months: scales vs common; theta-hat (LR) | LR 9.659 (p 0.008); 9.962 (6.268) | Rejection of equal attenuation not driven by the small months |
| Leave-one-2008-month-out theta-hat range (LR range) | 6.466 to 23.451 (4.31 to 17.045) | theta-hat is not an artefact of one month |
| Leave-one-month-out structural effect at contact 7, range | -0.14 to 1.177 | The fatigue/selection split in 2008 is unstable |
| Pooled gamma profile LR at theta = 1, 5, 40 | -91.698; -768.173; -2,775.88 | Global maximum at 0.074 over 0 to 40 |
| 2008 split population (continuous search): pi-hat, LR, p, interval; structural k7 | 0.363; 2.03; 0.0771; [0, 0.503]; -0.498 | Interior maximum, not significant, retains fatigue |
| 2009-10 split population: pi-hat, LR, p, interval; structural k7 | 0.016; 4.394; 0.018; [below 0.002, 0.043]; -0.296 | Small never-convert share; fatigue retained |
| One-sided 95% upper bounds (LR 2.71): pooled gamma; 2009-10 gamma | 0.196; 0.195 | Tighter than the 3.84 limits 0.221 and 0.222 |
| Regime-heterogeneity decomposition, structural effect at contact 7, KR 95% CI | [-0.424, 0.153] | Conditional on regime estimates; includes zero |
| Real-time-feasible client-only rule at rho = 0.02: vs bank; vs uniform cap | 7.802% [6.267, 9.614]; 1.632 [0.932, 2.376] per 1,000 | Feasible benchmark beside the 10.358% upper bound |
| Client-and-month rule at rho = 0.02: vs bank; vs uniform cap | 10.358% [8.554, 12.552]; 3.198 [2.362, 4.277] per 1,000 | Upper bound using same-month outcomes |
| Key pooled client coefficients (SE): previous success; telephone channel | 1.056 (0.057); -0.570 (0.058) | Largest precisely estimated client effects on the cloglog scale |
Caveats
- H1 holds with 26 month effects but is inconclusive with a single regime effect, fitted because month is recorded at the last contact; that model fits far worse and omits within-regime month variation, so its share is a conservative bound rather than an equally supported estimate.
- The 2008 attenuation test has wide standard errors on the client contrasts; it rejects equal attenuation but cannot reject the overall pattern implied by theta = 10.11.
- Leaving out any 2008 month keeps theta-hat large, but the implied structural effect at contact 7 changes sign across subsamples, so the 2008 fatigue/selection split is unstable and selection within 2008 cannot be excluded.
- In 2008 the indicator for two or more previous contacts is nearly empty and its coefficient diverges (the optimizer stops at -13.781), a separation artefact, so it is not estimable.
- Krinsky-Robb intervals for the regime-heterogeneity decomposition condition on the regime estimates of theta; the theta uncertainty is shown by the profiles.
- Profile limits inverting at 3.84 are conservative for parameters on the boundary; the 2.71 one-sided bounds are reported for the pooled and 2009-10 gamma variances.
- Assumption 1 is untestable; call-backs scheduled for interested clients would mimic wear-in, and truncation cannot value calling more than the bank did.
- Subscription is taken at the last contact; the data do not say whether a contact is an attempt or a conversation, and duration is recorded only for the last call.
- Month effects in the client-and-month rule use other clients of the same month, an upper bound for real-time use; the client-only rule is the feasible benchmark.
- The cost ratio is not observed; one bank, one product, 2008-2010; no identifier links clients across campaigns.
5.Paper writing
Manuscript drafted: “Contact Fatigue or Dynamic Selection? Decomposing the Declining Conversion Hazard in Repeated Telemarketing and Designing Client-Specific Stopping Rules” (≈ 9150 words).
≈ 9150 words · AI-drafted manuscript · not peer-reviewed
Contact Fatigue or Dynamic Selection? Decomposing the Declining Conversion Hazard in Repeated Telemarketing and Designing Client-Specific Stopping Rules
Abstract
Managers read the falling per-contact conversion rate in repeated telemarketing as contact fatigue; it may instead reflect dynamic selection, as receptive clients convert early, or the firm's choice of whom to keep calling. We model contacts to subscription as a grouped-time proportional hazard with campaign-month effects, client covariates and gamma or split-population heterogeneity, decompose the decline exactly into four parts, and characterize the profit-maximizing stopping rule. In a Portuguese bank's 2008–2010 campaigns (41,188 clients, 105,754 contacts), the hazard falls from 0.0558 at the first contact to 0.0110 at the tenth. With month effects, composition explains 77.7% (95% interval 64.0–99.2%) of the decline to the fourth contact; a single 2008/2009–2010 effect, which fits far worse (log-likelihood −15,943.12 against −15,236.87) and omits within-regime month variation, gives a conservative bound of 51.6% (41.8–67.4%). If continuation decisions ignore unobserved receptiveness, the second and third contacts are not detectably less productive than the first. Later contacts show a structural decline in every specification except a 2008 gamma model, whose implied equal attenuation of month and client contrasts is rejected (p = 0.009), though not its overall pattern (p = 0.194); the client contrasts are too imprecise to show or exclude attenuation. At a cost-to-margin ratio of 2%, rules evaluated exactly on logged calls raise profit over the bank's practice by 7.8% (6.3–9.6%) with client covariates only, feasible in real time, and by up to 10.4% (8.6–12.6%) adding same-month outcomes. Rules built under different heterogeneity assumptions differ by at most 0.22 conversion-equivalents per 1,000 clients.
Keywords: telemarketing; contact policy; optimal stopping; duration dependence; unobserved heterogeneity; dynamic selection; proportional hazards
1. Introduction
How many times should a firm call a prospect before giving up? Campaigns usually answer with a fixed cap, justified by the falling share of calls that convert: in the Portuguese bank studied here, 0.0558 at the first contact, 0.0326 at the fourth and 0.0110 at the tenth (Table 3). The tempting reading is contact fatigue: each call makes the client less receptive, so persistence destroys value and imposes costs on consumers (Goh et al. 2015, Ferecatu et al. 2024).
Two other explanations fit the curve. Under dynamic selection, receptive clients convert early, so the pool still being called becomes less receptive although no client changes (Vaupel and Yashin 1985, Fader and Hardie 2007). Under policy composition, the firm decides whom to call again; if it calls clients in poor campaigns longer, late contacts come from low-propensity months. Fatigue argues for caps and spacing, selection makes non-response informative (Schmittlein et al. 1987), and composition means the curve describes no client. Survey research faces the same stopping problem for call attempts (Elliott et al. 2000, Groves and Heeringa 2006).
We ask how much of the decline is composition, whether the within-client change is fatigue or unobserved selection, and which stopping rule the bank should use.
Our contributions are threefold. First, an exact decomposition of the log decline into the firm's continuation policy, selection on observables, selection on unobservables and the structural contact effect, with Krinsky–Robb intervals. Second, known results put to work on identification: if continuation decisions ignore unobserved receptiveness, selection cannot raise a conditional hazard, and gamma heterogeneity attenuates all covariate contrasts across contacts by a common factor. Used as a specification test, the latter indicates that the one estimate attributing the later decline to selection absorbs month-specific contact profiles. Third, the optimal stopping rule depends only on the conditional hazard path, and cross-fitted rules evaluated exactly by truncating the logged contact sequences earn almost the same profit under competing heterogeneity assumptions.
2. Literature Review
True versus spurious duration dependence. Neglected heterogeneity biases duration dependence downward (Lancaster 1979); Heckman and Borjas (1980) separate state dependence from heterogeneity, Vaupel and Yashin (1985) show how selection makes population hazards fall while individual hazards do not, and Elbers and Ridder (1982) identify the mixed proportional hazard (MPH) model from covariate variation. Identification is fragile: estimates depend on the heterogeneity distribution (Heckman and Singer 1984), the MPH information matrix is singular unless the baseline is restricted near zero (Ridder and Woutersen 2003), and nonparametric components are badly biased (Baker and Melino 2000). Grouped durations use the complementary log-log model (Prentice and Gloeckler 1978) on person-period data (Allison 1982, Jenkins 1995), its gamma extension (Meyer 1990, Han and Hausman 1990) and split-population models (Schmidt and Witte 1989, Sinha and Chandrashekaran 1992); Van den Berg (2001) reviews identification. In marketing, proportional hazards with heterogeneity model purchase timing (Jain and Vilcassim 1991, Seetharaman and Chintagunta 2003), sorting explains rising retention (Fader and Hardie 2007), inactivity reveals whether a customer is active (Schmittlein et al. 1987), and state dependence survives heterogeneity controls (Dubé et al. 2010).
Repetition and fatigue. Advertising shows wear-in, then wear-out (Calder and Sternthal 1980, Pechmann and Stewart 1988, Naik et al. 1998, Braun and Moe 2013, Chae et al. 2019). Charity mailings have dynamic effects on giving (Van Diepen et al. 2009a) and irritate without reducing it (Van Diepen et al. 2009b); marketing intensity hastens email opt-out (Kumar et al. 2014), solicitations raise panel attrition (Ferecatu et al. 2024), responsiveness declines over repeated interventions (Ebbes et al. 2026), and telemarketing opt-out imposes externalities on other consumers (Goh et al. 2015).
Contact policy and stopping. Contact policy is a classic Management Science problem: call planning (Lodish 1971), profit-based targeting (Bult and Wansbeek 1995), catalog mailing as a dynamic program (Bitran and Mondschein 1996, Gönül and Shi 1998, Simester et al. 2006) and panel invitations (Neslin et al. 2009). Stopping policies can be learned interpretably (Ciocan and Mišić 2022) or within a sales call (Manzoor et al. 2025). Targeting policies are evaluated by profit on experimental or logged data (Simester et al. 2020a, 2020b, Athey and Wager 2021, Hitsch et al. 2024); Gans et al. (2003) review call centres; the bank data have been a classification benchmark (Moro et al. 2014a). We know of no study that decomposes the contact-number gradient, tests the heterogeneity it attributes, or asks when the decomposition changes the stopping rule.
3. Model
3.1. Setting
A firm calls each client on a campaign list. Client has observed characteristics , the campaign month and client covariates , and unobserved receptiveness with distribution . At contact , a client who has not yet subscribed subscribes with probability
the grouped-time proportional hazard (Prentice and Gloeckler 1978). The structural contact effect is wear-in if positive and fatigue if negative; pure dynamic selection means with . We use gamma heterogeneity with mean one and variance , and split-population heterogeneity with . With and , survival through contacts is , which equals (gamma), (split population) and (). The conditional hazard is (Table 1).
Table 1. Notation
| Symbol | Meaning |
|---|---|
| ; | client; contact number (, and pooled) |
| campaign month and client covariates | |
| ; | unobserved receptiveness; its distribution |
| ; | gamma heterogeneity variance; never-convert share |
| ; ; | baseline of contact ; covariate coefficients; |
| cumulative baseline | |
| ; | structural hazard; conditional hazard given and no earlier subscription |
| probability of no subscription in the first contacts | |
| ; | clients called at least times; aggregate hazard at contact |
| ; ; | hazard of the pool had the firm dropped no one; with the contact-1 mix of ; with the contact-1 mix of and |
| variance of among clients in with characteristics | |
| ; | value of placing contact ; value of the option to continue at |
| ; | contacts in the campaign; subscription indicator |
| cost of one contact relative to the margin of one subscription | |
| number of contacts placed (decision variable) |
Notes. Profits are measured in conversion-equivalents: subscriptions minus times contacts.
3.2. Assumptions, Decision and Objective
Assumption 1 (Conditionally independent continuation). Given and , the firm's decision to place contact for a client who has not subscribed does not depend on .
Assumption 2 (Mixed proportional hazards). is independent of , and the contact effects are common to all clients.
Assumption 3 (No interference). A client's outcomes do not depend on how other clients are called.
Assumption 1 lets observed hazards be read as conditional hazards; the decomposition imposes Assumption 2's common contact effects (one in equation 1), while independence of and is needed only to split the within-cell change; Assumption 3 is used to evaluate policies. The first contact is given; after each failure the firm, knowing and the number of failures, decides whether to call again. Placing contacts yields expected profit
3.3. Decomposition and Structural Results
Let be the mean of over ; ; ; and . All four equal at , so
A cell is a set of clients with the same . With the third term vanishes and the fourth is the within-cell change .
Proposition 1 (Selection cannot raise a conditional hazard). Under Assumption 1, for every and ,
(i) A rise, , implies wear-in on average among the clients still called. (ii) Under pure selection, is non-increasing in . (iii) A within-cell decline overstates the average structural decline. The result holds cell by cell, for any , without Assumption 2.
Proposition 2 (Attenuation identifies heterogeneity). Under Assumption 2 with gamma heterogeneity, . Let . The elasticity of with respect to equals 1 for every if ; if it equals the average of over , which is below one and strictly decreasing in .
Because every covariate enters through , all contrasts shrink across contacts by approximately the same -dependent factor, an implication Section 5.4 tests; heterogeneity is well identified where hazards are high and weakly identified where they are low.
Proposition 3 (Stopping depends only on the conditional hazard path). Let , and . Placing contact if and only if for all maximizes (2) over . (a) The rule depends on only through , so structures that split a decline differently between fatigue and selection but share conditional hazards prescribe the same rule. (b) If is non-increasing for , the one-step look-ahead rule, call again if and only if , is optimal.
Corollary 1 (Exact evaluation by truncation). Suppose a subscription occurs at the client's last logged contact and Assumption 3 holds. For any rule , calling client times earns , computable from the logged data without a response model and without Assumptions 1 or 2. Rules that call more than the firm did cannot be evaluated this way.
Proposition 1 is the discrete-time form of the frailty selection identity behind Vaupel and Yashin (1985). Proposition 3(b) is the monotone case of optimal stopping, which rests on the 1961 martingale system theorem reprinted as Chow and Robbins (1985); the standard treatment, Chow, Robbins and Siegmund's 1971 monograph, is outside our retrieved sources. Corollary 1 avoids the randomization or reweighting that policy evaluation on logged data usually needs (Simester et al. 2020a, 2020b, Athey and Wager 2021). Proofs are in the Appendix. By Proposition 3(a), the fatigue/selection split matters only for counterfactuals that change the conditional hazard path, such as later re-contact or call spacing.
4. Analysis
Estimation. Client contributes to the likelihood, maximized by L-BFGS-B with analytic gradients; at this is the person-period complementary log-log model (Allison 1982, Jenkins 1995). Contacts 1–10 have their own baseline, contacts 11+ share one, standard errors use a heteroskedasticity-robust sandwich with one likelihood term per client, and Table 3's 40 tests are also Holm-adjusted.
Intervals. Decomposition intervals come from 1,000 parameter draws from the normal distribution with the sandwich covariance (Krinsky and Robb 1986), so share intervals reflect numerator and denominator jointly; the regime-heterogeneity decomposition uses 500 draws conditional on the regime estimates of . Raw log ratios use the delta method and hazards Wilson intervals.
Heterogeneity. We profile the likelihood over (or ) by bounded scalar search, test against (Self and Liang 1987), and invert the likelihood ratio at 3.84 (Murphy and van der Vaart 2000), adding one-sided bounds at 2.71 because 3.84 is conservative on the boundary. Search ranges for maxima and limits: pooled gamma [0, 0.3] and [0, 1]; pooled split [0, 0.2] and [0, 0.3]; regime gamma [0, 40] and [0, 100]; regime split [0, 0.95] and [0, 0.99]; completed profiles reach and . Each regime (2008; 2009–2010) has its own month effects, contact effects and coefficients; the leave-one-month-out analysis re-maximizes on [0, 40] with each 2008 month dropped.
Attenuation test. The person-period model gets scale factors at contacts 2–3 and 4+ (contact 1 normalized to 1) for the month index and the client index , either separate or common to both. Refitting it to each gamma model's conditional hazards over the observed risk sets gives the attenuation that model implies. A full covariate-by-bin interaction tests proportionality generally.
Algorithm 1 (numerical study). (1) Split clients into five folds (seed 20261024). (2) Fit each model on four folds; compute held-out clients' conditional hazard paths for contacts 1–20. (3) For each , set by backward induction (Proposition 3). (4) Evaluate by truncation (Corollary 1). (5) Choose uniform and month-specific caps in on the training folds by truncation profit. (6) Bootstrap held-out clients (500 Poisson reweightings).
Hypotheses (fixed before estimation). H1: at contact 4 the lower 95% limit of the composition share, at , exceeds 0.5 (conservative, since attributes most to fatigue). H2: within-cell effects at contacts 2 and 3 are not significantly negative. H3: at the upper 95% limit of the pooled gamma variance, the structural effect at contact 7 is negative. H4: the client-and-month rule beats the uniform cap at some , and rules under different heterogeneity assumptions earn the same profit within sampling error.
5. Numerical Study
5.1. Data and Design
We analyse the bank-additional-full file of the Bank Marketing dataset (UCI Machine Learning Repository, dataset 222, licence CC BY 4.0; Moro et al. 2014b), donated by S. Moro, P. Cortez and P. Rita and retrieved on 24 October 2026: 41,188 client records from a Portuguese retail bank's telemarketing campaigns for term deposits, ordered by date from May 2008 to November 2010. Agents called clients from lists and offered the deposit to clients who phoned for other reasons (Moro et al. 2014a); the file does not record call direction. campaign, the number of contacts with the client in this campaign including the last, is . The documentation does not say whether unanswered attempts count, and duration is recorded for the last contact only. A contact is thus a recorded attempt that may or may not have reached the client, so a contact effect mixes receptiveness with reachability, which weakens a fatigue reading, and prices an average attempt. For illustration only (neither figure is observed), a margin of €100 per deposit and €1–3 per attempt give –. y is , and a subscription is taken to occur at the last contact. Month and channel are recorded for the last contact only, so a spell spanning two months is assigned to the later one. The 26 campaign months are the distinct pairs of month and consumer price index, ordered by first appearance. The 31 client covariates are age in decades from 40 and its square; job, marital status, education, housing loan, contact channel and previous-campaign outcome as categories ('unknown' kept as a level); and indicators for a default status other than 'no', a personal loan and two or more previous contacts. We exclude call duration, day of the week and the macroeconomic indicators. No records are dropped (Table 2).
Table 2. Computational design
| Element | Setting |
|---|---|
| Sample | 41,188 clients, 105,754 contacts, 4,640 subscriptions; 2008: 27,690 clients, 1,339 subscriptions; 2009–2010: 13,498 clients, 3,301 subscriptions; 31 client covariates |
| Data and instances | One dataset; three cost ratios; by-month summaries over the 26 campaign months |
| Cost-to-margin ratio | 0.01, 0.02, 0.03 |
| Decision | contacts 2–20 (first contact given) |
| Cross-fitting | 5 folds, seed 20261024 |
| Benchmarks | bank's practice; best uniform cap; best month-specific cap; one-step look-ahead |
| Backward-induction rules | ; ; regime-specific ( in 2008, in 2009–2010); client covariates only |
| Inference | 500 Poisson-bootstrap replicates; 1,000 Krinsky–Robb draws (500 for Table A1) |
| Run time | 88.3 s for the whole analysis |
Notes. is the pooled upper 95% limit; regime values are regime estimates (Table 6). The manuscript reports results in tables only; figures are not part of this record, and Tables 3, 4 and 7 contain the hazard and profile values a figure would show.
5.2. The Hazard Profile
The pool shifts toward the low-conversion 2008 campaigns as contacts accumulate, from 0.672 of those at risk at contact 1 to 0.931 at contact 10 (Table 3). Within cells, the 95% intervals for the second and third contacts are [−0.002, 0.138] and [−0.023, 0.162], excluding declines larger than 0.2% and 2.3% of the first-contact hazard; H2 holds, though a small decline at contact 3 is not ruled out. This requires Assumption 1, since call-backs scheduled for clients who sounded interested would produce the same pattern. The within-2008 rise at contact 2 () and the pooled fall at contact 4 () do not survive Holm adjustment (adjusted and 0.720). The declines at contacts 5, 7, 8 and 11+ do (adjusted ).
Table 3. Aggregate hazard and within-cell contact effects by contact number
| At risk | 2008 share | Subscr. | [95% CI] | Month only | Month + client [95% CI] | Holm | 2008 | 2009–10 | |
|---|---|---|---|---|---|---|---|---|---|
| 1 | 41,188 | 0.672 | 2,300 | 0.0558 [0.0537, 0.0581] | 0 | 0 | – | 0 | 0 |
| 2 | 23,546 | – | 1,211 | 0.0514 [0.0487, 0.0543] | 0.065 (0.036) | 0.068 [−0.002, 0.138] | 0.7517 | 0.150 (0.070) | 0.041 (0.042) |
| 3 | 12,976 | – | 574 | 0.0442 [0.0408, 0.0479] | 0.040 (0.047) | 0.070 [−0.023, 0.162] | 1.0000 | 0.124 (0.085) | 0.052 (0.057) |
| 4 | 7,635 | 0.759 | 249 | 0.0326 [0.0289, 0.0368] | −0.178 (0.067) | −0.134 [−0.266, −0.001] | 0.7197 | −0.098 (0.114) | −0.141 (0.085) |
| 5 | 4,984 | – | 120 | 0.0241 [0.0202, 0.0287] | −0.400 (0.094) | −0.351 [−0.536, −0.165] | 0.0072 | −0.288 (0.147) | −0.377 (0.124) |
| 6 | 3,385 | – | 75 | 0.0222 [0.0177, 0.0277] | −0.379 (0.118) | −0.324 [−0.555, −0.094] | 0.1512 | −0.145 (0.163) | −0.482 (0.172) |
| 7 | 2,406 | 0.851 | 38 | 0.0158 [0.0115, 0.0216] | −0.604 (0.163) | −0.549 [−0.871, −0.228] | 0.0262 | −0.617 (0.234) | −0.439 (0.233) |
| 8 | 1,777 | – | 17 | 0.0096 [0.0060, 0.0153] | −1.012 (0.245) | −0.944 [−1.424, −0.465] | 0.0039 | −0.802 (0.292) | −1.153 (0.450) |
| 9 | 1,377 | – | 17 | 0.0123 [0.0077, 0.0197] | −0.697 (0.247) | −0.623 [−1.105, −0.141] | 0.2247 | −0.290 (0.254) | −2.236 (1.005) |
| 10 | 1,094 | 0.931 | 12 | 0.0110 [0.0063, 0.0191] | −0.776 (0.289) | −0.721 [−1.289, −0.153] | 0.2448 | −0.772 (0.357) | −0.487 (0.501) |
| 11+ | 5,386 | – | 27 | 0.0050 [0.0034, 0.0073] | −1.448 (0.200) | −1.405 [−1.798, −1.012] | 0.0000 | −1.307 (0.214) | −1.670 (0.600) |
Notes. At risk: clients contacted at least times (11+: contact-periods beyond the tenth); 2008 share: fraction of them from 2008 campaigns. Effects: at , with robust SE in parentheses or 95% CI in brackets; "Month only" omits the client covariates; regime columns use that regime alone. Holm adjusts the month + client test over all 40 tests in the table; the 2008 contact-2 test has unadjusted and adjusted . Log-likelihoods −15,532.45 (month only) and −15,236.87 (month + client); the latter reproduces raw log ratios within 0.0069 for .
5.3. Decomposition
At contact 4 the log decline of −0.538 splits into −0.027 from the bank's continuation policy, −0.391 from selection on observables and −0.120 within cells (Table 4). Composition therefore accounts for 0.777 (0.640–0.992), so H1 is supported, and month effects alone give 0.697 (0.572–0.891). By contact 7 the within-cell term (−0.502) rivals selection on observables (−0.576). These late rows rest on contact effects common to all months, which Table 3's regime columns question (contact 9: −0.290 in 2008, −2.236 in 2009–10). Because month is a last-contact attribute, Table 5 replaces the 26 month effects by one regime effect. The composition share then falls to 0.516 (0.418–0.674) at contact 4 and 0.424 (0.337–0.556) at contact 7, leaving H1 inconclusive under this coarser coding; but this model fits far worse (log-likelihood −15,943.12 against −15,236.87) and omits within-regime month variation, so 0.516 is a conservative bound rather than an equally supported estimate.
Table 4. Decomposition of with and month effects
| Raw | Policy | Selection on | Within cell | Composition share | Month-only share | |
|---|---|---|---|---|---|---|
| 2 | −0.082 [−0.150, −0.015] | 0.013 [0.006, 0.024] | −0.151 [−0.165, −0.143] | 0.061 [0.001, 0.123] | – | – |
| 3 | −0.233 [−0.322, −0.144] | −0.012 [−0.021, 0.007] | −0.279 [−0.301, −0.266] | 0.062 [−0.023, 0.139] | – | – |
| 4 | −0.538 [−0.666, −0.409] | −0.027 [−0.039, −0.007] | −0.391 [−0.418, −0.373] | −0.120 [−0.237, −0.003] | 0.777 [0.640, 0.992] | 0.697 [0.572, 0.891] |
| 5 | −0.841 [−1.022, −0.660] | −0.052 [−0.066, −0.028] | −0.471 [−0.503, −0.449] | −0.318 [−0.491, −0.146] | 0.622 [0.514, 0.780] | 0.562 [0.468, 0.725] |
| 6 | −0.924 [−1.152, −0.697] | −0.104 [−0.120, −0.076] | −0.520 [−0.557, −0.494] | −0.294 [−0.508, −0.076] | 0.680 [0.556, 0.892] | 0.621 [0.505, 0.805] |
| 7 | −1.263 [−1.581, −0.945] | −0.180 [−0.202, −0.150] | −0.576 [−0.616, −0.545] | −0.502 [−0.804, −0.204] | 0.601 [0.485, 0.784] | 0.553 [0.449, 0.730] |
| 8 | −1.764 [−2.239, −1.289] | −0.266 [−0.289, −0.230] | −0.623 [−0.669, −0.588] | −0.875 [−1.336, −0.426] | 0.504 [0.400, 0.672] | 0.462 [0.371, 0.623] |
| 9 | −1.509 [−1.983, −1.035] | −0.310 [−0.336, −0.272] | −0.632 [−0.677, −0.593] | −0.571 [−1.015, −0.142] | 0.623 [0.483, 0.866] | 0.573 [0.443, 0.802] |
| 10 | −1.627 [−2.192, −1.063] | −0.297 [−0.326, −0.253] | −0.663 [−0.714, −0.621] | −0.663 [−1.194, −0.152] | 0.592 [0.450, 0.855] | 0.553 [0.417, 0.808] |
Notes. Raw: life-table log ratio, delta-method 95% CI, treating and as independent (discrete-hazard estimators on nested risk sets are asymptotically uncorrelated). Other columns: estimate [2.5th, 97.5th percentiles of 1,000 Krinsky–Robb draws]; the three terms sum to the model-implied log ratio. Composition share within/total, omitted where the within-cell term is positive; month-only share uses the model without client covariates.
Table 5. Decomposition with one regime effect in place of the 26 month effects ()
| Model total | Policy | Selection on | Within cell | Composition share | Share with month effects (Table 4) | |
|---|---|---|---|---|---|---|
| 4 | −0.537 [−0.665, −0.412] | −0.003 [−0.012, 0.011] | −0.274 [−0.296, −0.257] | −0.260 [−0.386, −0.135] | 0.516 [0.418, 0.674] | 0.777 [0.640, 0.992] |
| 7 | −1.264 [−1.584, −0.942] | −0.138 [−0.151, −0.119] | −0.398 [−0.428, −0.374] | −0.728 [−1.040, −0.424] | 0.424 [0.337, 0.556] | 0.601 [0.485, 0.784] |
Notes. One indicator for 2009–2010 plus the 31 client covariates; 1,000 Krinsky–Robb draws. Log-likelihood −15,943.12, against −15,236.87 with month effects; the model omits within-regime month variation, so its composition shares are conservative bounds.
5.4. Fatigue or Unobserved Selection?
Pooled (Tables 6 and 7), gamma heterogeneity is tightly bounded (, LR , ; interval 0 to 0.221, one-sided bound 0.196), and the profile keeps falling up to . The split-population family fits better (, LR , ). Over both intervals the structural effect at contact 7 lies between −0.502 and −0.377 of a decline near −1.25, so under common heterogeneity fatigue drives most of the within-cell decline and H3 holds. On its own, 2009–2010 also bounds heterogeneity and keeps a structural decline at contact 7 (−0.391 to −0.184). In 2008 the gamma profile peaks at (LR , ; interval 1.89 to 24.97), where the within-cell decline is all selection. The split-population profile instead peaks inside its range (, LR , ) and keeps a structural decline (−0.498). Both see a large mass of very unreceptive 2008 clients but disagree on fatigue.
Table 6. Profile-likelihood analysis of unobserved heterogeneity
| Sample, family | Parameter | LR vs 0 () | Total at | Unobserved at | Structural at | ||
|---|---|---|---|---|---|---|---|
| Pooled, none | 0 | – | −0.134 | −0.549 | −1.258 | 0 | −0.502 |
| Pooled, gamma | 1.182 (0.138) | −0.096 | −0.504 | −1.254 | −0.081 | −0.456 | |
| Pooled, gamma | upper 0.221 | – | −0.028 | −0.426 | −1.251 | −0.218 | −0.377 |
| Pooled, gamma | 0.5 (outside) | −24.990 | 0.083 | −0.301 | −1.249 | −0.424 | −0.257 |
| Pooled, split | 4.575 (0.016) | −0.108 | −0.499 | −1.249 | −0.115 | −0.453 | |
| Pooled, split | upper 0.045 | – | −0.082 | −0.461 | −1.248 | −0.218 | −0.417 |
| 2008, gamma | 7.124 (0.0038) | 0.535 | 0.430 | −0.747 | −1.079 | 0.350 | |
| 2008, gamma | lower 1.89 | – | 0.024 | −0.413 | −0.744 | −0.313 | −0.378 |
| 2008, gamma | upper 24.97 | – | 1.411 | 1.895 | −0.745 | −1.873 | 1.131 |
| 2008, split | 2.030 (0.077) | – | – | −0.742 | −0.198 | −0.498 | |
| 2009–10, none | 0 | – | −0.141 | −0.439 | −0.901 | 0 | −0.391 |
| 2009–10, gamma | 0.412 (0.261) | −0.105 | −0.383 | −0.896 | −0.065 | −0.338 | |
| 2009–10, gamma | upper 0.222 | – | 0.012 | −0.216 | −0.888 | −0.260 | −0.184 |
| 2009–10, split | 4.394 (0.018) | – | – | −0.881 | −0.115 | −0.296 |
Notes. LR: twice the log-likelihood gain over no heterogeneity in the same sample; from ; limits invert the LR at 3.84 (pooled gamma 0 to 0.221, one-sided 2.71 bound 0.196; pooled split below 0.002 to 0.045; 2008 gamma 1.89 to 24.97; 2008 split 0 to 0.503; 2009–10 gamma 0 to 0.222, one-sided bound 0.195; 2009–10 split below 0.002 to 0.043). Totals are each sample's model-implied ; "–" marks values not computed for that row.
Table 7. Completed profile likelihoods (LR against no heterogeneity in the same sample)
| Profile | Values (parameter: LR) |
|---|---|
| Pooled, gamma | 1: −91.698; 2: −261.811; 5: −768.173; 10: −1,401.46; 20: −2,147.178; 40: −2,775.88 |
| Pooled, split | 0.1: −24.696; 0.2: −122.846; 0.5: −964.073; 0.9: −7,638.563 |
| 2008, gamma | 0.5: 1.154; 1: 2.048; 2: 3.412; 5: 5.889; 10: 7.124; 20: 5.004; 40: −1.870 |
| 2009–10, gamma | 0.1: 0.068; 0.5: −20.448; 1: −68.761; 5: −511.963; 40: −1,356.781 |
Notes. Profile maxima and limits are in Table 6. These rows replace the figure a profile plot would give.
Is the 2008 estimate heterogeneity? Table 8 tests the common attenuation that Proposition 2 implies. Pooled, neither index attenuates (LR , ), as implies. In 2008 the contrasts attenuate (LR , ), and the overall pattern is not rejected against the one implies (Wald ). The month contrasts attenuate (factors 0.313 and −0.373), but the client factors (1.012, SE 0.494; 1.331, SE 1.012) are too imprecise to show or exclude the attenuation implies (0.780 and 0.454). Equal attenuation of the month and client indices is rejected (LR , ), also without the two smallest months (); a full covariate-by-bin interaction is not significant (). Leaving out each 2008 month (Table 9) keeps between 6.466 and 23.451, so the estimate is not an artefact of one month. The structural effect at contact 7, however, then runs from −0.14 to +1.177.
We therefore read the 2008 gamma estimate as absorbing month-specific contact profiles, a non-proportionality that the MPH model can express only as heterogeneity, rather than as evidence of selection among clients. Every other specification keeps a structural decline at contact 7, so the evidence favours fatigue at later contacts, with campaign-specific profiles in 2008; selection within 2008 cannot be excluded.
Table 8. Proposition 2 test: attenuation of covariate contrasts across contact bins ( model; contact 1 normalized to 1)
| Scale factor | Pooled, observed (SE) | Implied, | Implied, | 2008, observed (SE) | Implied, | Implied, |
|---|---|---|---|---|---|---|
| Month index, contacts 2–3 | 0.983 (0.038) | 0.990 | 0.971 | 0.313 (0.271) | 0.722 | 0.934 |
| Month index, contacts 4+ | 1.008 (0.057) | 0.973 | 0.923 | −0.373 (0.658) | 0.433 | 0.844 |
| Client index, contacts 2–3 | 1.109 (0.094) | 0.974 | 0.927 | 1.012 (0.494) | 0.780 | 0.967 |
| Client index, contacts 4+ | 0.920 (0.140) | 0.949 | 0.870 | 1.331 (1.012) | 0.454 | 0.866 |
| Common factor, contacts 2–3 | 1.014 (0.026) | 0.986 | 0.959 | 0.695 (0.214) | 0.764 | 0.959 |
| Common factor, contacts 4+ | 0.991 (0.041) | 0.967 | 0.910 | 0.818 (0.301) | 0.449 | 0.861 |
| Wald test, observed = implied (4 df) | – | 3.056 (0.5485) | 8.423 (0.0773) | – | 6.069 (0.1941) | 11.499 (0.0215) |
| LR: separate factors vs none (4 df) | 2.635 (0.6206) | 12.158 (0.0162) | ||||
| LR: common factor vs none (2 df) | 0.487 (0.784) | 2.831 (0.2428) | ||||
| LR: separate vs common factor (2 df) | 2.149 (0.3415) | 9.327 (0.0094) | ||||
| LR: full covariate-by-bin interaction | – | 85.475 (0.1704), 74 df |
Notes. -values in parentheses. Implied factors are those the same model recovers when fitted to the conditional hazards of the fitted gamma model over the observed risk sets. Without October and December 2008 (27,613 clients), the 2008 factors are 0.318 (0.267), −0.394 (0.626), 0.971 (0.432) and 1.31 (0.894); common 0.700 (0.208) and 0.856 (0.287); LR separate vs none 12.457 (0.0143), common vs none 2.798 (0.2469), separate vs common 9.659 (0.008); (LR ).
Table 9. 2008 gamma profile leaving out one campaign month
| Month left out | Clients left out | LR vs 0 () | Total at | Unobserved at | Structural at | |
|---|---|---|---|---|---|---|
| May 2008 | 7,763 | 23.256 | 13.176 (0.0001) | −0.661 | −1.827 | 1.177 |
| Jun 2008 | 4,374 | 8.273 | 6.535 (0.0053) | −0.782 | −0.965 | 0.202 |
| Jul 2008 | 6,685 | 23.451 | 17.045 (0.0000) | −0.880 | −1.811 | 0.936 |
| Aug 2008 | 5,175 | 6.466 | 4.310 (0.0189) | −0.988 | −0.812 | −0.140 |
| Oct 2008 | 67 | 10.150 | 6.483 (0.0054) | −0.668 | −1.028 | 0.376 |
| Nov 2008 | 3,616 | 9.118 | 7.207 (0.0036) | −0.601 | −0.970 | 0.375 |
| Dec 2008 | 10 | 9.911 | 6.921 (0.0043) | −0.745 | −1.062 | 0.335 |
Notes. re-maximized on [0, 40]; from . With all 2008 months: , LR , structural effect 0.350 (Table 6).
5.5. Stopping Rules
At (Table 10) the bank earns 61.30 conversion-equivalents per 1,000 clients and the best uniform cap (six contacts in every fold) 64.45. The real-time-feasible rule, which uses client covariates only, earns 66.08, 7.80% (6.27–9.61%) more than the bank and 1.63 (0.93–2.38) more than the cap. The client-and-month rule earns 67.65, 10.36% (8.55–12.55%) more than the bank and 3.20 (2.36–4.28) more than the cap, with 64.56% of the bank's contacts and 89.48% of its subscriptions. Its month effects use other clients of the same month, so this is an upper bound for real-time use; a month-specific cap delivers 2.16 of it. The one-step look-ahead rule matches backward induction, and across months the gain over the cap averages 2.80 (range −1.37 to 16.51). Rules built at and with regime-specific heterogeneity earn 0.168 (0.016–0.329) and 0.223 (−0.268 to 0.706) less than the rule. The first difference refutes H4's strict invariance, but both are below 7% of the gain over the cap, so the ambiguity of Section 5.4 is nearly costless for stopping (Proposition 3(a)).
Table 10. Stopping rules at (cross-fitted, evaluated by truncation)
| Policy | Profit per 1,000 | vs bank, % [95% CI] | vs uniform cap per 1,000 [95% CI] | Contacts, % of bank | Subscriptions, % of bank |
|---|---|---|---|---|---|
| Bank's practice | 61.301 | 0 | −3.151 [−3.785, −2.550] | 100 | 100 |
| Uniform cap | 64.452 | 5.140 [4.061, 6.230] | 0 | 88.615 | 97.608 |
| Month-specific cap | 66.608 | 8.657 [6.904, 10.831] | 2.156 [1.363, 3.168] | 66.922 | 89.634 |
| One-step look-ahead, | 67.706 | 10.449 [8.609, 12.672] | 3.254 [2.388, 4.348] | 64.310 | 89.418 |
| Backward induction, | 67.650 | 10.358 [8.554, 12.552] | 3.198 [2.362, 4.277] | 64.560 | 89.483 |
| Backward induction, | 67.482 | 10.083 [8.294, 12.238] | 3.030 [2.228, 4.098] | 65.456 | 89.741 |
| Backward induction, regime | 67.427 | 9.994 [8.142, 12.350] | 2.975 [2.126, 3.997] | 64.238 | 89.138 |
| Backward induction, client only (real-time feasible) | 66.084 | 7.802 [6.267, 9.614] | 1.632 [0.932, 2.376] | 76.453 | 93.513 |
Notes. Profit per 1,000 clients over all 41,188 held-out predictions; 95% intervals from 500 Poisson-bootstrap replicates, conditional on the cross-fitted rules. The month-specific cap and all backward-induction and one-step rules except the client-only rule use month effects estimated from other clients of the same month, so their gains are upper bounds for real-time use; the uniform cap and the client-only rule use no month information.
Gains grow with (Table 11). At neither rule beats the uniform cap significantly. At , with the bank's profit down to 35.62 per 1,000 clients, the client-only and client-and-month rules earn 39.65% and 49.42% more than the bank.
Table 11. Sensitivity to the cost-to-margin ratio (backward induction, )
| Quantity | |||
|---|---|---|---|
| Bank's profit per 1,000 | 86.972 | 61.301 | 35.616 |
| Rule vs bank, % | 1.010 [0.506, 1.501] | 10.358 [8.554, 12.552] | 49.424 [43.217, 57.213] |
| Client-only rule vs bank, % | 1.015 [0.463, 1.478] | 7.802 [6.267, 9.614] | 39.654 [34.632, 45.961] |
| Rule vs uniform cap, per 1,000 | 0.357 [−0.049, 0.702] | 3.198 [2.362, 4.277] | 10.161 [8.911, 11.445] |
| Client-only rule vs uniform cap, per 1,000 | 0.362 [−0.081, 0.740] | 1.632 [0.932, 2.376] | 6.680 [5.688, 7.833] |
| Month cap vs uniform cap, per 1,000 | −0.121 [−0.457, 0.194] | 2.156 [1.363, 3.168] | 9.764 [8.429, 11.040] |
| Rule vs month cap, per 1,000 | 0.478 [0.103, 0.861] | 1.043 [0.475, 1.589] | 0.397 [−0.060, 0.843] |
| Rule minus rule at | −0.003 [−0.165, 0.128] | 0.168 [0.016, 0.329] | 0.031 [−0.141, 0.196] |
| Rule minus regime- rule | 0.110 [−0.186, 0.379] | 0.223 [−0.268, 0.706] | 0.177 [−0.174, 0.495] |
| Rule minus one-step look-ahead | 0.038 [−0.106, 0.194] | −0.056 [−0.126, 0.039] | −0.006 [−0.089, 0.085] |
| Contacts / subscriptions, % of bank | 89.298 / 98.341 | 64.560 / 89.483 | 52.258 / 82.974 |
| vs uniform cap by month: mean, worst, best | 0.302, −0.215, 5.780 | 2.801, −1.373, 16.512 | 12.150, −1.172, 38.708 |
| Months with a loss vs uniform cap | 8 | 6 | 1 |
| Uniform cap chosen (contacts) | 7 or 11 | 6 | 4 |
Notes. "Rule" is the client-and-month backward-induction rule. Differences per 1,000 clients with 95% Poisson-bootstrap intervals; by-month rows summarize the 26 campaign months.
6. Managerial Insights
First, do not read the aggregate decline as fatigue. Through the fourth contact, as late contacts come mostly from campaigns that converted poorly, composition accounts for 0.777 of the decline with month effects and 0.516, a conservative bound, with a far worse-fitting single-regime model that omits within-regime month variation. Second, persisting through the third contact is a model-based conjecture, not a tested policy. Within cells, and only under Assumption 1, the second and third calls are not detectably less productive, but truncation cannot value calling more than the bank did (Corollary 1). Third, call again only if the predicted conversion probability of the next call exceeds the cost-to-margin ratio. With client covariates alone, feasible in real time, this earns 7.80% more than the bank at , and with same-month information at most 10.36%. Monitoring a campaign's conversion level as it runs is worth testing, since a month-specific cap captures 2.16 of the 3.20 per 1,000 clients that client-and-month rules gain over a uniform cap. Fourth, the one-step rule matched backward induction at every . Fifth, the fatigue question matters across campaigns: re-contacting a client who failed many times pays only if she was tired rather than unreceptive. The data favour fatigue at later contacts but cannot exclude selection in 2008, and a randomized test of re-contact and spacing would settle it.
7. Conclusion
A repeated-call campaign's falling conversion curve mixes the firm's choice of whom to keep calling, the sorting of receptive clients and what each call does. In a Portuguese bank's 2008–2010 campaigns, composition explains much of the early decline: 0.777 with month effects and 0.516, a conservative bound from a far worse-fitting single-regime model that omits within-regime month variation. Under Assumption 1, untestable here, the second and third calls are not detectably less productive than the first. Later contacts show a structural decline in every specification except a 2008 gamma model, whose implied equal attenuation of month and client contrasts is rejected (), though not its overall pattern (). Because the optimal rule depends only on conditional hazards, the remaining ambiguity costs little. At the real-time-feasible client-only rule earns 7.802% more than the bank. Client-and-month rules earn 9.994% to 10.358% more whichever heterogeneity model generates them, an upper bound because they use same-month outcomes.
If agents stopped calling clients who sounded uninterested, the policy and selection terms absorb selection on unrecorded information, and the apparent wear-in may be scheduled call-backs. Subscription is assumed at the last contact, a contact is a recorded attempt that may include unflagged inbound offers, and month and channel are last-contact attributes. The fatigue/selection split rests on the proportional-hazards form and the heterogeneity family, and the 2008 client contrasts are too imprecise to show or exclude attenuation. Truncation cannot value extra calls, two rules fix heterogeneity at full-sample estimates, and intervals condition on the cross-fitted rules. The cost ratio is unobserved, and the data cover one bank, product and period, with no identifier linking clients across campaigns. Future work should randomize the number and spacing of calls, link clients across campaigns, since multiple spells identify heterogeneity under weaker conditions (Van den Berg 2001), and add within-call signals (Manzoor et al. 2025).
Appendix
A.1. Lemma 1 and its proof. Lemma 1 (Selection identity): under Assumption 1, the mean of among clients with characteristics still at risk at is , where . Under Assumption 1 the firm's continuation decision for non-subscribers with characteristics is independent of , so the density of among clients in is proportional to times its density in . Hence , with expectations over .
A.2. Proof of Proposition 1. By definition ; subtracting Lemma 1 gives the identity, and gives the inequality. (i) follows directly. (ii) Under pure selection , so the left side is zero and . (iii) restates the identity. No step uses Assumption 2.
A.3. Proof of Proposition 2. From , . Since , the elasticity is . The intervals are consecutive and move to the right as grows, and is strictly decreasing, so the average strictly decreases. At , , whose elasticity is 1.
A.4. Proof of Proposition 3 and Corollary 1. Since , (2) depends on the primitives only through , which gives (a). With contact 1 given, ; backward induction from shows that continuing at is optimal exactly when . For (b), if and is non-increasing, then for all , so for and ; if then . For Corollary 1, if the realized sequence equals the logged one; if , the first contacts are logged failures, so the client does not subscribe and costs ; Assumption 3 rules out effects on other clients.
Table A1. Pooled decomposition with regime-specific heterogeneity
| Total | Policy | Selection on | Selection on | Structural | Structural, 2008 lower limit | |
|---|---|---|---|---|---|---|
| 4 | −0.538 [−0.666, −0.411] | −0.022 [−0.033, −0.002] | −0.373 [−0.413, −0.340] | −0.203 [−0.236, −0.182] | 0.060 [−0.060, 0.177] | −0.087 |
| 7 | −1.260 [−1.559, −0.989] | −0.209 [−0.296, −0.115] | −0.580 [−0.682, −0.461] | −0.343 [−0.478, −0.247] | −0.128 [−0.424, 0.153] | −0.388 |
| 10 | −1.621 [−2.156, −1.010] | −0.408 [−0.691, −0.096] | −0.662 [−0.842, −0.401] | −0.438 [−0.735, −0.238] | −0.113 [−0.604, 0.527] | −0.443 |
Notes. Gamma heterogeneity at the regime estimates (10.11 for 2008, 0.05 for 2009–2010), each regime with its own month effects, contact effects and coefficients; brackets: 2.5th and 97.5th percentiles of 500 Krinsky–Robb draws, conditional on the regime estimates of , whose own uncertainty is shown in Table 6. The last column uses 1.89 for 2008 and 0 for 2009–2010.
Table A2. Coefficients of the model with month effects (robust SE)
| Parameter | Pooled | 2008 | 2009–2010 |
|---|---|---|---|
| Contact 1 | −4.035 (0.128) | −4.691 (0.304) | −1.837 (0.143) |
| Contact 2 | −3.967 (0.130) | −4.540 (0.310) | −1.796 (0.145) |
| Contact 3 | −3.966 (0.133) | −4.567 (0.310) | −1.786 (0.149) |
| Contact 4 | −4.169 (0.142) | −4.789 (0.321) | −1.979 (0.162) |
| Contact 5 | −4.386 (0.158) | −4.979 (0.339) | −2.215 (0.185) |
| Contact 6 | −4.359 (0.172) | −4.836 (0.345) | −2.319 (0.220) |
| Contact 7 | −4.585 (0.204) | −5.308 (0.379) | −2.276 (0.267) |
| Contact 8 | −4.980 (0.270) | −5.493 (0.410) | −2.990 (0.466) |
| Contact 9 | −4.659 (0.275) | −4.981 (0.394) | −4.074 (1.012) |
| Contact 10 | −4.756 (0.315) | −5.462 (0.466) | −2.325 (0.518) |
| Contacts 11+ | −5.440 (0.233) | −5.997 (0.361) | −3.507 (0.613) |
| Jun 2008 | 0.131 (0.098) | 0.128 (0.098) | – |
| Jul 2008 | −0.012 (0.096) | 0.116 (0.135) | – |
| Aug 2008 | −0.177 (0.107) | 0.061 (0.158) | – |
| Oct 2008 | 4.213 (0.177) | 4.350 (0.181) | – |
| Nov 2008 | 0.221 (0.114) | 0.451 (0.164) | – |
| Dec 2008 | 1.930 (1.020) | 2.180 (1.019) | – |
| Mar 2009 | 2.095 (0.128) | – | baseline |
| Apr 2009 | 1.465 (0.097) | – | −0.600 (0.106) |
| May 2009 | 0.595 (0.095) | – | −1.446 (0.104) |
| Jun 2009 | 2.201 (0.104) | – | 0.097 (0.110) |
| Jul 2009 | 2.033 (0.147) | – | −0.085 (0.151) |
| Aug 2009 | 2.024 (0.107) | – | −0.094 (0.113) |
| Sep 2009 | 2.291 (0.130) | – | 0.163 (0.134) |
| Oct 2009 | 2.397 (0.114) | – | 0.280 (0.119) |
| Nov 2009 | 2.278 (0.118) | – | 0.156 (0.124) |
| Dec 2009 | 2.150 (0.140) | – | 0.028 (0.144) |
| Mar 2010 | 2.518 (0.123) | – | 0.406 (0.132) |
| Apr 2010 | 2.348 (0.143) | – | 0.222 (0.144) |
| May 2010 | 2.442 (0.126) | – | 0.324 (0.131) |
| Jun 2010 | 2.279 (0.131) | – | 0.180 (0.138) |
| Jul 2010 | 2.320 (0.118) | – | 0.203 (0.122) |
| Aug 2010 | 2.294 (0.132) | – | 0.179 (0.139) |
| Sep 2010 | 2.251 (0.123) | – | 0.136 (0.129) |
| Oct 2010 | 2.212 (0.139) | – | 0.098 (0.144) |
| Nov 2010 | 2.265 (0.155) | – | 0.139 (0.158) |
| Age (decades from 40) | −0.034 (0.019) | −0.104 (0.036) | −0.008 (0.023) |
| Age squared | 0.023 (0.008) | 0.025 (0.034) | 0.021 (0.009) |
| Job: blue-collar | −0.094 (0.060) | 0.144 (0.104) | −0.217 (0.077) |
| Job: entrepreneur | −0.074 (0.100) | 0.198 (0.149) | −0.234 (0.135) |
| Job: housemaid | −0.092 (0.106) | −0.185 (0.211) | −0.046 (0.127) |
| Job: management | −0.019 (0.062) | 0.209 (0.114) | −0.111 (0.074) |
| Job: retired | 0.069 (0.082) | 0.407 (0.173) | −0.038 (0.093) |
| Job: self-employed | −0.058 (0.087) | 0.024 (0.156) | −0.083 (0.106) |
| Job: services | −0.140 (0.066) | 0.126 (0.111) | −0.276 (0.082) |
| Job: student | 0.065 (0.078) | 0.479 (0.286) | 0.049 (0.084) |
| Job: technician | 0.002 (0.053) | 0.041 (0.099) | −0.006 (0.063) |
| Job: unemployed | −0.007 (0.088) | −0.197 (0.213) | 0.037 (0.099) |
| Job: unknown | −0.115 (0.184) | 0.287 (0.285) | −0.281 (0.225) |
| Married | 0.054 (0.050) | −0.022 (0.089) | 0.072 (0.060) |
| Single | 0.058 (0.058) | −0.011 (0.103) | 0.075 (0.071) |
| Marital status unknown | 0.149 (0.297) | 0.639 (0.473) | −0.117 (0.356) |
| Education: basic 6-year | 0.106 (0.090) | 0.247 (0.149) | 0.040 (0.113) |
| Education: basic 9-year | 0.044 (0.071) | 0.214 (0.121) | −0.042 (0.090) |
| Education: high school | 0.050 (0.069) | 0.176 (0.129) | 0.007 (0.082) |
| Education: illiterate | 0.572 (0.518) | 0.980 (1.034) | 0.577 (0.594) |
| Education: professional course | 0.025 (0.077) | 0.073 (0.145) | 0.015 (0.091) |
| Education: university degree | 0.068 (0.067) | 0.173 (0.129) | 0.033 (0.080) |
| Education: unknown | 0.075 (0.084) | 0.033 (0.178) | 0.079 (0.097) |
| Housing loan: unknown | −0.087 (0.104) | −0.360 (0.221) | −0.006 (0.121) |
| Housing loan: yes | −0.033 (0.030) | −0.053 (0.055) | −0.026 (0.036) |
| Channel: telephone | −0.570 (0.058) | −0.435 (0.127) | −0.613 (0.066) |
| No previous campaign | 0.300 (0.048) | 0.545 (0.218) | 0.286 (0.050) |
| Previous campaign: success | 1.056 (0.057) | 1.023 (0.622) | 1.043 (0.057) |
| Default status not 'no' | −0.199 (0.054) | −0.037 (0.070) | −0.432 (0.091) |
| Personal loan | −0.055 (0.042) | 0.005 (0.076) | −0.083 (0.050) |
| Two or more previous contacts | 0.008 (0.060) | not estimable | 0.009 (0.061) |
Notes. Contact baselines are (complementary log-log scale); month effects are relative to May 2008 (pooled, 2008) or March 2009 (2009–2010); "–" marks a parameter not in that model. Client categories are relative to job admin., divorced, basic 4-year education, no housing loan, cellular channel and a failed previous campaign. In 2008 the indicator for two or more previous contacts is nearly empty and its coefficient diverges (separation), so it is not estimable.
References
Allison PD (1982) Discrete-time methods for the analysis of event histories. Sociological Methodology 13:61–98. https://doi.org/10.2307/270718
Athey S, Wager S (2021) Policy learning with observational data. Econometrica 89(1):133–161. https://doi.org/10.3982/ECTA15732
Baker M, Melino A (2000) Duration dependence and nonparametric heterogeneity: A Monte Carlo study. J. Econometrics 96(2):357–393. https://doi.org/10.1016/S0304-4076(99)00064-0
Bitran GR, Mondschein SV (1996) Mailing decisions in the catalog sales industry. Management Sci. 42(9):1364–1381. https://doi.org/10.1287/mnsc.42.9.1364
Braun M, Moe WW (2013) Online display advertising: Modeling the effects of multiple creatives and individual impression histories. Marketing Sci. 32(5):753–767. https://doi.org/10.1287/mksc.2013.0802
Bult JR, Wansbeek T (1995) Optimal selection for direct mail. Marketing Sci. 14(4):378–394. https://doi.org/10.1287/mksc.14.4.378
Calder BJ, Sternthal B (1980) Television commercial wearout: An information processing view. J. Marketing Res. 17(2):173–186. https://doi.org/10.1177/002224378001700202
Chae I, Bruno HA, Feinberg FM (2019) Wearout or weariness? Measuring potential negative consequences of online ad volume and placement on website visits. J. Marketing Res. 56(1):57–75. https://doi.org/10.1177/0022243718820587
Chow YS, Robbins H (1985) A martingale system theorem and applications. Herbert Robbins Selected Papers (Springer, New York), 413–424. Reprint of the 1961 paper. https://doi.org/10.1007/978-1-4612-5110-1_38
Ciocan DF, Mišić VV (2022) Interpretable optimal stopping. Management Sci. 68(3):1616–1638. https://doi.org/10.1287/mnsc.2020.3592
Dubé JP, Hitsch GJ, Rossi PE (2010) State dependence and alternative explanations for consumer inertia. RAND J. Econom. 41(3):417–445. https://doi.org/10.1111/j.1756-2171.2010.00106.x
Ebbes P, Ascarza E, Netzer O (2026) Learning from many experiments: A hierarchical Bayesian framework for decomposing marketing treatment heterogeneity. Working paper, SSRN. https://doi.org/10.2139/ssrn.6213079
Elbers C, Ridder G (1982) True and spurious duration dependence: The identifiability of the proportional hazard model. Rev. Econom. Stud. 49(3):403–409. https://doi.org/10.2307/2297364
Elliott MR, Little RJA, Lewitzky S (2000) Subsampling callbacks to improve survey efficiency. J. Amer. Statist. Assoc. 95(451):730–738. https://doi.org/10.1080/01621459.2000.10474261
Fader PS, Hardie BGS (2007) How to project customer retention. J. Interactive Marketing 21(1):76–90. https://doi.org/10.1002/dir.20074
Ferecatu A, De Bruyn A, Mukherjee P (2024) Silently killing your panelists one email at a time: The true cost of email solicitations. J. Acad. Marketing Sci. 52(4):1216–1239. https://doi.org/10.1007/s11747-023-00992-w
Gans N, Koole G, Mandelbaum A (2003) Telephone call centers: Tutorial, review, and research prospects. Manufacturing Service Oper. Management 5(2):79–141. https://doi.org/10.1287/msom.5.2.79.16071
Goh KH, Hui KL, Png IPL (2015) Privacy and marketing externalities: Evidence from Do Not Call. Management Sci. 61(12):2982–3000. https://doi.org/10.1287/mnsc.2014.2051
Gönül F, Shi MZ (1998) Optimal mailing of catalogs: A new methodology using estimable structural dynamic programming models. Management Sci. 44(9):1249–1262. https://doi.org/10.1287/mnsc.44.9.1249
Groves RM, Heeringa SG (2006) Responsive design for household surveys: Tools for actively controlling survey errors and costs. J. Roy. Statist. Soc. Ser. A 169(3):439–457. https://doi.org/10.1111/j.1467-985X.2006.00423.x
Han A, Hausman JA (1990) Flexible parametric estimation of duration and competing risk models. J. Appl. Econometrics 5(1):1–28. https://doi.org/10.1002/jae.3950050102
Heckman JJ, Borjas GJ (1980) Does unemployment cause future unemployment? Definitions, questions and answers from a continuous time model of heterogeneity and state dependence. Economica 47(187):247–283. https://doi.org/10.2307/2553150
Heckman J, Singer B (1984) A method for minimizing the impact of distributional assumptions in econometric models for duration data. Econometrica 52(2):271–320. https://doi.org/10.2307/1911491
Hitsch GJ, Misra S, Zhang WW (2024) Heterogeneous treatment effects and optimal targeting policy evaluation. Quant. Marketing Econom. 22(2):115–168. https://doi.org/10.1007/s11129-023-09278-5
Jain DC, Vilcassim NJ (1991) Investigating household purchase timing decisions: A conditional hazard function approach. Marketing Sci. 10(1):1–23. https://doi.org/10.1287/mksc.10.1.1
Jenkins SP (1995) Easy estimation methods for discrete-time duration models. Oxford Bull. Econom. Statist. 57(1):129–136. https://doi.org/10.1111/j.1468-0084.1995.tb00031.x
Krinsky I, Robb AL (1986) On approximating the statistical properties of elasticities. Rev. Econom. Statist. 68(4):715–719. https://doi.org/10.2307/1924536
Kumar V, Zhang X, Luo A (2014) Modeling customer opt-in and opt-out in a permission-based marketing context. J. Marketing Res. 51(4):403–419. https://doi.org/10.1509/jmr.13.0169
Lancaster T (1979) Econometric methods for the duration of unemployment. Econometrica 47(4):939–956. https://doi.org/10.2307/1914140
Lodish LM (1971) CALLPLAN: An interactive salesman's call planning system. Management Sci. 18(4-part-II):P25–P40. https://doi.org/10.1287/mnsc.18.4.P25
Manzoor E, Ascarza E, Netzer O (2025) Learning when to quit in sales conversations. Working paper, arXiv:2511.01181. https://arxiv.org/abs/2511.01181
Meyer BD (1990) Unemployment insurance and unemployment spells. Econometrica 58(4):757–782. https://doi.org/10.2307/2938349
Moro S, Cortez P, Rita P (2014a) A data-driven approach to predict the success of bank telemarketing. Decision Support Systems 62:22–31. https://doi.org/10.1016/j.dss.2014.03.001
Moro S, Rita P, Cortez P (2014b) Bank Marketing [Dataset]. UCI Machine Learning Repository. https://doi.org/10.24432/C5K306
Murphy SA, van der Vaart AW (2000) On profile likelihood. J. Amer. Statist. Assoc. 95(450):449–465. https://doi.org/10.1080/01621459.2000.10474219
Naik PA, Mantrala MK, Sawyer AG (1998) Planning media schedules in the presence of dynamic advertising quality. Marketing Sci. 17(3):214–235. https://doi.org/10.1287/mksc.17.3.214
Neslin SA, Novak TP, Baker KR, Hoffman DL (2009) An optimal contact model for maximizing online panel response rates. Management Sci. 55(5):727–737. https://doi.org/10.1287/mnsc.1080.0969
Pechmann C, Stewart DW (1988) Advertising repetition: A critical review of wearin and wearout. Current Issues Res. Advertising 11(1-2):285–329. https://doi.org/10.1080/01633392.1988.10504936
Prentice RL, Gloeckler LA (1978) Regression analysis of grouped survival data with application to breast cancer data. Biometrics 34(1):57–67. https://doi.org/10.2307/2529588
Ridder G, Woutersen TM (2003) The singularity of the information matrix of the mixed proportional hazard model. Econometrica 71(5):1579–1589. https://doi.org/10.1111/1468-0262.00460
Schmidt P, Witte AD (1989) Predicting criminal recidivism using 'split population' survival time models. J. Econometrics 40(1):141–159. https://doi.org/10.1016/0304-4076(89)90034-1
Schmittlein DC, Morrison DG, Colombo R (1987) Counting your customers: Who are they and what will they do next? Management Sci. 33(1):1–24. https://doi.org/10.1287/mnsc.33.1.1
Seetharaman PB, Chintagunta PK (2003) The proportional hazard model for purchase timing: A comparison of alternative specifications. J. Bus. Econom. Statist. 21(3):368–382. https://doi.org/10.1198/073500103288619025
Self SG, Liang KY (1987) Asymptotic properties of maximum likelihood estimators and likelihood ratio tests under nonstandard conditions. J. Amer. Statist. Assoc. 82(398):605–610. https://doi.org/10.1080/01621459.1987.10478472
Simester D, Timoshenko A, Zoumpoulis SI (2020a) Targeting prospective customers: Robustness of machine-learning methods to typical data challenges. Management Sci. 66(6):2495–2522. https://doi.org/10.1287/mnsc.2019.3308
Simester D, Timoshenko A, Zoumpoulis SI (2020b) Efficiently evaluating targeting policies: Improving on champion vs. challenger experiments. Management Sci. 66(8):3412–3424. https://doi.org/10.1287/mnsc.2019.3379
Simester DI, Sun P, Tsitsiklis JN (2006) Dynamic catalog mailing policies. Management Sci. 52(5):683–696. https://doi.org/10.1287/mnsc.1050.0504
Sinha RK, Chandrashekaran M (1992) A split hazard model for analyzing the diffusion of innovations. J. Marketing Res. 29(1):116–127. https://doi.org/10.1177/002224379202900110
Van den Berg GJ (2001) Duration models: Specification, identification and multiple durations. Handbook of Econometrics, vol. 5 (Elsevier), 3381–3460. https://doi.org/10.1016/S1573-4412(01)05008-5
Van Diepen M, Donkers B, Franses PH (2009a) Dynamic and competitive effects of direct mailings: A charitable giving application. J. Marketing Res. 46(1):120–133. https://doi.org/10.1509/jmkr.46.1.120
Van Diepen M, Donkers B, Franses PH (2009b) Does irritation induced by charitable direct mailings reduce donations? Internat. J. Res. Marketing 26(3):180–188. https://doi.org/10.1016/j.ijresmar.2009.03.007
Vaupel JW, Yashin AI (1985) Heterogeneity's ruses: Some surprising effects of selection on population dynamics. Amer. Statist. 39(3):176–185. https://doi.org/10.1080/00031305.1985.10479424
6.Review
Automated review: Accept (novelty 3, rigor 4, clarity 4, reproducibility 5).
| Novelty 3/5 | Rigor 4/5 | Clarity 4/5 | Reproducibility 5/5 |
Refereed by Claude Fable 5.1, a different model from the one that wrote the paper (Claude Opus 5.5).
Final round of a manuscript that decomposes the declining per-contact conversion hazard in a Portuguese bank's telemarketing (41,188 clients, 105,754 contacts, UCI bank-additional-full) into continuation policy, selection on observables, unobserved selection and the structural contact effect, profiles gamma and split-population heterogeneity pooled and by regime, tests the heterogeneity it attributes with Proposition 2, and evaluates cross-fitted stopping rules exactly by truncating logged contact sequences. Across three rounds the authors answered every required change: the overclaims of the first version were removed, the feasible gain now stands beside the upper bound, the heterogeneity analysis was completed and probed, coefficient tables and intervals were added, and the remaining wording was brought into line with what the tables support. The record is unchanged since round 2 and the manuscript matches it in every number I checked; the reference list, cut by the review prompt in round 2, is complete. The remaining limits are inherent to single-spell campaign data without call outcomes and are stated plainly. I recommend acceptance. This report is reissued unchanged in substance on a record-alignment pass: the manuscript resubmitted here is byte-identical to the version accepted in round 3 and RESULT_JSON is unchanged, the run having been sent back only so that the pipeline's analysis record could be aligned with the accepted wording.
Strengths
- Round 3, text-only revision. All four required changes of my round-2 report are made as requested, the experiment record is unchanged (same executed script and RESULT_JSON, run time 88.282 s), and every number in the manuscript still matches it: the life table, the forty contact effects with Holm-adjusted p-values and the 95% intervals at contacts 2-3, every term, share and interval in Tables 4, 5 and A1, the profile estimates, limits and one-sided bounds in Tables 6-7, the attenuation factors and tests in Table 8, the leave-one-month-out rows in Table 9, the policy profits and intervals in Tables 10-11, and the 175 coefficients in Table A2. The reference list is now complete (52 entries, all retrieved sources with resolving DOIs, including Van Diepen et al. 2009a, 2009b and Vaupel and Yashin 1985).
- The attenuation finding is now stated exactly as Table 8 supports it, in the abstract, Section 5.4 and the conclusion: equal attenuation of the month and client indices is rejected (LR 9.327, p = 0.009), the overall pattern implied by theta = 10.11 is not (Wald p = 0.194), and the 2008 client factors (1.012, SE 0.494; 1.331, SE 1.012) are too imprecise to show or exclude the implied 0.780 and 0.454. 'Client contrasts do not' and 'contradicts the frailty it posits' are gone.
- The regime-coding composition share (51.6%) is reported with its caveat in the same sentence wherever it appears (abstract, Sections 5.3 and 6, conclusion, Table 5 note): the single-regime model fits far worse (log-likelihood -15,943.12 against -15,236.87) and omits within-regime month variation, so the share is a conservative bound, not an equally supported estimate.
- Table A2 reports the separated 2008 coefficient as not estimable and keeps the separation note; the trimming of repeated clauses (about 12 words of body prose) removed no number, claim or qualification, since the 'never call more than the firm did' restriction remains in Corollary 1 itself.
- The paper's substantive merits from the earlier rounds stand: an exact decomposition with Krinsky-Robb intervals; correct and honestly attributed propositions; Proposition 2 turned into a specification test that changed the paper's own reading of the 2008 estimate; completed profiles with all search bounds stated; cross-fitted stopping rules evaluated exactly by truncation with the real-time-feasible gain (7.80%, 6.27-9.61%) reported beside the same-month upper bound (10.36%, 8.55-12.55%); managerial insights stated as conditional conjecture where the design cannot test them. Management Science conventions (numbered propositions with proofs, notation table, percentage gaps, actionable insights, first person plural) are followed.
Weaknesses
- Limits that are inherent to the data and are stated in the paper: Assumption 1 (continuation independent of unobserved receptiveness) is untestable, so the within-cell wear-in at contacts 2-3 and the second managerial insight remain conditional; single-spell identification of heterogeneity is fragile and the title's fatigue-versus-selection question is answered only for specifications other than the 2008 gamma model; truncation cannot value calling more than the bank did; policy intervals condition on the cross-fitted rules and two rules fix heterogeneity at full-sample estimates; the cost ratio is unobserved.
- Smaller technical points in the revision analyses, acceptable as they stand: the Wald tests treat the frailty-implied factors as known; the 74-df full-interaction test in 2008 is low-powered and its degrees of freedom are approximate because of near-empty cells; the Holm family mixes nested specifications; the non-proportionality finding rests on the single-index scale test rather than the general interaction test.
- No figure: the author states that figures are not part of this record and that Tables 3, 4 and 7 hold the values a plot would show. The tables are complete, but a Management Science reader would expect the hazard profiles and the profile likelihoods drawn; this limits clarity, not correctness.
- Novelty remains moderate: the decomposition is an index-number identity and Propositions 1 and 3 restate known results, as the paper says; the specification test and the exact truncation evaluation are the distinctive contributions.
Required changes
—
Record
§Event timeline
- Run createdQuestion: “Contact fatigue or dynamic selection? Decomposing the declining conversion hazard across repeated telemarketing contacts and designing client-specific stopping rules' language=en discipline=management venue='Management Science” · output language: English · models: agent:claude-opus-5-5
- Stage started · Literature reviewAttempt 1 · agent:claude-opus-5-5
- awaiting_agent · Literature reviewsearch response for literature_review: literature_review-search-1.response.json
- Resume requested
- Stage started · Literature reviewAttempt 2 · agent:claude-opus-5-5
- Search complete · Literature review69 sources · 19,212 characters of notes
- awaiting_agent · Literature reviewcomplete_json response for literature_review: literature_review-complete_json-1.response.json
- Resume requested
- Stage started · Literature reviewAttempt 3 · agent:claude-opus-5-5
- Search complete · Literature review69 sources · 19,212 characters of notes
- Stage completed · Literature review
- Stage started · IdeationAttempt 1 · agent:claude-opus-5-5
- awaiting_agent · Ideationcomplete_json response for ideation: ideation-complete_json-1.response.json
- Resume requested
- Stage started · IdeationAttempt 2 · agent:claude-opus-5-5
- Awaiting approval · Ideation
- Approved · IdeationIdea 1 selected
- Stage started · ExperimentationAttempt 1 · agent:claude-opus-5-5
- Dataset provisioned · Experimentationbank_marketing_calls · 41,188 observations · 21 variables · sha256 e0bf5f5de5b8
- awaiting_agent · Experimentationcomplete response for experimentation: experimentation-complete-1.response.txt
- Resume requested
- Stage started · ExperimentationAttempt 2 · agent:claude-opus-5-5
- Dataset provisioned · Experimentationbank_marketing_calls · 41,188 observations · 21 variables · sha256 e0bf5f5de5b8
- Experiment round · ExperimentationRound 1: exit 0; result parsed (41.17s)
- Stage completed · Experimentation
- Stage started · AnalysisAttempt 1 · agent:claude-opus-5-5
- awaiting_agent · Analysiscomplete_json response for analysis: analysis-complete_json-1.response.json
- Resume requested
- Stage started · AnalysisAttempt 2 · agent:claude-opus-5-5
- Stage completed · Analysis
- Stage started · Paper writingAttempt 1 · agent:claude-opus-5-5
- awaiting_agent · Paper writingcomplete response for paper_writing: paper_writing-complete-1.response.txt
- Resume requested
- Stage started · Paper writingAttempt 2 · agent:claude-opus-5-5
- Stage completed · Paper writing
- Stage started · ReviewAttempt 1 · agent:claude-opus-5-5
- awaiting_agent · Reviewcomplete_json response for review: review-complete_json-1.response.json
- Resume requested
- Stage started · ReviewAttempt 2 · agent:claude-opus-5-5
- Awaiting approval · Review
- Sent back for revision · ReviewRe-running from Paper writing: “Referee recommendation: minor_revision. Make every required change below using only results already in RESULT_JSON; do not add analyses. Required changes (quoted): (1) "Correct Section 3.2 on where As”
- Stage started · Paper writingAttempt 3 · agent:claude-opus-5-5
- awaiting_agent · Paper writingcomplete response for paper_writing: paper_writing-complete-1.response.txt
- Resume requested
- Stage started · Paper writingAttempt 4 · agent:claude-opus-5-5
- Stage completed · Paper writing
- Stage started · ReviewAttempt 1 · agent:claude-opus-5-5
- awaiting_agent · Reviewcomplete_json response for review: review-complete_json-1.response.json
- Resume requested
- Stage started · ReviewAttempt 2 · agent:claude-opus-5-5
- Awaiting approval · Review
- Approved · Review
- Run completereport=/Users/haodingresearch/Desktop/auto-research/runs/studies-opus55/management-en/runs/eae11b02ad52/REPORT.md
- Resume requested
- Run completereport=/Users/haodingresearch/Desktop/auto-research/runs/studies-opus55/management-en/runs/eae11b02ad52/REPORT.md
- Reopened after sign-off · ReviewReopened after sign-off; re-running from Review: “Independent review: the paper is re-refereed by an independent referee (Claude Fable 5.1) that took no part in writing it.”
- Stage started · ReviewAttempt 3 · agent:claude-fable-5-1
- awaiting_agent · Reviewcomplete_json response for review: review-complete_json-1.response.json
- Resume requested
- Stage started · ReviewAttempt 4 · agent:claude-fable-5-1
- Awaiting approval · Review
- Sent back for revision · ReviewRe-running from Experimentation: “Independent referee (Claude Fable 5.1): major_revision. Required changes: (1) [text] Report the real-time-feasible gain beside the headline wherever the 10.4% appears (abstract, Sections 5.5 and 6, co”
- Stage started · ExperimentationAttempt 3 · agent:claude-opus-5-5
- Dataset provisioned · Experimentationbank_marketing_calls · 41,188 observations · 21 variables · sha256 e0bf5f5de5b8
- awaiting_agent · Experimentationcomplete response for experimentation: experimentation-complete-1.response.txt
- Resume requested
- Stage started · ExperimentationAttempt 4 · agent:claude-opus-5-5
- Dataset provisioned · Experimentationbank_marketing_calls · 41,188 observations · 21 variables · sha256 e0bf5f5de5b8
- Experiment round · ExperimentationRound 1: exit 0; result parsed (88.875s)
- Stage completed · Experimentation
- Stage started · AnalysisAttempt 1 · agent:claude-opus-5-5
- awaiting_agent · Analysiscomplete_json response for analysis: analysis-complete_json-1.response.json
- Resume requested
- Stage started · AnalysisAttempt 2 · agent:claude-opus-5-5
- Stage completed · Analysis
- Stage started · Paper writingAttempt 5 · agent:claude-opus-5-5
- awaiting_agent · Paper writingcomplete response for paper_writing: paper_writing-complete-1.response.txt
- Resume requested
- Stage started · Paper writingAttempt 6 · agent:claude-opus-5-5
- Stage completed · Paper writing
- Stage started · ReviewAttempt 1 · agent:claude-fable-5-1
- awaiting_agent · Reviewcomplete_json response for review: review-complete_json-1.response.json
- Resume requested
- Stage started · ReviewAttempt 2 · agent:claude-fable-5-1
- Awaiting approval · Review
- Sent back for revision · ReviewRe-running from Paper writing: “Independent referee (Claude Fable 5.1), round 2: minor_revision. Required changes: (1) [text] In the abstract, the end of Section 5.4 and the conclusion, replace 'client contrasts do not [attenuate]' ”
- Stage started · Paper writingAttempt 7 · agent:claude-opus-5-5
- awaiting_agent · Paper writingcomplete response for paper_writing: paper_writing-complete-1.response.txt
- Resume requested
- Stage started · Paper writingAttempt 8 · agent:claude-opus-5-5
- Stage completed · Paper writing
- Stage started · ReviewAttempt 1 · agent:claude-fable-5-1
- awaiting_agent · Reviewcomplete_json response for review: review-complete_json-1.response.json
- Resume requested
- Stage started · ReviewAttempt 2 · agent:claude-fable-5-1
- Awaiting approval · Review
- Sent back for revision · ReviewRe-running from Analysis: “Editorial alignment after acceptance (no new referee request): make the analysis record agree with the accepted manuscript. In 2008 equal attenuation of the month and client indices is rejected (LR 9.”
- Stage started · AnalysisAttempt 3 · agent:claude-opus-5-5
- awaiting_agent · Analysiscomplete_json response for analysis: analysis-complete_json-1.response.json
- Resume requested
- Stage started · AnalysisAttempt 4 · agent:claude-opus-5-5
- Stage completed · Analysis
- Stage started · Paper writingAttempt 9 · agent:claude-opus-5-5
- awaiting_agent · Paper writingcomplete response for paper_writing: paper_writing-complete-1.response.txt
- Resume requested
- Stage started · Paper writingAttempt 10 · agent:claude-opus-5-5
- Stage completed · Paper writing
- Stage started · ReviewAttempt 1 · agent:claude-fable-5-1
- awaiting_agent · Reviewcomplete_json response for review: review-complete_json-1.response.json
- Resume requested
- Stage started · ReviewAttempt 2 · agent:claude-fable-5-1
- Awaiting approval · Review
- Approved · Review
- Run completereport=/Users/haodingresearch/Desktop/auto-research/runs/studies-opus55/management-en/runs/eae11b02ad52/REPORT.md