Dead Reckoning: Counting Your Customers Who Never Say Goodbye

Battle Count: 50

The paper is primarily about customer-base analytics and marketing, not financial markets or trading. However, it has tangential relevance: (1) the partial identification framework and calibration audit methodology could inform model risk management in quantitative finance; (2) the drift detection and regime-change alarm concepts parallel regime-switching detection in trading; (3) the critique of extrapolation-based point estimates (dead reckoning) resonates with concerns about overfitting and model uncertainty in quantitative strategies; (4) the customer-based corporate valuation connection touches on fundamental analysis inputs. The core contribution is not directly applicable to trading strategy development.

The conceptual framework (limit identity, partial identification) is straightforward. The calibration audit grid (vintage-by-horizon with CORP reliability diagrams, Brier scores, moving-block bootstrap) uses standard tools but requires careful implementation of the grid structure and censoring-aware scoring. The dynamic layer (isotonic shape + level tilt from partially observed cohorts) is moderate complexity, analogous to actuarial reserving. The main implementation challenges are: (1) maintaining the vintage-by-horizon audit infrastructure in production, (2) correctly handling overlapping windows and serial dependence in bootstrap inference, (3) the drift statistic computation, and (4) integrating the calibration layer with existing BTYD model fitting pipelines. The paper provides sufficient methodological detail for reimplementation.

The paper provides a replication on the publicly available CDNOW benchmark dataset and describes the audit protocol in detail. However, the primary empirical data comes from MakerStock (proprietary, author is co-founder). The author states replication materials accompany the paper. The dynamic layer and calibration methodology are described with sufficient detail for reimplementation, but the 13-quarter backtest and specific parameter choices depend on the proprietary panel. The CDNOW replication provides an independent verification path.