Home Professional Work Case Studies Research & Collaboration Contact
Back to Research & Collaboration

Undergraduate Research · Customer Analytics · Forecasting

BTYD Transaction Modeling

Organization
Inha University
Period
2026 – Present
Role
Undergraduate Research Assistant
Key Result
−0.758 Holdout MAE
Type
Conditional Model Evaluation

Built a continuous-time Neural Jump ODE that preserves BTYD-style purchase intensity, churn risk, and P(alive), then tested when a calendar covariate path adds measurable forecast value.

Background & Problem

BTYD models give useful customer-state outputs in non-contractual settings, but their static-rate assumption can miss calendar-driven changes in transaction behavior. The question was not whether a more complex model always wins, but whether a diagnostic can identify when additional time-varying information is useful.

Research Strategy

I first used calibration-only diagnostics to separate individual non-stationarity from cohort seasonality. I then compared an autonomous continuous-time model with a version that incorporates a calendar covariate path, using holdout purchase-count error rather than training fit as the decision metric.

Execution

  • Customer-state modeling: Represented transaction histories through continuous-time purchase intensity, churn risk, and P(alive) outputs while retaining the BTYD likelihood structure.
  • Calibration-only gate: Tested for seasonal cohort structure before observing holdout outcomes, so the model choice is tied to an interpretable diagnostic rather than post-hoc selection.
  • Holdout evaluation: Compared the calendar-aware model against the autonomous ODE with user-level holdout MAE and paired uncertainty intervals.
−0.758
MAE on seasonal panel
2
Reported evaluation panels
95% CI
[−1.079, −0.440]
P(alive)
Customer-state output retained

Result & Interpretation

On dunnhumby, where the gate detected cohort seasonality, the calendar covariate path reduced holdout MAE by 0.758 versus the autonomous ODE (95% CI [−1.079, −0.440]). On CDNOW, where the same gate found no cohort seasonality, the change was −0.041 MAE and not material.

The result supports a conditional deployment rule rather than a blanket replacement claim: add calendar information when the calibration data shows the structure it can explain. The reported continuous-time model remained statistically comparable with the BTYD baseline on both panels.