PPC Curve Autoresearch Dashboard

Autonomous ML experiment loop — advertising response curve fitting

1,073 product families

WMAPE Progress

Experiment Phases

1

Daily Data Fitting

0.3412 → 0.3195   |   Experiments 1–57

WLS with 1/sales weighting, gradient descent tuning, Huber loss, LR decay optimization. Discovered that reducing GD iterations improved WMAPE by reducing overfitting.

2

Model Refinement

0.3195 → 0.3143   |   Experiments 58–335

Dropped Hill/Exp models entirely — linear-only WLS won. Added power law, recency weighting, quadratic, saturation models, cached blending. Extensive hyperparameter micro-optimization converging on diminishing returns.

3

Temporal Aggregation Breakthrough

0.3143 → 0.2175   |   Experiment 336

Weekly aggregation of daily data produced a massive 30.9% WMAPE reduction in a single experiment. Smoothing noisy daily ad data removed measurement noise that was limiting model quality.

4

Monthly Aggregation & Holdout Tuning

0.2175 → 0.1322   |   Experiments 337–382

Biweekly then monthly (4-week) aggregation. Holdout split optimization (15% optimal). Recency and blend parameter refinement on the new data granularity.

5

Final Refinement

0.1322 → 0.1228   |   Experiments 383–384+

Selection margin removal, ensemble methods, three-way blending, cubic recency. Additional post-TSV experiments brought WMAPE from 0.1322 to 0.1228 (current best).

Experiment Log

# Commit WMAPE Best Time(s) Status Description

Key Insights

Biggest Single Improvement

4-week temporal aggregation collapsed noisy daily data into cleaner monthly signals. 0.189 → 0.169 (-10.6%)

Paradigm Shift

Weekly aggregation harness was the most impactful single change in the entire experiment series. 0.314 → 0.217 (-30.9%)

Simplification Wins

Dropping Hill and Exp models and using linear-only WLS improved WMAPE. Fewer GD iterations also improved results by reducing overfitting. Simpler is better.

Convergence Proof

250+ discarded experiments in the micro-optimization plateau proved the current approach is near-optimal for this data. The search space has been thoroughly explored.