Hall of Mistakes & Root Cause Audits
Scientific trust is not built by hiding errors. We publish our largest model misses, dissect their feature failures, document the algorithmic fixes applied, and verify their out-of-sample resolution.
Historical Error Case Studies & Algorithmic Fixes
Standard Double Poisson assumption treated home and away goals as strictly independent random variables. Failed to account for low-probability dependence parameter (rho) during heavy low-block tactical setups.
Implemented Dixon-Coles bivariate goal correction factor (rho = -0.041) to correctly inflate low-score probabilities (0-0, 0-1) when defensive low-block rating exceeds 1.4 stddev.
Feature engine used raw 5-match rolling averages without adjusting for 3-day turnaround cross-border travel fatigue vectors.
Integrated exponential time-decay vector and FIFA calendar distance decay metric into team offensive strength calculations.
Raw model probabilities over-weighted home advantage in high-variance rivalry matches, leading to an Expected Calibration Error (ECE) > 0.05 in derby cohorts.
Added non-linear Platt scaling and Isotonic Regression calibration layer trained on derby cohort historical out-of-sample data.