Quantitative Backtesting Standards

1. The Golden Rule of Time

"Thou shalt not peek at the future."

• •Random shuffling (train_test_split) is FORBIDDEN for time-dependent data.
• •All validation must be Walk-Forward: Train on $t_{0} \dots t_{n}$, Test on $t_{n+1}$.

2. Walk-Forward Validation (Rolling Window)

Simulate the production reality of predicting the "next" season.

Fold Structure (Example)

• •Fold 1: Train [2011-2018] -> Test [2019]
• •Fold 2: Train [2011-2019] -> Test [2020]
• •...
• •Fold N: Train [2011-2024] -> Test [2025]

3. The "Red Team" Metrics

A model is only valid if it survives the Temporal Degradation Test:

• •Does RMSE spike in later years? (Architecture rot)
• •Does the model fail to adapt to regime changes (e.g., Cap Spikes)?

4. Feature Lag Integrity

• •Ensure lag_1 features for Year $Y$ strictly strictly come from Year $Y-1$.
• •Any "future" features (e.g., 'career_total_at_retirement') are strictly prohibited.