v2.0 — Multi-Model Ensemble

Date: February 17, 2026

What Changed

Multi-Model Ensemble

Replaced the single HistGBT model with three gradient boosting implementations:

Model	Library	Key Strength
HistGBT	scikit-learn	Fastest CPU training
LightGBM	lightgbm	Native categorical features, GPU support
XGBoost	xgboost	Strong regularization, GPU support

All three train independently on the same features with MSE loss. Predictions are combined via equal-weight averaging for the ensemble.

Conformal Prediction

Added split conformal prediction for uncertainty quantification:

• Collects out-of-fold residuals during cross-validation
• Groups residuals by horizon bucket (day 1-7)
• Computes 50% and 90% confidence intervals per bucket
• Wider intervals for distant horizons, narrower for near-term

Walk-Forward Backtesting

Implemented a backtesting framework that simulates production conditions:

• Trains on data available up to each test date
• Generates predictions using only past information
• Stores results with _backtest suffix for separate evaluation

Naive Benchmarks

Added persistence and weekly seasonal baselines:

• Persistence: Tomorrow’s price = today’s same hour
• Weekly seasonal: Next week’s price = this week’s same hour and day
• Skill scores measure improvement over these naive baselines

Impact

• Ensemble reduces variance — individual model errors cancel out
• Confidence intervals enable risk-aware decision making
• Backtesting provides honest out-of-sample performance estimates
• Skill scores contextualize whether the model adds value vs simple heuristics