Methodology
Two LightGBM models trained independently with 30-trial Optuna Bayesian search over hyperparameters. Strict chronological train/holdout split prevents lookahead bias.
CV precision
54.5%
5-fold TimeSeriesSplit on train
Holdout precision
53.2%
Strictly out-of-sample
CV RMSE
0.0279
Log-return
Holdout RMSE
0.0266
Train / holdout split
Cutoff date
2024-05-14
Holdout fraction
20%
Train rows
4,980
Holdout rows
1,248
Hyperparameter search and final model fitting occur only on rows before the cutoff. All metrics reported on the holdout are true out-of-sample.
Top features · Classifier
Top features · Regressor
Signal rules
- BUY — classifier prob_up > 0.55
- SELL — classifier prob_up < 0.45
- HOLD — otherwise
- Target price — current_price × exp(pred_log_return)