SPY ETF Close Price Regression

1. Feature Preprocessing (fitted on training data only)

The preprocessing pipeline is modular and configurable via pp_steps. Each step can either overwrite ('o') or append ('a') features:

Default pipeline: cov_select → zscore → winsor → pca

Key design for avoiding data leakage:

• All preprocessing steps are fitted only on training data
• The same fitted transformers are applied to test data without refitting
• Each step fits on the data before any transformation from the current step

2. TimeSeries Conversion

• Convert preprocessed features and targets to Darts TimeSeries objects
• Use integer index (instead of datetime frequency) to handle trading days with irregular gaps

3. Model Training

• Train a Temporal Fusion Transformer (TFT) model from the Darts library
• Key hyperparameters:
  • input_chunk_length: Number of historical time steps as input (default: 30)
  • output_chunk_length: Forecast horizon (default: 1)
  • hidden_size, lstm_layers, num_attention_heads: Model architecture
  • Optimizer: AdamW with gradient clipping
• Model weights are saved to results/<exp_name>/tft_model.pt

4. Evaluation

• Use historical_forecasts to generate rolling 1-step-ahead predictions on both train and test sets
• Convert predicted log returns back to price space using from_logret()
• Calculate MSE/RMSE in price space
• Save results:
  • metrics.json: Train/Test MSE and RMSE
  • predictions.csv: Date, ground truth, prediction, and split label
  • predictions_plot.png: Visualization of predictions vs ground truth

5. Avoiding Future Data Leakage

• Preprocessing: All scalers/selectors are fitted only on training data
• Target transformation: Log returns are computed on the full dataset before splitting, ensuring proper continuity
• Evaluation: historical_forecasts with retrain=False ensures no future information is used during prediction
• NaN handling: Only forward-fill (ffill) is used; backward-fill is avoided to prevent leaking future data

Pretrained Model Performance

The baseline model is trained with the default configuration and achieves the following performance on the test set:

• Training MSE: 29.103605906890895
• Test MSE: 89.18892824798843

The plot below visualizes the model’s predictions against the ground truth prices on the test set: