major-year-project/tft-option-pricing

TFT-Lite -- Option Price Prediction (RegimeWise)

Part of RegimeWise -- A Deep Learning Option Pricing Framework (Phase 2: Baseline Model Comparison).

A dependency-free, pure-PyTorch Temporal Fusion Transformer ("TFT-Lite") that predicts SPY option mark price from option-chain, Greek, and liquidity features. Built with a Variable Selection Network, an LSTM encoder, and interpretable multi-head self-attention -- no pytorch_forecasting / pytorch_lightning dependency.

Data & preprocessing

Same dataset, feature set, and preprocessing as the project's LSTM baseline (hf://datasets/major-year-project/stock-data/spy_options_2024.csv), with a 70/30 chronological train/test split and 10-day per-contract input sequences.

Features (12): implied_volatility, delta, gamma, theta, vega, rho, tte_days, log_moneyness, strike, type_enc, volume, open_interest Target: mark

Architecture

• Variable Selection Network (per-feature Gated Residual Networks + learned softmax weights)
• LSTM encoder (hidden size 64, 1 layer(s))
• Multi-head self-attention (4 heads) with gated residual skip connections
• Linear output head -> single-step mark price prediction

Training

• Epochs: 30
• Optimizer: Adam (lr=0.001, weight_decay=1e-5), ReduceLROnPlateau
• Loss: MSE
• Batch size: 512

Test-set results (original price scale)

Metric	Value
RMSE	11.3021
MAE	5.4133
R2	0.9843

Files

• model.pt -- PyTorch state_dict for TFTLite (see config.json for architecture args)
• feature_scaler.joblib, target_scaler.joblib -- fitted StandardScalers (train split only)
• config.json -- architecture + training configuration, needed to reconstruct the model
• metrics.csv, test_predictions.csv, feature_importance.csv -- evaluation outputs