Multi-Target Transformer for Financial Forecasting (From Scratch)

Published a thoroughly documented implementation of a Multi-Target Prediction (MTP) Time-Series Transformer built entirely from scratch in PyTorch. This project features a custom dual-output decoder architecture capable of simultaneous regression for log returns and classification for volatility regimes. To ensure model stability, the system trains strictly on stationary-engineered features and incorporates statistical validation via Spearman Rank Correlation analysis to prune redundant features during the EDA phase.