Intraday Direction Intelligence Platform

Highlights

• Originated and scoped the project; made final training-strategy and cross-fold validation decisions jointly with Om Patel
• 28 engineered features: 5 stationary OHLCV transforms, 18 technical indicators, and 5 temporal cyclic features
• EWMA volatility-normalized labels (lambda 0.94 RiskMetrics) with adaptive per-fold quantile thresholds, on length-60 sliding windows
• Leakage controls built into label and split design: session-safe boundaries, fold-scoped normalization, and time-aware splits
• Implemented three architectures under one protocol: LSTM with Bahdanau temporal attention, a Temporal Fusion Transformer with variable-selection gating, and a dilated CNN-BiLSTM
• Specified evaluation against a majority-class control with McNemar significance testing and Monte-Carlo-dropout confidence-gated backtesting

How it was built

I proposed and scoped this course project: predicting 5-minute intraday direction (down, neutral, up) for 30 S&P 500 stocks. Yash Jain collected the data and set up the project; Om Patel built the dashboard, and Om and I made the final training and cross-fold validation decisions together. The pipeline builds 28 engineered features, labels moves by EWMA-normalised volatility with per-fold thresholds, and trains an LSTM, a Temporal Fusion Transformer and a dilated CNN-BiLSTM under one protocol with leakage controls. Limits: the repository records no test results, so this is a design, not a performance claim.