Deep Learning to Predict Bank Nifty’s Next-Day Opening Direction

Summary

This project describes two classifiers intended to predict whether Bank Nifty and its leading constituents would open higher or lower on the following trading day. The stock models use daily OHLCV history and technical indicators for five constituents; the index model uses selected first- and last-hour one-minute OHLC data from the previous day. The author selects features with an XGBoost classifier, then trains artificial neural networks, scales inputs, and uses validation splits and early stopping to limit overfitting. A proposed combined signal would require the index and constituent predictions to agree, though that combination was not implemented.

The reported test-set AUC is about 83% for the stock models and about 60% for the index model. The project uses historical training and later test periods, but provides no evidence here of live trading profitability, transaction costs, or risk-adjusted returns. The author describes it as a constrained framework and notes limited intraday data and computing capacity; results may not generalize beyond the sample or remain useful in changing markets.

Key ideas

• The stock models classify next-day opening direction for five Bank Nifty constituents using daily OHLCV data and technical indicators.
• The index model uses the prior day’s first- and last-hour one-minute OHLC observations as inputs.
• Feature selection uses XGBoost, followed by neural networks with scaling, validation splits, and early stopping.
• The reported test AUC is about 83% for constituent models and about 60% for the index model.
• The proposed agreement filter between stock and index predictions was not implemented, and live profitability is not established.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.