A Two Stage LightGBM Framework for Intraday Bank Nifty Close Price Prediction Using Direction Guided Log Return Estimation

Abstract

Purpose: Accurately predicting intraday prices of equity indices represents a fundamental challenge in quantitative finance, due to high noise levels, nonlinear relationships, and the nonstationary characteristics of financial time series. This study presents a two-stage machine learning framework to predict the Bank Nifty closing price, India's most liquid derivatives benchmark, using data only until 12:00 PM on the prediction day.

Methods: In the initial stage, a LightGBM classifier predicts the intraday directional movement (bullish or bearish) of the final close relative to the day's opening price. This predicted direction is subsequently employed as a conditioning feature in the second stage, where a LightGBM regressor forecasts the log return of the close price from the open. The final closing price is calculated by multiplying the opening price with the exponential value of the predicted log return. Seventy-six features derive from 1-minute and 15-minute and daily OHLC data including momentum indicators, volatility measures, trend indicators, and market structure features.

Results: Tested on Bank Nifty 1-minute data from January 2015 to March 2024 (851,392 bars across 2,271 days), the framework yields 76.32% directional accuracy, Mean Absolute Error of 177.68 INR, RMSE of 236.00 INR, and MAPE of 0.4055% . It explains 98.77% of closing price variance (R² = 0.9877), with Cohen's Kappa of 0.5168 confirms significant predictive skill beyond random chance.

Conclusion: The two-stage framework surpasses single-stage baselines in directional accuracy while maintaining competitive error metrics, underscoring the effectiveness of directional conditioning for reliable stock price movement prediction