A Hybrid ARIMA-LSTM-Random Forest Framework for Predictive Analytics of NIFTY 50 Index Using Technical Indicators

Abstract

Predicting the stock indices is a challenging task. The prices are an outcome of decisions taken by millions of agents, and there does not exist a modelling technique that takes into account all aspects of such a process. This paper addresses the above difficulty by developing a forecasting framework using a combination of three completely different models in one stack. In particular, we suggest a new approach called HALRF that combines ARIMA, two layers of LSTM, and Random Forest for daily predictions of the NIFTY 50 index. While ARIMA deals with the linear autoregressive component of prices, LSTM models deep non-linear temporal dependencies and Random Forest introduce cross-sectional information based on 25 technical indicators (RSI, MACD, Bollinger bands, ATR, OBV, ADX, CCI, Williams%R, Stochastic Oscillator, Ichimoku Cloud, etc.). The whole ensemble uses ridge regression as a meta-modeling technique, which optimally balances contributions from different base models. Applying this method to 3,737 data points between 2010 and 2024 resulted in RMSE = 87.34, MAE =63.12, MAPE =1.24% and 79.6% accuracy in predicting price changes. None of the base models and any of their pairwise hybrids performed equally well.