Predictive Analytics for Stock Market Trends: A Comparative Study of SARIMA and LSTM Architectures for NIFTY Index Forecasting

Abstract

Reliable forecasts of equity-index behaviour underpin risk management, portfolio allocation and market oversight. This paper presents a controlled comparison of a classical linear model, the Seasonal Autoregressive Integrated Moving Average model with exogenous regressors (SARIMAX), and a deep-learning sequence model, the Long Short-Term Memory (LSTM) network, applied to the daily closing price of the NIFTY 50 index of India's National Stock Exchange. A six-stage pipeline - data ingestion, preprocessing, exploratory diagnostics, parallel SARIMAX and LSTM branches, and a shared evaluation layer - was designed and implemented. An AIC/BIC grid search selected SARIMAX (5,1,2) for the univariate closing-price series, while a stacked LSTM network was trained on scaled multivariate Open-High-Low-Volume features. Both models were evaluated on an identical seventeen-day hold-out window immediately following the COVID-19 crash (15 April-8 May 2020) using RMSE, MAE and R2. LSTM achieved RMSE = 260.99, MAE = 190.96, R2 = 0.972, while SARIMAX returned RMSE = 361.17, MAE = 299.94 and a negative R2 of -1.959, performing worse than a naive mean forecast. A residual-correction hybrid SARIMA-LSTM architecture is proposed at the design level but left for empirical evaluation because of the compounding-error risk of its recursive multi-step structure. The results support the broader view that nonlinear, feature-rich deep-learning models tend to dominate linear statistical models during structural market disruption, while the confounding effects of the unequal input sets and the atypical test window are discussed explicitly as limitations.