A Novel Machine Learning Approach for Predicting the NIFTY50 Index in India

Over the past decade, extensive research on stock market prediction using machine learning models has been conducted. In this framework, different approaches for data standardisation methods have been used for financial time series analysis and to assess the impact of data standardisation on the final prediction outcome. The paper uses the feature-level optimal rolling-window batch data standardisation method to improve the machine learning model's predictive power significantly. Along with the standardisation method, the paper explores the performance of the automated feature interactions learner (Deep Cross Networks) effect on a plethora of technical indicators aiming at predicting the movements of the NIFTY50 index in India, as these predicted changes are reflected in options contracts. Feature-level optimal rolling window data standardisation can identify the optimal window of time such that the correlation between features and the response variable is maximized, with most features correlating at 0.7. In the experiment, 48% of important technical indicators negatively correlated with the response variable. The Deep Cross Network regression model combined with the optimal rolling window batch data standardisation method outperformed all other model configurations at weekly and monthly data frequency. It achieved a directional hit rate of 69.52% (weekly) and 79.17% (monthly) and root mean square error of 2.82 (weekly) and 5.01 (monthly), generating a profit 5.5 times (weekly) and 2.85 times (monthly) greater than the benchmark buy-and-hold strategy providing opposing evidence to the sub-martingale model.