Stock Market Forecasting: Comparative analysis of SARIMA, CNN and LSTNet Models

Abstract

Indian stock market with market capitalization of $2.3 trillion, over 5500 listed firms on BSE and over 1700 listed firm on NSE has been a very promising market for investors all over the world. This market is tracked by two prominent indices Sensex and Nifty. Forecasting stock indices has been everyday struggle for financial analysts. The unpredictable nature of news and its impact on stock indices makes this task very difficult. Similarly, prediction of stock market crash, has been subject of study for decades. The aim of this work is to compare and examine accuracy of Long-and Short-term Time-arrangement Network (LSTM) with Seasonal ARIMA and Convolutional Neural Network (CNN) for forecasting returns of Indian stock market. For this study, historical data has been collected of Indian stock market from 2000 to 2020 and predictions are made for test data which is subset of collected data. Various models are compared using mean squared error and features of time series data. Stock return data has inherent characteristics such as its temporal nature, sequential nature, memory. Some of these characteristics are captured in LSTM and CNN which makes them effective for stock market forecasting. This study finds LSTM to yield least mean square error, thus making it most accurate amongst CNN and SARIMA in contrast to other published findings. This study provides comparison between statistical methods usually employed by finance analyst for analyzing stock returns. It highlights the fact the models used for individual company prediction may not be most accurate for index prediction. This study would be helpful in picking stocks for daily trade. Setting upper and lower limit for weekly trades.