State-of-the-art vs prominent models: An empirical analysis of various neural networks on stock market prediction

Abstract

Stock trading has always been a crucial and risky way of making money, requiring a profound understanding of the market and the data. Hence, stock market price prediction has always been a topic of interest for the research community. Existing literature has used legions of ways to accurately predict the price of stocks using sentimental analysis and fundamental and technical indicators, combined with the multitudinous linear, machine, and deep learning models. The existing research primarily focuses on classic univariate linear models like ARIMA, machine learning models including regression analysis and classification strategies, and traditional deep learning methods like LSTM; these have been the celebrity models in the stock price prediction problem. In the past few years, Recurrent Neural Networks (RNN) models have become synonymous with time series forecasting. Moreover, numerous new state-of-the-art models have also been gaining focus for time-series forecasting. This paper compares these new state-of-the-art models, including Temporal Fusion Transformer (TFT), N-BEATS, and Temporal Convolution Network, with prominent models, LSTM and GRU. Two years’ worth of historical data from different securities listed on the National Stock Exchange(NSE) of India is fed into these models to predict near-future closing prices. The comparison of the models is made using four performance metrics: Root Mean Square Error, Mean Absolute Error, Mean Absolute Percentage Error, and R Squared Score. The results showed that LSTM and GRU outperformed every other model with the slightest error. Moreover, TFT outperformed the state-of-the-art models and had somewhat comparable performance with LSTM and GRU, but not better!