Association mining based deep learning approach for financial time-series forecasting

Abstract

Stock market plays a vital role in a country’s economy, serving as a platform for companies to raise capital and enabling investors to share in their growth and success. The market is very unpredictable, characterized by non-linear variations and sudden fluctuations driven by a multitude of external factors. In the past, several traditional, deep learning, machine learning-based, and hybrid solutions have been put forth to estimate stock trends accurately. The existing techniques fail to achieve the desired accuracy due to the complex, non-linear, and random behavior of the stock time series. Moreover, the hybrid approaches in this domain have several shortcomings, such as shifting procedure dependency, optimal component count, computationally expensive, and many more. To address these challenges, and improve both accuracy and reliability, the current approach proposes a hybrid approach integrating association rule mining with deep learning models. The association rule mining aims to quantify the impact of a critical external factor (other companies’ stock trends) on a target company stock. The identified associated companies define the data to be fed to the neural models. The current approach implements a multivariate long short-term memory neural architecture for the stock price prediction task. Through performance analysis conducted on the National Stock Exchange (NIFTY50) dataset focusing on four well-known companies, the proposed hybrid approach demonstrates significantly improved prediction accuracy compared to benchmark methods. The proposed approach has demonstrated an average 8%–10% improvement in the prediction results of the benchmark approach. Moreover, the prediction performance of the proposed approach is statistically verified using the T-test.