Estimating the Movement of Stock Trends Employing Regularizing Momentum based Learning for Deep Neural Networks
Stock market forecasting has remained one of the most critical time series forecasting problems owing to the fact that serious financial repercussions may arise out of inaccuracies of forecasting models. However, achieving high forecasting accuracy pertaining to stock prices is challenging since stock prices exhibit an extremely volatile time series nature and dependence of a multitude of financial and non-financial parameters. It is commonplace to find a noisy baseline average for the time series statistical data which makes accurate forecasting challenging. This paper presents a data cleaning technique based on the discrete wavelet transform to offset the noisy nature of stock prices. Subsequently, a deep neural model has been proposed for pattern recognition and forecasting based on the gradient descent with momentum training algorithm. The evaluation of the proposed system has been done based on the values of the mean absolute percentage error, mean square error, regression and accuracy. A comparative analysis has been made with existing baseline techniques for S&P benchmark datasets. It has been shown that the proposed system outperforms the baseline techniques in terms of forecasting accuracy.