Stock Market Prediction Using Principal Component Analysis, Iterative Filtering And Ensemble Deep Neural Networks

Abstract

Stock market behaviour is extremely volatile and complex in nature due to the randomness of the governing parameters. This makes attaining a high degree of accuracy in stock market forecasting extremely challenging. This paper presents a combination of discrete wavelet transform and gradient boosting approach for stock market prediction. A recursive wavelet decomposition approach is employed to filter the noisy data and subsequently a gradient boost approach is applied to different co-efficient values of the wavelet transform. Principal component Analysis has also been used for dimensional reduction. A composite predicted output them computed from the individual neural networks trained with different wavelet co-efficient. The day-wise data for shares has been considered with opening prices, closing prices, average price and volume as the temporal parameters. The performance evaluation has been done based on the mean absolute percentage error, mean square error, accuracy of prediction and regression values. It has been shown that the proposed approach and attains an average accuracy of 88.97% and a regression value of 0.99 for multiple benchmark datasets. A comparative analysis of the proposed technique shows that the method outperforms existing baseline techniques.