Close-Price Prediction using Online Kernel Adaptive Filtering

Stock price prediction is a challenging and a tedious task. Although various methods have been developed for issue, an investigation on accurate and low latency methods is not given much attention. In addition, traditional regression and classification methods require batch-oriented and independent training. Thus, they are not suitable for stock price prediction as the data we are working with is non-stationary with so many confluencing factors. In this paper, we propose an online learning-based kernel adaptive filtering approach for stock price prediction. Specifically, we work with ten different kernel filtering algorithms and propose a method to predict the next closing price. The idea is tested on fifty stocks of the NSE index with nine different time-windows such as one-minute, five-minutes, ten-minutes, fifteen-minutes, twenty-minutes, thirty-minutes, one hour, and one day. To this, it should be noted here that this article is the first wherein a stock is analyzed by looking at these different time windows. Moreover, the empirical results suggest that Kernel adaptive filtering is an efficient tool for high-frequency trading as well. The work presented here shows the predictive capability and superiority of the kernel adaptive filtering class of algorithms over classical regression and classification methods.