Prediction and Trend Analysis of Financial Time Series Deep Learning and Kernel Adaptive Filtering Perspective

Abstract

Stock price prediction is one of the important applications of time series analysis. Its success stems from its ability to reduce asset management costs, market impacts, and volatility risks. Stock price prediction is a significant challenge, and a plethora of techniques have been investigated. Nevertheless, a true solution is yet to be found. Despite a lot of effort, research has demonstrated that forecasting a stock’s price is difficult, especially when considering the nonlinear and non-stationary behavior of financial time series. The prediction of stock prices is traditionally done with regression and classification, thereby requiring a large set of batch-oriented and independent training samples. The work presented in this thesis takes on this challenge and present an automated framework to address two important issues in the field: i) Stock trend analysis and ii) Stock price prediction. For the former issue, we propose a 3D convolutional neural network based approach to classify the directional trends in a stock’s price. In contrast to existing literature, where work emphasizes upon predicting the direction in a single stock, we focus on a particular sector. This is done to analyze and capture the influence of one company on another. Further, multiple technical indicators are chosen, and the stock prices are converted into a 3D image. To find the best features, we experiment with hierarchical clustering. Lastly, we also complement the 3D convolutional neural network via the application of 3D ensemble learning. With extensive numerical investigation performed on forty-five different stocks, we found that the proposed work achieved up to 35% returns in some cases, with the average being 9.19%. To tackle the latter issue (stock price prediction), we focus our efforts on Kernel Adaptive Filtering. During experimentation on the previous problem, we discovered that deep learning based techniques require a lot of computational resources during training. Moreover, relying on offline-trained models and expecting them to perform well in real-world trading like circumstances is a strong assumption. Therefore, to take on the second challenge, we propose the idea of online Kernel Adaptive Filtering based approach to predict a stock’s prices. Specifically, we experiment with ten different kernel adaptive filtering algorithms to analyze a stocks’ performance and show the efficacy of the proposed work. The experiments are performed with quotes gathered at the window of one minute, five minutes, ten minutes, fifteen minutes, twenty minutes, thirty minutes, one hour, and one day. These time windows represent some of the standard windows used by high-frequency traders. The proposed framework was tested on fifty different stocks of the Indian stock index: Nifty-50. In terms of performance and compared to existing methods, the proposed method achieved a 66% probability of correctly predicting a stock’s next up or a down movement. To the best of our knowledge, the work presented in this thesis is the first wherein we analyze the price of a stock at these time windows. Furthermore, this article is the first to test the application of the KAF class of algorithms on all fifty stocks of the Nifty-50. Lastly, we also complement a stock’ price with additional statistical features to make the predictions more accurate. The experimental findings show that kernel adaptive filtering is a better option for predicting stock prices and may also be used as an alternative in high-frequency trading.