Financial time series analysis and prediction using Chaos theory, HHT and SVR.

Abstract

Stock market prediction is a very complex and therefore well studied area of economics and applied mathematics. The stock market prediction is often termed as a non-solvable problem precisely because as cited many times by various authors that the probability of correct prediction is no less than the probability of success of a fair coin toss. In this thesis, we exploit the presence of chaos in stock market data; in particular, we use the Bombay Stock Exchange data for explanation, along with results of using different datasets of different countries, and use a novel de-noising algorithm, based on the Hilbert-Huang Transform (HHT), and apply it to the - Support Vector Regression (SVR) for prediction of the pre-processed time series data. We compare the results with the existing techniques based on wavelet denoising. The purpose of this thesis is two-fold. Firstly, it deals with the verification of Takens’ embedding theorem as applied to chaotic time series data and its denoising and prediction. The work provides an experimental proof that indeed prediction of financial time series is possible via machine learning. On the other hand, it also gives a brief review of the existing techniques in various areas of data analysis and prediction so that the algorithm used can be fully justified. The algorithm presented here achieves an error of less than 1.5 % which is an improvement on the other previously existing techniques.