Penalty function in SVM classifiers applied to automated trading strategies

Kamil Piotr Żbikowski

Abstract

This thesis is dedicated to the problem of machine learning algorithms applications to the area of financial time series prediction.

A new formula for constructing a penalty function that varies over time was proposed. It was used then to create an upgraded Support Vector Machine classifier. The idea of coupling penalty function and the underlying times series is motivated by the specific characteristics of financial time series. Its distribution, expected value and variance change over time. These properties directly cause difficulties in static analysis of classifications errors. Much more robust techniques should be applied.

Analysis of such data requires a dedicated computational system that would allow to adequately test the proposed classifier with the use of walk forward analysis method. Such a system was developed for the purpose of verification of the hypothesis presented in the presented thesis.

Experiments that were conducted showed that the results of short-term trends predictions with the use of proposed penalty function are better than with the use of plain penalty function . The statistical significance of the results was verified. The penalty function was then used for the purpose of constructing trading strategy. The data used for the experiments consisted of US stock and bitcoin quotations. Experiments show significant improvement in terms of the annual rate of return and the risk measured by the maximal drawdown over the test period.