Multiple Kernel Learning for stock price direction prediction

Abstract

Unstable and assumptive aspects of the securities makes it hard to predict the next day stock prices. There is no absolute indicator for financial forecasting but there are many technical indicators like simple moving average, exponential moving average, stochastic fast and slow, on balance volume for better accomplishment. It is important to have a significant and well-constructed set of features to elaborate stock trends. In this paper, we have proposed a Multiple Kernel Learning Model which predicts the daily trend of stock prices such as up or down, it comprises of 2-tier framework. In first tier, we extracted some technical indicators based on five raw elements- opening price, daily high price, daily low price, closing price and trading volume. In second tier, we built different base kernels on the extracted feature set and then combined these base kernels through Multiple Kernel learning, we have trained the model through walk forward method and predicted the movement of daily stock trend such as up or down, and then evaluated its performance. Experiment results shows that our proposed solution performs well consistently than baseline methods (Support Vector Machine) in terms of prediction accuracy for two commodities in stock market.