Computational intelligence for the prediction of financial markets

Abstract

Purpose: In spite of the inherent randomness in the behavior of financial markets, the potential profitability of autonomous trading systems motivates many researchers to develop forecasting models based on machine learning technologies, which attempt to make predictions for future price movements. State of the art algorithms have been proven to achieve profitable trading results, but still lack the accuracy that make these systems reliable. This research aims to introduce a new algorithm that recognizes these weaknesses of the forecasting models and filters potentially incorrect predictions to increase the reliability of such systems. Design: The research performed in this dissertation aims to improve neural network based time-series forecasting by retrieving reliability information from the networks past performance to filter out inaccuracies. The approach combines a customized classification algorithm with a neural network based forecasting model. The algorithm estimates the prediction performance for the next input by evaluating the performance of similar patterns. Findings: The results show that the proposed algorithms are able to significantly improve the accuracy of the predictions, leading the system to greatly reduced risk and more stable profits. In many cases the algorithm correctly classified incorrect predictions, thus avoiding drawdowns and decreasing volatility. Conclusions: Reducing the number of effective predictions in financial forecasting by means of filtering is a powerful methodology that increases the reliability of a trading system in terms of accuracy. The introduced classification algorithm satisfyingly performed this task. Implications: While the filtering mechanisms do not replace research on more accurate machine learning algorithms, it can extend future forecasting systems with capabilities of handling the particular nature of financial data.