Adaptive algorithms for classification on high-frequency data streamsApplication to finance

Abstract

In recent years, the problem of concept drift has gained importance in the financial domain. The succession of manias, panics and crashes have stressed the nonstationary nature and the likelihood of drastic structural changes in financial markets. The most recent literature suggests the use of conventional machine learning and statistical approaches for this. However, these techniques are unable or slow to adapt to non-stationarities and may require re-training over time, which is computationally expensive and brings financial risks. This thesis proposes a set of adaptive algorithms to deal with high-frequency data streams and applies these to the financial domain. We present approaches to handle different types of concept drifts and perform predictions using up-to-date models. These mechanisms are designed to provide fast reaction times and are thus applicable to high-frequency data. The core experiments of this thesis are based on the prediction of the price movement direction at different intraday resolutions in the SPDR S&P 500 exchange-traded fund. The proposed algorithms are benchmarked against other popular methods from the data stream mining literature and achieve competitive results. We believe that this thesis opens good research prospects for financial forecasting during market instability and structural breaks. Results have shown that our proposed methods can improve prediction accuracy in many of these scenarios. Indeed, the results obtained are compatible with ideas against the efficient market hypothesis. However, we cannot claim that we can beat consistently buy and hold; therefore, we cannot reject it