서강대학교

초록

This paper evaluates the value of alternative time delineations of financial data under machine learning by analyzing whether imbalance-bars, particularly DIBs (Dollar Imbalance Bars), capture the presence of informed trading in cash equity markets in Korea. This is done by comparing future market variable forecasting performances between RFs (Random Forests) that use time bars and DIBs with identical and minimal features. Forecasting algorithms should have enhanced performances in instances it finds given samples more informative in formulating forecasts of various market variables. Time bars are established as the benchmark, which in design are incapable of carrying out designated details of DIBs. Tick level data is utilized on four of the most heavily traded stocks in the Korean Exchange (KRX), Samsung Electronics, KaKao corp., SK Hynix, and Celltrion. Rolling window learning and predictions with hyperparameter optimized RF algorithms are conducted. The results of the analysis display increased forecasting performance metrics under the usage of information-based bars, implying that the DIBs do indeed contain more coherent and predictive information content that aids learning and predicting capabilities of RFs. In turn, it is inferred DIBs are able to capture underlying trading behaviors of cash equity market participants as such in index futures markets, along with varying information arrivals.