A1213
Title: Modeling intraday local extrema with order book copula-based probabilities and machine learning models
Authors: Chun Fai Carlin Chu - The Hang Seng University of Hong Kong (Hong Kong) [presenting]
Po Kin David Chan - The Hang Seng University of Hong Kong (Hong Kong)
Ho Yin Willy Yue - The Hang Seng University of Hong Kong (Hong Kong)
Junxiang Jerry Peng - The Hang Seng University of Hong Kong (Hong Kong)
Abstract: Recent literature demonstrated that machine learning models can provide satisfactory performance on predicting local extrema in financial series. However, most studies did not incorporate L2 order book data due to its inherited noisiness, non-stationarity, and periodic variation characteristics. The aim is to exploit hidden information from L2 data using copula-based conditional probabilities with the considerations of empirically observed intraday persistent patterns. High-fidelity data is resampled in multiple frequencies for building empirical intraday pattern profiles, and time series data is then decomposed into persistent and varying components using the profile information. Several copula distribution functions have been examined to capture the dynamics among the varying components. The modeled dynamics are shown to improve the pseudo-local extrema identification. Bayesian-optimized machine learning models are employed to incorporate the extracted information. The models are optimized for VWAP values to avoid the occurrence of unfavorable trades. As a result, the models are tuned to achieve maximum precision. The resultant pseudo-extrema prediction models are evaluated by a threshold-based trading algorithm for performance evaluation. The analysis is evaluated with 30-day real market data from two stock exchanges. The empirical findings are discussed.
