A1652
Title: Prediction of local extrema in financial time series with multiple timeframe extreme gradient boosting method
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)
Abstract: The prediction of maximum and minimum values in a financial time frame is crucial for developing automatic trading strategies. The existence of noise and non-stationarity inherited in the series challenges the effectiveness of traditional time series methods, and recent literature has demonstrated that the use of machine learning methods with richer input features is more promising. The financial time series is first processed by smoothing functions to alleviate the influence from random noise. The resultant series is processed for the identification of a set of pseudo local extrema, and their locations are subsequently modelled by multiple Bayesian optimized XGBoost models using a set of Shapley additive explanations (SHAP) selected features derived from level 2 data. The models are tuned to achieve maximum precisions, minimizing the chance of unfavourable trades. Afterwards, the extrema prediction model is integrated with a trading algorithm for the development of customer-centric strategies. The proposed method considers extrema prediction, market condition and customer risk collectively. It is evaluated empirically with 15-day real market data, and its VWAP outperforms several benchmarks. On the other hand, an effective data handling procedure to consolidate irregularly observed level 1 and level 2 data with the consideration of sampling bias will be addressed in this research work.
