A functional mixture prediction model for dynamically forecasting cumulative intraday returns of crude oil futures

Since the high-frequency crude oil futures price data from intraday trading sessions exhibit continuous functional characteristics, we propose a functional mixture prediction (FMP) model for real-time forecasting of crude oil cumulative intraday returns (CIDR). The core idea of FMP is dynamic forecasting after adaptive classification. Specifically, we develop an adaptive functional clustering algorithm to identify the distinct patterns of CIDR curves and establish a probabilistic discriminant model to estimate their cluster membership probabilities. The mixture prediction of a new partially observed CIDR is obtained by weighting its predicted trajectory in each cluster with its estimated membership probabilities. Moreover, we design an adaptive information updating mechanism to further improve the accuracy of intraday forecasts. Empirical results from applying FMP to forecast the CIDR of China’s crude oil futures show that the proposed FMP not only outperforms several competing forecasters but also provides additional insights into CIDR analysis by revealing distinct patterns in daily CIDR curves of similar variability and typical temporal trends. Furthermore, we provide evidence that FMP can achieve greater gains for traders with different risk preferences based on our designed intraday trading strategies.