HRFT: Mining High-Frequency Risk Factor Collections End-to-End via Transformer

Abstract

In quantitative trading, it is common to find patterns in short term volatile trends of the market. These patterns are known as High Frequency (HF) risk factors, serving as key indicators of future stock price volatility. Traditionally, these risk factors were generated by financial models relying heavily on domain-specific knowledge manually added rather than extensive market data. Inspired by symbolic regression (SR), which infers mathematical laws from data, we treat the extraction of formulaic risk factors from high-frequency trading (HFT) market data as an SR task. In this paper, we challenge the manual construction of risk factors and propose an end-to-end methodology, Intraday Risk Factor Transformer (IRFT), to directly predict complete formulaic factors, including constants. We use a hybrid symbolic-numeric vocabulary where symbolic tokens represent operators/stock features and numeric tokens represent constants. We train a Transformer model on the HFT dataset to generate complete formulaic HF risk factors without relying on a predefined skeleton of operators. It determines the general shape of the stock volatility law up to a choice of constants. We refine the predicted constants (a, b) using the Broyden Fletcher Goldfarb Shanno algorithm (BFGS) to mitigate non-linear issues. Compared to the 10 approaches in SRBench, a living benchmark for SR, IRFT gains a 30% excess investment return on the HS300 and SP500 datasets, with inference times orders of magnitude faster than theirs in HF risk factor mining tasks.