Machine Trading: Theory, Advances, and Applications

Abstract

Dynamic contributions to trading are evaluated using covariations between position and price changes over a horizon. Other performance measures such as Sharpe ratios, gain–loss ratios, acceptability indexes, and drawdowns are also employed. Machine learning strategies based on Gaussian process regression (GPR) are compared with least squares (LSQ). Furthermore, both are generalized by invoking conservative valuation schemes that lead to the study of conservative conditional expectations modeled by distorted expectations. The latter lead to the development of distorted least squares (DLSQ) and distorted Gaussian process regression (DGPR) as the associated estimation or prediction schemes. Trading strategies are executed for nine sectors of the US economy using 14 different predictive factor sets. Results indicate improvements are made by GPR, DGPR over LSQ, and DLSQ, with the distorted versions also favorably affecting the drawdowns.

TOPICS: Statistical methods, simulations, big data/machine learning

Key Findings

• • The authors demonstrate how finance theory reduced the potential contributions of data science by assumption.

• • The concept of conservative conditional expectations is defined, developed, and implemented to enhance predictive technologies such as least squares and Gaussian process regression.

• • Using factor attributes as input, the trading of stock return predictions classically and conservatively illustrates the practical benefits of the advocated theoretical advances.