Trading With Trees: Exploiting Market Inefficiencies With Random Forests

Abstract

Random forests are an ensemble learning method popular for their ability to capture nonlinear relationships and interactions between input features. This thesis analyzes the capacity of random forests to exploit patterns in market data for the purpose of statistical arbitrage. We formulate two candidate random forest models to use as predictors of relative stock performance. The first model is trained only on historical stock returns, and the second model is trained on historical stock returns plus a set of additional economic factors. We use forecasts generated by these models to simulate dollar-neutral trading strategies on the universe of S&P 500 Index constituents using data from 1993 to 2022.

Our analysis is twofold. First, we empirically evaluate the profitability of the trading strategies we simulate and assess their robustness to changing market conditions. Second, we deconstruct the random forests deployed in our simulations and reveal the mechanisms that influence their forecasting capacity.