Umeå universitets logga

Short-term mean reversions in equities presents trading opportunities if they can be distinguished from structural repricings. This study developed a binary classification system to determine whether machine learning architectures can systematically isolate profitable short-term reversal trading opportunities in US large-cap equities using option contracts. In the study, two separate supervised architectures were evaluated: Extreme Gradient Boosting (XGBoost) and a Multi-Layer Perceptron(MLP) neural network. The feature matrix included 34 technical and macroeconomic indicators captured in relation to the 53 (at the time of the study) largest components of the S&P 500 index.

To eliminate look-ahead bias and simulate trade execution with real market prices, the training and filtration system accessed only historical data and used chronological walk-forward dynamic retraining. Model performance was assessed using statistical metrics (ROC-AUC and PR-AUC) paired with monetary trade expectancy. The best performing model was ultimately integrated into a capital-constrained portfolio simulator to analyze performance across three risk management profiles (Conservative, Balanced, and Aggressive). Although real option market prices were used, transaction costs and spreads were not considered, thus real world execution would yield lower returns.

The findings demonstrated that trading options with an unfiltered baseline mean reversion strategy was unviable, yielding a negative expected return of -2.58% per trade and causing capital depletion in the portfolio simulation. However, the XGBoost architecture successfully extracted a persistent statistical edge, outperforming both the baseline and the continuous neural network approximator. Diagnostic attributions via Information Gain and SHAP values revealed that the tree ensemble based its decision boundaries on broader market context, specifically S&P 500 price dislocations and systemic volatility levels, to separate idiosyncratic liquidity vacuums from trend persistent market regimes. Ultimately, when synergy between capital allocation restrictions and the machine learning filter was achieved, the terminal cumulative return reached +194.24% over the tested period of two years and showed the potential viability of algorithmic options trading for a non-institutional market participant.