Evaluating the Effect of Meta-Labeling on Equity Market Neutral Strategy

Abstract: This thesis aims to construct an Equity Market Neutral (EMN) strategy framework to predict intraday excess returns of stocks within the S&P 500 index by utilizing machine learning techniques proposed by (López de Prado, 2018). The constructed EMN strategies within the framework utilizes techniques such as Stacked Single Feature Importance (SSFI), sample weighting, Probabilistic Sharpe Ratio (PSR), and meta-labeling. The objective of this paper is twofold: to construct a EMN strategy framework appropriate for financial data, and to analyze the effect of meta-labeling by comparing the investment skill generated with and without its implementation. The out-of-sample results of the meta-label model that uses only regime features as predictor variables indicate an absence of significant investment proficiency in terms of PSR. However, adding primary model predictor variables to the meta-labeling model shows a significant investment skill above 90%. Furthermore, the non-meta model displays the best results with a PSR of 99% and achieves gains during and as well as after the COVID-19 crisis with solid risk-adjusted returns.