An Explainable Machine Learning Framework for Cross-Sectional Forecast-Based Fund Selection

Abstract

Since the 1990’s the global hedge fund industry has seen a rapid expansion. Its growing presence in financial markets ranging from equity, fixed income and derivative markets has inextricably linked it to the broader financial industry, with larger funds effectively acting as a market makers and liquidity providers in many markets. For both academics and practitioners, the space has established itself as a key area of research given the vast heterogeneity of investment styles and the high mutability of the industry. In this thesis, a cross sectional fund selection approach which builds upon the paradigm of explainable machine learning is proposed in a fully systematic setting. Four fund performance metrics, Sharpe and Sortino ratios, fund alpha and its t-statistic, are used as the ranking and selection metric, at investable inter-regime forecast horizons of 24 and 36 months. We find that quintile portfolios constructed from machine learning and deep learning approaches outperform linear models and benchmark portfolios constructed exclusively based on historical realizations of the forecast metric, in terms of absolute and risk-adjusted performance. We find that the extreme quintile portfolios realize a high (resp. low) value of the performance metric employed as forecast metric in model training. We find forecasting on the Sortino ratio to yield the most consistent overall performance, and find particular benefit in employing machine learning methods for bottom quintile fund selection (consistent identification of under-performers) in the case of forecasting on fund alpha. Explainability, achieved via the use of SHAP values further serves the purpose of outlining feature importance both at the aggregate and the individual fund level. At the aggregate level, all methods agree on a subset of statistically consistent predictors across investment style and forecast horizon; with discernible relevance of predictors constructed from interactions of fund returns with nowcasters, and management quality indicators. This consistency enables a discretionary fund selection process to be complemented by model forecasts and SHAP value-based feature importance delineations. There is thus evidence that proposed approach may be valuable for a discretionary fund manager looking to incorporate machine learning based signals into their selection process.