Evaluating Effectiveness of Nonlinear Feature Extraction in Hedge Funds’ Returns Forecasting

Abstract

Hedge funds (HF) are actively managed investment vehicles employing diverse and often complex strategies. Accurate returns forecasting is essential for optimizing their performance and managing risk. This paper investigates the application of nonlinear dimensionality reduction (DR) methods in forecasting HF strategy performance, building upon prior work in financial time series analysis. We evaluate the effectiveness of Kernel Principal Component Analysis (KPCA), t-Distributed Stochastic Neighbor Embedding (t-SNE), Uniform Manifold Approximation and Projection (UMAP), and autoencoders on predictive performance of machine learning models. The extracted features are fed into several forecasting models, Support Vector Machine (SVM) with linear and nonlinear kernels, Neural Network (NN), and Extreme Gradient Boosting (XGB), to predict returns of five diverse HF investment strategies: Commodity Trading Advisors, Equity Long Short, Equity Market Neutral, Fixed Income Arbitrage, and Global Macro. The results demonstrate that nonlinear DR methods, particularly autoencoders, and KPCA combined with NN, significantly outperform other techniques. Our findings highlight the value of nonlinear transformations in enhancing predictive accuracy for HF returns time series.