Challenges of conventional feature extraction techniques for multivariate time-bars assets

Abstract This study investigates the inherent limitations of conventional dimensionality reduction techniques when applied to time-bars financial datasets. Such datasets are characterized by low correlation, pronounced heteroscedasticity, and non-Gaussian return distributions—properties that often violate the assumptions underpinning traditional methods. Our empirical findings reveal that these techniques tend to exhibit inflated generalization performance on out-of-sample tests, yet fall short in generating interpretable signals for financial machine learning applications. To rigorously examine this issue, we evaluate 14 feature extraction models on time-bars asset data, focusing on their ability to produce robust informational signals. While the Denoising Autoencoder outperforms several competing methods with respect to covariance-based metrics, further statistical analysis indicates that false discoveries may compromise its apparent efficacy. Despite strong cross-validation results, we perform a portfolio optimization backtesting using features derived from both the original and reconstructed datasets within a defined market regime. The near-identical cumulative returns observed across both strategies reinforce our central hypothesis: the marginal utility of conventional feature extraction methods in financial contexts is limited, particularly when they are deployed without addressing the structural idiosyncrasies of financial data.