Physics-Informed Singular-Value Learning for Cross-Covariances Forecasting in Financial Markets

Battle Count: 90

Highly relevant to quantitative trading. The paper directly addresses cross-covariance estimation for portfolio replication with tracking-error minimization, a core quantitative trading problem. The NN estimator achieves lowest OOS tracking error in every year from 2017-2024, with ~10% improvement over best non-neural benchmarks. The method is dimension-agnostic, handles non-stationarity, and remains stable as universe size grows. The portfolio replication application with long-only constraints is directly implementable. The connection to index tracking, cross-asset momentum, and principal portfolios broadens applicability.

Moderate complexity. The architecture is relatively simple (331K parameters: 2-layer MLP encoder with 16 hidden units, 2-layer bidirectional LSTM with 128/64 hidden sizes, 2-layer MLP head with 252 hidden units). However, the full pipeline requires: SVD computation, marginal projection calculation, proper padding for rectangular matrices, walk-forward training protocol, integration with RIEnet for end-to-end portfolio optimization, and constrained quadratic programming at evaluation. The dimension-agnostic design and permutation equivariance add conceptual complexity. Open-source code (CRIENet) reduces implementation burden.

Open-source TensorFlow/Keras implementation provided via CRIENet package (github.com/bongiornoc/CrossRIEnet). Detailed synthetic data-generating processes, training protocols, hyperparameters, and universe filtering criteria are described. Walk-forward evaluation protocol is fully specified. Bootstrap inference procedures are detailed. However, the full U.S. equity dataset (CRSP/Compustat) requires commercial access.