Compact Portfolio Network Cuts Parameters in 2025 Study
A paper published at ICAIF 2025 and uploaded to arXiv on July 25, 2026 describes a portfolio-optimization network that reduces learnable parameters from 39,586 to 2,175. The authors report lower realized variance than their selected benchmarks, but independent replication was not retrieved and the linked Compact-RIEnet repository is now deprecated in favor of the maintained RIEnet package.
A seven-page paper published in the 2025 ACM International Conference on AI in Finance proceedings was uploaded to arXiv on July 25, 2026. The study describes a compact neural architecture for global minimum-variance portfolio optimization that reduces its learnable parameter count from 39,586 to 2,175.
This is a 2026 arXiv release of a 2025 conference paper, not a newly completed 2026 study. Its performance results are author-reported, and no independent replication was retrieved.
How the architecture becomes smaller
The network replaces a 2,400-parameter lag-transformation layer with a five-parameter hyperbolic weighted moving average and a saturating exponential. It then combines a bidirectional GRU module for eigenvalue cleaning with a smaller marginal-volatility network.
The authors designed the parameter count to remain independent of both the look-back window and the number of assets. That can make the same architecture easier to apply across universes and sampling frequencies, although deployment still depends on training data, transaction assumptions and portfolio constraints.
In the paper's out-of-sample experiments, the compact model produced the lowest realized portfolio variance among the tested nonlinear-shrinkage and risk-parity benchmarks without reducing expected return. Under long-only constraints, the authors report that the lower variance supported more leverage with similar drawdown control. A trading simulator with margin-call dynamics was used to test resilience under aggressive leverage.
Reproducibility and limits
The linked GitHub repository contains the reference implementation, tests and paper-oriented defaults. It also carries a current warning that Compact-RIEnet is deprecated and that active development has moved to the RIEnet repository and package. Researchers reproducing the paper should therefore distinguish the legacy code path from the maintained package.
The results do not establish that the model will outperform in live portfolios. The retrieved evidence does not provide independent replication, trading-cost sensitivity across venues or evidence covering every market regime. For quantitative ML teams, the practical contribution is the size-invariant architecture and available code, while the reported performance should remain a hypothesis to test against local data, constraints and execution costs.
Key Points
- 1The 2025 ICAIF paper was uploaded to arXiv on July 25, 2026; it is not a newly completed 2026 study.
- 2The authors reduce learnable parameters from 39,586 to 2,175 and report lower realized variance than their selected benchmarks.
- 3The linked Compact-RIEnet repository is deprecated in favor of RIEnet, and no independent replication was retrieved.
Scoring Rationale
The paper offers a reproducible parameter-efficiency idea for neural covariance estimation, but it is a 2025 conference study newly uploaded to arXiv, with author-reported results, a deprecated linked repository and no retrieved independent replication.
Sources
Primary source and supporting public references used for this report.
Practice with real FinTech & Trading data
90 SQL & Python problems · 15 industry datasets
250 free problems · No credit card
See all FinTech & Trading problems
