Deep Portfolio Manager Shows Double Risk‑Adjusted Returns in Futures Backtests

Global: Deep Portfolio Manager Shows Double Risk‑Adjusted Returns in Futures Backtests

A new deep‑learning macro portfolio manager, trained end‑to‑end to maximize a robust, risk‑adjusted utility, reportedly delivers net risk‑adjusted returns roughly twice those of traditional trend‑following strategies and passive benchmarks. The authors evaluated the system on daily closing prices of 50 diversified futures over the period 2010‑2025, incorporating realistic transaction costs. Their findings suggest strong performance across multiple market regimes, including the 2010s CTA winter, the post‑2020 volatility shift, and the pandemic‑induced inflation shocks.

Model Architecture and Causal Learning

According to the researchers, the model resolves the asynchronous “ragged filtration” problem through a Directed Delay, or Causal Sieve, mechanism. This component prioritizes causal impulse‑response learning over mere information freshness, allowing the system to focus on temporally relevant signals while mitigating lag‑induced noise.

Macroeconomic Graph Prior

The study introduces a Macroeconomic Graph Prior that regularizes cross‑asset dependencies based on established economic principles. The authors claim this graph‑based regularization improves signal extraction in environments characterized by low signal‑to‑noise ratios.

Robust Optimization Objective

To address distributional uncertainty, the authors employ a smooth worst‑window penalty as a differentiable proxy for Entropic Value‑at‑Risk (EVaR). They describe this approach as a window‑robust utility that encourages strong performance during the most adverse historical subperiods.

Backtesting Results

In large‑scale backtests spanning 2010‑2025, the system reportedly achieved net risk‑adjusted returns roughly twice those of classical trend‑following strategies and passive benchmarks, using only daily closing prices. The authors note that the model maintained consistent performance through the pandemic, inflation shocks, and the subsequent higher‑for‑longer environment.

Comparative Performance

The authors state that the new manager improves upon the state‑of‑the‑art Momentum Transformer architecture by roughly fifty percent, highlighting gains in both return and risk metrics.

Ablation Findings and Generalization

Controlled ablation studies, as described by the authors, identify strictly lagged cross‑sectional attention, the graph prior, principled transaction‑cost treatment, and robust minimax optimization as primary contributors to the model’s generalization capability.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.