[Paper] DeePM: Regime-Robust Deep Learning for Systematic Macro Portfolio Management

Source: arXiv - 2601.05975v1

Overview

The paper introduces DeePM (Deep Portfolio Manager), a deep‑learning system designed to allocate capital across a broad set of macro‑futures contracts while staying robust to changing market regimes. By tackling data‑timing issues, noisy signals, and the need for risk‑aware objectives, DeePM delivers risk‑adjusted returns that are roughly double those of classic trend‑following rules and even outperforms the recent Momentum‑Transformer benchmark.

Key Contributions

• Causal Sieve (Directed Delay) layer – solves the “ragged filtration” problem by forcing the model to learn from truly lagged, causally‑valid information rather than inadvertently peeking at future data.
• Macroeconomic Graph Prior – injects economic knowledge (e.g., commodity‑currency‑interest‑rate linkages) as a regularizer, improving cross‑asset signal extraction in a low‑signal‑to‑noise environment.
• Distributionally robust utility – implements a smooth worst‑window penalty that approximates Entropic Value‑at‑Risk (EVaR), encouraging strong performance during the toughest historical sub‑periods.
• End‑to‑end training with realistic transaction‑cost modeling – the cost model is baked into the loss, so the network learns to trade efficiently from day one.
• Empirical validation on 50 diversified futures (2010‑2025) – demonstrates consistent out‑performance across multiple regime shifts (CTA winter, pandemic, post‑2020 inflationary environment).
• Ablation study – isolates the impact of each architectural component, confirming that lag‑only attention, graph prior, cost‑aware loss, and robust optimization are the primary drivers of generalization.

Methodology

1. Data pipeline – Daily closing prices for 50 futures are transformed into lagged returns, macro‑economic indicators, and a sparse adjacency matrix encoding known economic relationships (e.g., oil ↔ USD, bond yields ↔ equity indices).
2. Model architecture
   • Directed Delay (Causal Sieve): a custom masking layer that only permits information older than a configurable delay to flow into the attention mechanism, guaranteeing causality.
   • Cross‑sectional attention with graph regularization: the attention scores are penalized according to the macro‑graph, nudging the network to respect economically plausible co‑movements.
   • Portfolio head: outputs raw position weights that are passed through a softmax‑like normalization and a transaction‑cost layer (proportional to turnover).
3. Training objective – The loss combines the negative of a risk‑adjusted utility (expected return minus a risk‑aversion term) with a smooth worst‑window penalty that approximates EVaR. This creates a minimax problem: the model seeks weights that perform well even under the most adverse historical windows.
4. Optimization – Stochastic gradient descent with Adam, using mini‑batches of rolling windows to preserve temporal structure. Early stopping is based on out‑of‑sample utility on a validation period (2018‑2019).

Results & Findings

• Robustness across regimes: Performance stayed within ±1 % of the 2010‑2025 average Sharpe during the CTA winter (2014‑2016), COVID‑19 crash (Mar‑2020), and the 2022‑2023 inflation‑driven volatility spike.
• Ablation insights: Removing the graph prior dropped Sharpe by ~0.3, while replacing the causal sieve with a standard look‑ahead mask caused a 30 % performance collapse due to leakage.
• Transaction‑cost sensitivity: Even with realistic slippage (0.5 bps per contract) and a 5 bps turnover cost, DeePM retained a Sharpe >1.5, indicating the model learned low‑turnover, high‑conviction signals.

Practical Implications

• For quant teams: DeePM offers a plug‑and‑play architecture that can be trained on any set of liquid futures or ETFs, with the macro‑graph easily customized to reflect the assets you trade.
• Risk‑aware portfolio construction: The EVaR‑style worst‑window penalty provides a differentiable way to embed tail‑risk considerations directly into the loss, reducing the need for post‑hoc risk overlays.
• Regime‑agnostic deployment: Because the model learns to rely on lagged, causally‑valid signals and respects economic linkages, it is less prone to over‑fitting to a single market environment—a common pain point for deep‑learning traders.
• Cost‑efficient execution: By training with an explicit turnover penalty, the resulting strategy naturally limits unnecessary trades, translating into lower execution costs in production.
• Extensibility: The Directed Delay concept can be applied to other time‑series domains (e.g., demand forecasting, energy load balancing) where asynchronous data streams cause “ragged” inputs.

Limitations & Future Work

• Data horizon: The study uses only daily closing prices; intraday information (order‑book depth, volume spikes) could further boost performance but also re‑introduces leakage risks.
• Graph prior staticity: The macro‑graph is hand‑crafted and fixed; learning a dynamic graph from data (e.g., via graph neural networks) could capture evolving economic relationships.
• Computational cost: Training the full model on 50 assets over 15 years requires several GPU‑days; lighter‑weight variants may be needed for rapid prototyping.
• Regulatory & interpretability concerns: While the graph prior adds some explainability, the internal attention weights remain a black‑box; future work could integrate attention‑visualization tools for compliance reporting.

Bottom line: DeePM demonstrates that a carefully engineered deep‑learning pipeline—one that respects causality, embeds economic structure, and optimizes a robust risk metric—can deliver truly regime‑robust macro portfolio performance, opening the door for more resilient AI‑driven trading systems in production.

Authors

• Kieran Wood
• Stephen J. Roberts
• Stefan Zohren

Paper Information

• arXiv ID: 2601.05975v1
• Categories: q-fin.TR, cs.LG, stat.ML
• Published: January 9, 2026
• PDF: Download PDF