💰 quantitative finance

End-to-End Parametric Portfolio Policies for Cross-Asset Futures Timing: When Do AI Models Beat Simple Rules?

This paper demonstrates that end-to-end AI models, particularly transformer-based architectures trained with a differentiable Sharpe ratio, can outperform simple rules-based strategies in cross-asset futures timing by directly mapping market states to portfolio weights while effectively managing transaction costs.

Original authors: Austin Pollok, Kevin Robik

Published 2026-07-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Austin Pollok, Kevin Robik

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the captain of a massive ship carrying cargo across six different oceans: stocks, bonds, currencies, energy, metals, and farm goods. Your goal is to decide how much cargo to put in each hold to maximize your profit while avoiding storms.

For a long time, captains have used two main ways to make these decisions:

The Simple Rules: "Put equal amounts in every hold," "Balance the weight so no single hold is too heavy," or "Follow the wind (momentum)." These are easy to follow but sometimes miss opportunities.
The Complex AI: A super-smart computer that tries to predict the weather and the market, then calculates the perfect cargo mix. The problem is, these computers are often so complex they make mistakes, or they get so excited they change their minds too often, wasting fuel (money) on unnecessary moves.

This paper asks a very practical question: When is the super-smart computer actually better than the simple rules, and when is it just a waste of time?

The Experiment: A High-Speed Race

The authors set up a race using the 16 most popular "futures" contracts (think of these as bets on the future price of things like oil, gold, or the S&P 500). They tested two types of AI captains against three simple rule-followers:

The Simple Rules: Equal weight, Risk Parity, and Time-Series Momentum (following trends).
The AI Captains:
- The LSTM: A computer that remembers the past like a diary, reading one day at a time.
- The Transformer: A more advanced computer (like the ones powering modern chatbots) that can look at the whole picture at once, spotting patterns across different oceans simultaneously.

They trained these AI captains using a "Sharpe Ratio" scorecard, which measures how much profit you get for every unit of risk you take.

The Big Findings

1. The "Smart" AI isn't always the "Fast" AI
The most surprising discovery was about turnover, or how often the captain changes their mind.

The LSTM was like a nervous captain who constantly rearranges the cargo. It found some good opportunities, but it moved the cargo so much that the "fuel costs" (transaction fees) ate up all its profits.
The Transformer was like a calm, strategic captain. It found similar opportunities but made fewer moves. Because it didn't waste fuel on constant rearranging, it kept more of its profit. In the end, the Transformer was the winner because it was efficient, not just smart.

2. The AI wins in some oceans, but not all
The AI didn't beat the simple rules everywhere.

In the "Stock" ocean: The simple rules (just buying everything equally) were already doing such a great job that the AI couldn't find much extra value.
In the "Agriculture" and "Cross-Asset" (all oceans combined) oceans: The AI found a slight edge. It could navigate the choppy waters of farm goods and the complex mix of all assets better than the simple rules.
In the "Interest Rate" and "Currency" oceans: The simple "follow the trend" rule was actually the best. The AI couldn't improve on it.

3. The "Magic" might just be a different kind of exposure
When the AI did well in the stock market, the authors dug deeper. They found the AI wasn't necessarily predicting the future better than a human. Instead, it was just holding the stocks in a slightly more efficient way, taking on a bit less risk for the same reward. It wasn't "magic timing"; it was just "efficient steering."

The Verdict: When to Use the AI?

The paper concludes that the AI is a useful tool, but it's not a magic wand that works everywhere.

Use the AI (specifically the Transformer) when you are managing a mix of many different assets (like a global portfolio) or in choppy markets where simple rules get confused. The AI's ability to stay calm and trade less is its superpower.
Stick to the Simple Rules when the market is trending strongly (like in interest rates) or when the market is already very efficient (like in major stock indices). In these cases, the AI's extra complexity doesn't pay off.

The Bottom Line

Think of this like choosing a car.

Simple Rules are like a reliable, fuel-efficient sedan. They get you where you need to go 90% of the time without fuss.
The AI is like a high-performance sports car. It can go faster and handle tricky corners better, but only if you have a skilled driver (the Transformer) who knows when to shift gears and doesn't burn too much gas. If you try to drive the sports car on a straight, empty highway, the sedan is actually the better choice.

The paper teaches us that in the world of investing, complexity doesn't automatically equal better results. Sometimes, the best strategy is the one that knows when not to move.

Technical Summary: End-to-End Parametric Portfolio Policies for Cross-Asset Futures Timing

1. Problem Statement

The paper addresses the challenge of cross-asset timing in systematic multi-asset investing (global-macro and managed-futures strategies). While timing decisions drive realized risk and return, they are notoriously difficult to execute. The standard quantitative approach relies on a two-step "predict-then-optimize" framework: forecasting expected returns and covariances, then optimizing weights via mean-variance optimization. The authors argue this pipeline is fragile; noisy return forecasts can compound into unstable portfolio weights.

The central research question is practical: When does an AI-based, end-to-end allocation policy outperform simple, rules-based strategies (such as equal weighting, risk parity, and time-series momentum) enough to justify its added complexity? The study focuses on a highly liquid, investable universe where illiquidity premia are not a source of alpha, isolating the pure timing skill of the models.

2. Methodology

Data and Universe

The study utilizes end-of-day data for the sixteen most liquid CME futures across six asset classes: Equity Indices, Interest Rates, Foreign Exchange, Energy, Metals, and Agriculturals. The data spans 2001–2024. The authors use continuous contract series to ensure gap-free histories and avoid look-ahead bias. The feature set includes raw daily returns and a set of economically interpretable engineered features (trend, mean reversion, Hurst exponent, volatility, correlation, and higher moments), though the primary out-of-sample analysis relies on the cross-section of daily returns as the state.

Policy Framework

The authors employ an end-to-end parametric portfolio policy approach. Instead of forecasting returns, a neural network maps market states ( $X_t$ ) directly to portfolio weights ( $w_t$ ) via a differentiable policy function $\pi_\theta(X_t)$ .

Objective Function: The models are trained to maximize a differentiable Sharpe ratio (minimizing the negative Sharpe ratio). This collapses prediction and optimization into a single step.
Constraints: Long/short positions are permitted with unit gross exposure. The long/short budget constraint is embedded within the policy architecture rather than imposed as a post-hoc constraint.
Transaction Costs: A cost-aware objective penalizes turnover ( $\lambda$ ), set at a realistic 2 basis points (bp) for liquid futures, though sensitivity analysis extends to 10 bp.

Architectures

Two neural network architectures are compared:

LSTM (Long Short-Term Memory): Serves as the neural-network-based baseline for sequence modeling.
Transformer: A custom architecture following recent literature on attention-based futures momentum. It includes:
- Time2Vec Embedding: To capture periodic structures since standard attention is order-agnostic.
- Encoder-Decoder: Four layers of multi-head self-attention with gated residual networks (GRN).
- Causal Masking: Ensures weights depend only on information available at the time of position formation.
- Output Layer: A signed-softmax layer converts decoder scores into weights, normalizing total absolute exposure.

Experimental Design

The evaluation uses an expanding-window walk-forward design. Models are trained on an initial 40% of the sample and re-fit in roughly four-year blocks with a 90/10 train-validation split. Performance is benchmarked against three standard rules: Equal Weighting (1/N), Risk Parity, and Time-Series Momentum (TSMOM).

3. Key Results

Performance Across Asset Classes

Overall Ranking: The learned policies (Transformer and LSTM) generally rank above simple rules in the pooled cross-asset portfolio and in several single-asset-class sleeves (e.g., Agriculture, Metals), but not uniformly.
Asset Class Specifics:
- Equity Indices: The LSTM achieves the highest Sharpe (0.87), outperforming equal weighting (0.78) and risk parity (0.79). However, factor decomposition reveals this is largely due to market exposure ( $\beta$ ) rather than residual alpha.
- Interest Rates & FX: Time-series momentum remains the dominant strategy; learned models struggle to beat TSMOM in these regimes.
- Agriculture: The Transformer shows the most significant improvement, more than doubling the Sharpe of the best long-only strategy (0.34 vs. 0.16).
Statistical Significance: After Bonferroni correction for multiple testing, only the equity-index results for the learned models remain statistically distinguishable from zero. The cross-asset Transformer's Sharpe (0.55) is slightly above equal weighting (0.52) but not statistically significant in a direct comparison.

The Impact of Transaction Costs

The most critical finding concerns turnover and cost sensitivity:

Turnover: The Transformer exhibits very low turnover (~~0.02 daily), whereas the LSTM trades aggressively (~~0.17 daily in the cross-asset portfolio).
Net Performance: After applying a 2 bp transaction cost, the LSTM's Sharpe ratio drops significantly (from 0.50 to 0.33 in the cross-asset portfolio), while the Transformer's remains stable (0.55 to 0.54).
Conclusion: The Transformer generates the strongest net policy because it trades less, preserving its gross advantage. The LSTM's higher turnover erodes its alpha once costs are implemented.

Complexity and Robustness

Seed Sensitivity: The Transformer is sensitive to random initialization. Seed averaging (averaging three independently seeded models) is the only enhancement that reliably improves stability and competitiveness.
Complexity: More complex structures (mixture-of-experts, larger feature sets, cross-model ensembles) did not reliably improve out-of-sample performance. The simple, seed-averaged Transformer proved superior to these additions.

Skill vs. Exposure

In the equity index, the Transformer's performance is attributed primarily to risk-efficient directional exposure (sub-unit beta) rather than market-neutral alpha. The LSTM showed some residual alpha in equities, but this was not consistent across other asset classes. No universe showed a residual alpha distinguishable from zero after rigorous testing.

4. Significance and Claims

The paper concludes that the value of end-to-end AI allocation is conditional, not universal.

When it works: End-to-end policies add value where simple rules are weaker (e.g., choppy commodity markets) and where the model can exploit breadth across a large cross-section of contracts (the 16-contract portfolio).
When it fails: The approach is less compelling where naive diversification (equal weighting) or standard trend rules (TSMOM) already capture most available returns (e.g., Interest Rates, FX).
The "Transformer" Advantage: The Transformer architecture is preferred over the LSTM not necessarily for superior raw prediction, but for its low turnover, which allows it to survive realistic transaction costs.
Practitioner Takeaway: The study advocates for benchmarking AI models against the specific simple rules a practitioner would actually run, rather than just against equal weighting. It emphasizes that turnover and cost sensitivity are as critical as gross Sharpe ratios. The results suggest that while AI can produce a useful, low-turnover allocation rule, it does not automatically generate market-neutral alpha and must be evaluated with rigorous attention to transaction costs and seed stability.

The authors frame this as a "conservative test" of the framework in a liquid, daily-frequency setting, noting that higher frequencies or wider universes might offer more signal for these models to exploit.