Cross-Asset Attention Networks

• Cross-Asset Attention Networks are neural modules that model inter-asset relationships, addressing selection bias in quantitative trading strategies.
• They integrate LSTM-based per-asset embeddings with a scaled dot-product self-attention mechanism to generate context-aware winner scores for asset ranking.
• Empirical evaluations show that CAAN improves portfolio Sharpe ratios by 20–30% and reduces maximum drawdown, ensuring robust long-short strategies.

A Cross-Asset Attention Network (CAAN) is a neural network module designed to model and exploit interrelationships among assets in portfolio management, specifically to address the challenge of selection bias in quantitative trading strategies. Unlike traditional frameworks that evaluate each asset in isolation, CAAN introduces a mechanism by which every asset’s representation is directly compared, via learned attention weights, to all others in the cross-section. This context-aware modeling enables the generation of robust “winner” scores reflective of each asset’s relative standing, foundational for buying-winners-and-selling-losers (BWSL) strategies in financial markets (Wang et al., 2019).

1. Architecture and Integration within AlphaStock

The CAAN architecture operates as the central relational layer within the AlphaStock reinforcement learning pipeline. AlphaStock comprises three primary modules:

1. Per-Asset Embedding: For each asset $i$, a Long Short-Term Memory with History Attention (LSTM-HA) encodes a K-period look-back of engineered features including return, volatility, volume, price/earnings ratio, and more. The output is a fixed-length embedding $\mathbf{r}^{(i)} \in \mathbb{R}^{D_r}$.
2. Cross-Asset Attention Network: The set $\{\mathbf{r}^{(1)},\ldots,\mathbf{r}^{(I)}\}$ (for $I$ assets) is jointly processed. CAAN produces for each asset a scalar “winner score” $s^{(i)} \in (0,1)$, reflecting its relative potential.
3. Portfolio Generator: Assets are ranked by $s^{(i)}$, with allocations constructed to long the top-G and short the bottom-G assets, satisfying a zero-investment (market-neutral) constraint. This allocation determines the agent’s trading policy $\pi_\theta$, trained via policy-gradient reinforcement learning to directly optimize a risk-adjusted return, measured by Sharpe ratio.

CAAN thus forms the core relational comparator, receiving feature embeddings and outputting context-dependent scores that drive the explicit long/short composition of the portfolio (Wang et al., 2019).

2. Mathematical Mechanism of Cross-Asset Attention

CAAN implements a scaled dot-product self-attention mechanism, optionally enhanced with a rank-based prior reflecting similarity of recent returns. The procedure is as follows (time index omitted for clarity):

1. Linear Projections:

$$Q^{(i)} = W^{(Q)} r^{(i)} \in \mathbb{R}^{D_q}, \quad K^{(i)} = W^{(K)} r^{(i)} \in \mathbb{R}^{D_k}, \quad V^{(i)} = W^{(V)} r^{(i)} \in \mathbb{R}^{D_v}$$

Each asset embedding is projected to query, key, and value vectors.

1. Scaled Dot-Product with Optional Rank Prior:
   • For each pair $(i, j)$, compute base compatibility:

   $$\beta^0_{ij} = \frac{Q^{(i)\top} K^{(j)}}{\sqrt{D_k}}$$

• Optionally, for rank-based prior, calculate

  $\mathbf{r}^{(i)} \in \mathbb{R}^{D_r}$0

  where $\mathbf{r}^{(i)} \in \mathbb{R}^{D_r}$1 is the last-period return rank. Retrieve an embedding $\mathbf{r}^{(i)} \in \mathbb{R}^{D_r}$2 and process:

  $\mathbf{r}^{(i)} \in \mathbb{R}^{D_r}$3

  The final unnormalized attention:

  $\mathbf{r}^{(i)} \in \mathbb{R}^{D_r}$4

  else set $\mathbf{r}^{(i)} \in \mathbb{R}^{D_r}$5 if no prior.

1. Attention Weights:

$\mathbf{r}^{(i)} \in \mathbb{R}^{D_r}$6

2. Context Vector:

$\mathbf{r}^{(i)} \in \mathbb{R}^{D_r}$7

3. Winner-Score Head:

$\mathbf{r}^{(i)} \in \mathbb{R}^{D_r}$8

CAAN is implemented as single-head self-attention with typical projection dimensions $\mathbf{r}^{(i)} \in \mathbb{R}^{D_r}$9, scaling as $\{\mathbf{r}^{(1)},\ldots,\mathbf{r}^{(I)}\}$0 in line with standard attention practice (Wang et al., 2019).

3. Input Feature Construction and Embedding

Each asset is mapped from raw historical data to a dense representation as follows:

• Feature Engineering: Each asset at time $\{\mathbf{r}^{(1)},\ldots,\mathbf{r}^{(I)}\}$1 is represented over a K-month rolling window, yielding a temporal sequence $\{\mathbf{r}^{(1)},\ldots,\mathbf{r}^{(I)}\}$2 with $\{\mathbf{r}^{(1)},\ldots,\mathbf{r}^{(I)}\}$3 cross-sectionally Z-scored features (monthly return, volatility, volume, market capitalization, P/E, B/M, dividend yield).
• Temporal Encoding and History Attention: This sequence is processed by an LSTM with hidden size $\{\mathbf{r}^{(1)},\ldots,\mathbf{r}^{(I)}\}$4, yielding a series of hidden states $\{\mathbf{r}^{(1)},\ldots,\mathbf{r}^{(I)}\}$5. History attention computes a convex combination:

$\{\mathbf{r}^{(1)},\ldots,\mathbf{r}^{(I)}\}$6

where $\{\mathbf{r}^{(1)},\ldots,\mathbf{r}^{(I)}\}$7, $\{\mathbf{r}^{(1)},\ldots,\mathbf{r}^{(I)}\}$8, and $\{\mathbf{r}^{(1)},\ldots,\mathbf{r}^{(I)}\}$9 are trainable parameters. The output $I$0 is the fixed-size embedding for asset $I$1.

This design ensures that the CAAN operates on context-rich, temporally informed representations.

4. Mitigating Selection Bias and Modeling Asset Interrelationships

CAAN addresses the selection bias endemic to per-asset evaluation by explicitly calculating for each asset $I$2 a score $I$3 that aggregates information from all other assets’ value vectors $I$4. Key properties include:

• Cross-sectional Comparisons: The attention mechanism enables each asset to “query” the latent factors of all others, forcing relative, rather than absolute, evaluation.
• Competitive and Hedging Effects: Attention weights $I$5 upweight assets with similar recent performance or those that may serve as hedges.
• Concentration Control: Assets that are latent analogs of many others may be down-weighted, promoting diversification.
• Relative Scoring: Winner scores embody not just absolute but relative standing, reducing propensity to select assets that are local maxima in isolation but poorly ranked in the global cross-section.

This relational encoding enables the model to systematically avoid pitfalls associated with market regime changes, such as inadvertently “buying losers” during slumps (Wang et al., 2019).

5. Implementation Hyperparameters and Practical Details

Salient implementation aspects are summarized below:

Component	Typical Value/Setting	Remarks
Attention Heads	Single	Multi-head possible
Projection Dimensions	$I$6
Scaling Factor	$I$7	As in canonical transformers
Rank Prior Bin Width $I$8	10 (tunable)	Via validation
Rank Prior Embedding$I$9	8–16	Learned embedding table
Dropout	0.1–0.3	On keys/values or context vectors
Winner Score Head	$s^{(i)} \in (0,1)$0	Dimensionality matches value projection

CAAN is trained end-to-end with the policy-gradient RL agent, leveraging differentiability of the entire architecture for Sharpe ratio maximization (Wang et al., 2019).

6. Portfolio Generation and Reinforcement Learning

Post-attention, the assets are allocated as follows:

1. Scores $s^{(i)} \in (0,1)$1 are sorted.
2. Top-G assets are allocated to the long book with weights

$s^{(i)} \in (0,1)$2

and bottom-G to the short book with

$s^{(i)} \in (0,1)$3

1. The resulting zero-investment vector $s^{(i)} \in (0,1)$4 parameterizes $s^{(i)} \in (0,1)$5, the asset-level buy/sell policy.
2. The reward for the RL agent is the realized Sharpe ratio over episode length $s^{(i)} \in (0,1)$6:

$s^{(i)} \in (0,1)$7

With baseline $s^{(i)} \in (0,1)$8, a REINFORCE-style update is performed:

$s^{(i)} \in (0,1)$9

This process ensures end-to-end training of CAAN parameters to maximize portfolio-wide risk-adjusted return (Wang et al., 2019).

7. Empirical Impact and Significance

By integrating cross-asset awareness directly into the scoring, CAAN enables regime-robust winner identification. Empirically, this leads to a realized Sharpe ratio improvement of 20–30% over non-attention baselines, with maximum drawdown reduced by half. These effects are attributed to the avoidance of mis-ranking assets under turbulent conditions, demonstrating robust gains in both risk and return dimensions (Wang et al., 2019). A plausible implication is that attention-based relational modeling offers a scalable, generalizable framework adaptable beyond BWSL strategies, wherever cross-sectional dependencies are salient.

Reference: See "AlphaStock: A Buying-Winners-and-Selling-Losers Investment Strategy using Interpretable Deep Reinforcement Attention Networks" (Wang et al., 2019) for the canonical implementation and analysis of Cross-Asset Attention Networks.