Decision-Informed Neural Networks with Large Language Model Integration for Portfolio Optimization

In one sentence

Summary

The paper addresses the critical disconnect between prediction accuracy and decision quality in portfolio optimization, arguing that minimizing mean squared error alone leads to suboptimal portfolio weights due to estimation sensitivity. It positions its approach as a bridge between advanced representation learning and decision-focused learning, leveraging the representational power of Large Language Models (LLMs) to capture complex market dynamics that traditional statistical methods miss. The core problem is that conventional two-stage processes fail to account for how prediction errors propagate into portfolio construction, often resulting in poor out-of-sample performance despite high in-sample predictive accuracy. The proposed Decision-Informed Neural Network (DINN) architecture integrates LLM-based semantic embeddings with time-series data through a cross-attention mechanism. Input embeddings are generated by converting pairwise asset relationships, sector-level yields, and irregularly sampled macroeconomic variables into textual prompts, which are then embedded using a pretrained LLM. These embeddings are fused with normalized and decomposed return series (separating long-term trends from short-term dynamics) via cross-attention layers. The model outputs predicted returns and covariance factors, which are fed into a differentiable convex optimization layer. This layer computes optimal portfolio weights end-to-end, allowing gradients from the portfolio performance loss to backpropagate through the optimization layer to the forecasting network, thereby aligning prediction objectives with decision objectives. Extensive experiments on S&P100 and DOW30 datasets demonstrate that DINN consistently outperforms state-of-the-art deep learning baselines in terms of Sharpe ratio, Sortino ratio, and terminal wealth. Gradient-based analyses reveal that the model prioritizes forecasting accuracy for assets with high sensitivity to portfolio weights, effectively mitigating the impact of estimation errors on decision quality. The findings underscore the value of integrating decision objectives into the learning process, showing that context-aware, decision-focused models are more robust to market regime shifts and non-stationary relationships than purely predictive models. Limitations include the computational overhead of LLM embedding generation and the reliance on specific risk-aversion parameters for the optimization layer.

Research problem

The paper addresses the fundamental disconnect between prediction accuracy and decision quality in portfolio optimization. Traditional methods minimize prediction error (e.g., MSE) in a two-stage process, which fails to account for the sensitivity of optimal portfolio weights to estimation errors, often leading to suboptimal investment decisions. The problem is to develop a framework that directly aligns predictive modeling with portfolio optimization objectives, leveraging the rich representational capabilities of Large Language Models to capture complex, non-linear market dynamics and macroeconomic contexts that traditional models miss.

Core contributions

• Proposes a Decision-Informed Neural Network (DINN) framework that integrates LLMs with decision-focused learning for portfolio optimization.
• Introduces an attention mechanism that processes LLM-enhanced semantic embeddings of asset relationships, temporal dependencies, and macroeconomic variables.
• Develops a hybrid loss function that combines statistical prediction accuracy with decision-focused portfolio performance metrics.
• Demonstrates theoretically and empirically that minimizing prediction error alone leads to suboptimal portfolio decisions.
• Shows that gradient-based analysis reveals the model prioritizes assets crucial to decision-making, mitigating the effects of prediction errors.

Method

• Input embedding process using LLMs to encode pairwise asset relationships, sector yields, and macroeconomic variables into textual prompts.
• Time-series normalization and multi-scale decomposition to separate long-term trends from short-term dynamics.
• Cross-attention mechanism to fuse LLM embeddings with decomposed time-series features.
• Differentiable convex optimization layer to compute optimal portfolio weights from predicted returns and covariance factors.
• Hybrid loss function combining prediction loss (MSE) and decision-focused loss (portfolio performance metrics).
• End-to-end training allowing gradients from the optimization layer to update the forecasting network.

Results & findings

• DINN consistently outperforms state-of-the-art deep learning models on S&P100 and DOW30 datasets.
• The model achieves higher Sharpe and Sortino ratios, indicating better risk-adjusted returns.
• Gradient analysis shows that assets with higher sensitivity to portfolio weights have lower forecasting errors, confirming decision-focused prioritization.
• The approach mitigates the negative impact of estimation errors on portfolio performance compared to prediction-only models.
• LLM embeddings effectively capture inter-asset and macroeconomic contexts, enhancing the model's ability to handle complex market dynamics.
• The hybrid loss function successfully bridges the gap between statistical accuracy and economic meaningfulness.

Data & experiments

• Experiments conducted on S&P100 and DOW30 datasets covering multiple market regimes including COVID-19 and various macroeconomic indicators.
• Comparison against state-of-the-art deep learning models for time-series forecasting and portfolio optimization.
• Evaluation metrics include Annualized Return, Standard Deviation, Sharpe Ratio, Sortino Ratio, Maximum Drawdown, Value-at-Risk, and Terminal Wealth.
• Gradient-based sensitivity analysis to examine how prediction errors correlate with asset importance in portfolio construction.
• Hyperparameter sensitivity analysis for risk-aversion parameters in the optimization layer.

Limitations & caveats

• The computational overhead of generating LLM embeddings for pairwise asset relationships and macroeconomic variables may be significant.
• The model's performance is sensitive to the choice of risk-aversion parameter in the optimization layer.
• The approach relies on the quality of LLM embeddings, which may vary depending on the pretrained model and prompt design.
• The study focuses on large-cap indices (S&P100, DOW30), and generalizability to smaller assets or other markets is not fully explored.
• The differentiable optimization layer assumes convexity, which may limit applicability to more complex portfolio constraints.

Curator notes

Important distinction between prediction accuracy and portfolio utility.