TL;DR
This paper investigates how Decision-Focused Learning (DFL) fundamentally alters stock return prediction models for Mean-Variance Optimization (MVO). It reveals that DFL, by incorporating the MVO objective into its training, strategically biases predictions: overestimating returns for assets likely to be included in a portfolio and underestimating those likely to be excluded. This "feature, not a flaw" bias explains DFL's superior portfolio performance despite potentially higher prediction errors.
What Problem Does This Paper Address?
The cornerstone of modern portfolio theory, Mean-Variance Optimization (MVO), relies heavily on accurate estimates of expected returns, variances, and covariances. However, these parameters are inherently uncertain in real-world financial markets. While machine learning models are often used to predict these parameters, traditional methods like Mean Squared Error (MSE) treat prediction errors for each asset independently. This can lead to suboptimal investment decisions because MVO is sensitive to the relative accuracy of predictions and the interdependencies between assets. Decision-Focused Learning (DFL) has emerged as a promising approach to address this by integrating the optimization objective directly into the learning process. However, the precise mechanisms by which DFL reshapes prediction models for MVO, and how these changes lead to improved outcomes, remained largely unexplored. This paper aims to bridge this gap by dissecting how DFL modifies prediction models for MVO.
Methods
The core of this paper's contribution lies in its theoretical and empirical analysis of how DFL modifies prediction models for MVO.
- Theoretical Analysis: The authors derive the gradient of the DFL loss function for an unconstrained MVO problem. They show that this gradient inherently includes a term that tilts the prediction errors by the inverse covariance matrix ($\Sigma^{-1}$). This is a crucial departure from MSE, which treats errors independently. The $\Sigma^{-1}$ term explicitly incorporates inter-asset correlations into the learning process, suggesting that DFL learns how assets move together.
- DFL Framework: The paper implements DFL by defining a combined loss function that balances the MVO loss (regret) with the traditional MSE loss. This is achieved using a hyperparameter $\alpha$, where $\alpha=0$ represents pure MSE and $\alpha=1$ represents pure MVO loss. The model is trained end-to-end, allowing gradients from the MVO objective to flow back and influence the prediction model.
- Prediction Model: A standard 4-layer Multi-Layer Perceptron (MLP) is used as the prediction model, taking historical daily returns as input to predict future returns.
Datasets
The experiments were conducted using two well-established asset universes:
- DOW 30: A dataset comprising 30 stocks.
- S&P 100: A dataset comprising 100 stocks.
The data covers the period from January 1, 2010, to December 31, 2024, with daily returns used as input features. Historical covariances were calculated using a 60-day rolling look-back period.
Key Insights and Ideas
- Strategic Bias, Not Error: The most striking insight is that the "higher prediction errors" often associated with DFL are not flaws but rather strategic biases that enhance decision quality. DFL intentionally overestimates returns for assets likely to be included in the portfolio and underestimates those likely to be excluded.
- The Power of $\Sigma^{-1}$: The theoretical derivation highlights the inverse covariance matrix ($\Sigma^{-1}$) as the key mechanism through which DFL reshapes predictions. This term forces the model to learn asset interdependencies, moving beyond independent error minimization.
- Polarization of Predictions: As the influence of the MVO loss (parameter $\alpha$) increases, DFL leads to a polarization of prediction biases. Assets with higher predicted portfolio weights (IN group) receive increasingly positive biases, while those with lower weights (OUT group) receive increasingly negative biases.
- DFL Induces Concentration: In contrast to MSE-trained models that tend to diversify, DFL consistently leads to portfolio concentration, utilizing the minimum number of assets allowed by constraints. This suggests DFL identifies extreme concentration as optimal for MVO performance under certain conditions.
- Trade-off Between Accuracy and Decision Quality: The study confirms a common DFL observation: increasing decision focus (higher $\alpha$) improves portfolio performance (lower regret, higher Sharpe ratio) but often leads to higher MSE loss, highlighting a fundamental trade-off.
Key Results
- Superior Portfolio Performance: DFL models (with $\alpha > 0$) consistently outperformed pure MSE models ($\alpha = 0$) across both DOW 30 and S&P 100 datasets, achieving higher Sharpe ratios and often lower maximum drawdowns.
- Extreme Portfolio Concentration: Table 1 demonstrates that DFL models consistently select the minimum number of assets allowed by constraints, regardless of risk aversion ($\lambda$), whereas MSE models diversify more as $\lambda$ increases.
- Significant Reduction in Regret: Figure 2 shows that DFL models achieve substantially lower portfolio regret on the test set compared to MSE models.
- Predictive Bias Polarization: Figure 3 and Figure 4 visually confirm the polarization of prediction biases. As $\alpha$ increases, the separation between predicted returns for "IN" and "OUT" assets widens, with "IN" assets showing positive bias and "OUT" assets showing negative bias.
- Weight-Dependent Bias: Figure 5 illustrates that DFL creates a clear distinction in prediction bias based on an asset's average portfolio weight. High-weight assets exhibit strong positive biases, while low-weight assets exhibit strong negative biases.
Limitations and Future Work
- Unconstrained MVO Focus: The theoretical analysis primarily focused on the unconstrained MVO problem. Extending this to constrained optimization settings (e.g., budget limits, turnover restrictions) is a natural next step for practical relevance.
- Granular Asset Analysis: While group-level biases were analyzed, a more granular understanding of which individual assets are most susceptible to strategic bias could inform more refined model architectures.
- Alternative DFL Approaches: The paper suggests exploring simpler alternatives, such as a covariance-weighted MSE loss, to see if similar benefits can be achieved with potentially more computationally efficient hybrid methods.
- Understanding $\lambda$ Interaction: While the interaction with risk aversion ($\lambda$) was observed, further research could delve deeper into how DFL's strategic biases adapt to different levels of risk aversion.
Summary updated