Paper: arXiv 2409.09684
Authors: Junhyeong Lee, Haeun Jeon, Hyunglip Bae, Yongjae Lee
Abstract
Markowitz laid the foundation of portfolio theory through the mean-variance optimization (MVO) framework. However, the effectiveness of MVO is contingent on the precise estimation of expected returns, variances, and covariances of asset returns, which are typically uncertain. Machine learning models are becoming useful in estimating uncertain parameters, and such models are trained to minimize prediction errors, such as mean squared errors (MSE), which treat prediction errors uniformly across assets. Recent studies have pointed out that this approach would lead to suboptimal decisions and proposed Decision-Focused Learning (DFL) as a solution, integrating prediction and optimization to improve decision-making outcomes. While studies have shown DFL’s potential to enhance portfolio performance, the detailed mechanisms of how DFL modifies prediction models for MVO remain unexplored. This study investigates how DFL adjusts stock return prediction models to optimize decisions in MVO. Theoretically, we show that DFL’s gradient can be interpreted as tilting the MSE-based prediction errors by the inverse covariance matrix, effectively incorporating inter-asset correlations into the learning process, while MSE treats each asset’s error independently. This tilting mechanism leads to systematic prediction biases where DFL overestimates returns for assets included in portfolios while underestimating excluded assets. Our findings reveal why DFL achieves superior portfolio performance despite higher prediction errors. The strategic biases are features, not flaws.
Complexity vs Empirical Score
- Math Complexity: 8.5/10
- Empirical Rigor: 4.0/10
- Quadrant: Lab Rats — high math complexity, limited empirical testing
Why this score: The paper is mathematically dense with detailed derivations of gradients, inverse covariance matrices, and optimization theory, placing it in the high math category. While it presents empirical results (like Table 1), the summary suggests a focus on theoretical mechanisms rather than extensive backtesting or implementation-heavy datasets, aligning with lab-oriented research.
Research Flowchart
flowchart TD A["Research Goal<br>How does DFL modify prediction models<br>to optimize MVO decisions?"] --> B["Data & Inputs<br>Asset return data for MVO<br>ML model for return prediction"] B --> C["Methodology<br>Compare: MSE-based (baseline) vs.<br>Decision-Focused Learning (DFL)"] C --> D["Computational Process<br>Backpropagation: Theoretical analysis<br>of gradients via chain rule"] D --> E["Key Finding 1: Gradient Mechanism<br>MSE Error ∝ Inverse Covariance<br>(Incorrelates inter-asset correlation)"] D --> F["Key Finding 2: Strategic Biasing<br>DFL systematically overestimates<br>returns for included assets"] E --> G["Outcome<br>DFL yields superior portfolio performance<br>despite higher MSE prediction errors"] F --> G style A fill:#e1f5fe style G fill:#f1f8e9
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes a Decision-Focused Learning (DFL) approach that adjusts stock return prediction models to improve mean-variance portfolio decisions.
Source sentence
This study investigates how DFL adjusts stock return prediction models to optimize decisions in MVO.
Assumes 5
- The analysis focuses on the unconstrained mean-variance optimization problem. model
Source sentence
Our theoretical analysis focused on the unconstrained MVO problem, which allowed us to derive closed-form solutions and clear insights.
- Sample covariance matrices computed from historical daily returns are used as the true covariance for MVO. data
Source sentence
For covariances, we use the sample covariance matrix ˆΣ calculated using historical data.
- Expected returns are predicted from a 60‑day look‑back window of daily returns. data
Source sentence
Prediction models were trained to predict the returns of
- Daily rolling forecasts are used for both expected returns and covariances. execution
Source sentence
Both the prediction of expected returns and the covariances are calculated on a daily rolling basis.
Show 1 more
- No transaction costs or market impact are incorporated in the reported portfolio performance. execution
Source sentence
Our results demonstrate that intermediate α values consistently outperform both extremes.
Data
2010-01 – 2024-12 · US equities (DOW 30) · equities · daily
- DOW 30 market-prices public
Source sentence
For the dataset, we used the DOW30 and S&P100 asset universes, covering the period from 01-Jan-2010 to 31-Dec-2024.
- S&P 100 market-prices public
Source sentence
For the dataset, we used the DOW30 and S&P100 asset universes, covering the period from 01-Jan-2010 to 31-Dec-2024.
Costs
Not modeled
Finds 5
Sharpe 1.30
- Models trained with DFL (α>0) consistently achieve higher returns, Sharpe ratios and lower maximum drawdowns than pure MSE models (α=0). performance
Source sentence
Models with α> 0 consistently outperform pure MSE (α= 0), with optimal performance typically at intermediate values (α∈[0.25, 0.75]).
- DFL induces systematic prediction biases: it overestimates returns for assets included in the portfolio and underestimates those excluded. predictive-accuracy
Source sentence
This tilting mechanism leads to systematic prediction biases where DFL overestimates returns for assets included in portfolios while underestimating excluded assets.
- The DFL gradient can be interpreted as tilting the MSE error vector by the inverse covariance matrix, incorporating inter‑asset correlations into learning. theoretical
Source sentence
DFL's gradient can be interpreted as tilting the MSE-based prediction errors by the inverse covariance matrix, effectively incorporating inter-asset correlations into the learning process.
- Increasing the strength of the MVO loss (higher α) raises the MSE loss but improves decision quality, confirming the trade‑off between prediction accuracy and portfolio regret. statistical
Source sentence
Figure 2 shows that increasing the strength of the MVO loss leads to a consistent rise in the MSE loss. This trend is particularly
- The risk‑aversion parameter λ behaves as expected: higher λ reduces portfolio drawdowns and volatility, with the most pronounced effects at λ=5.0. performance
Source sentence
The risk aversion parameter λ exhibits expected behavior across all models. Increasing λ reduces portfolio drawdowns and volatility, with the most pronounced effects at λ= 5.0.
Checked by others
Supported by
- Decision-Induced Ranking Explains Prediction Inflation and Excessive Turnover in SPO-Based Portfolio Optimization
Compare the findings
This paper: DFL induces systematic prediction biases: it overestimates returns for assets included in the portfolio and underestimates those excluded.
That paper: SPO‑trained portfolios produce highly inflated return predictions and excessive turnover across all datasets.
4 papers in the archive
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.