Paper: arXiv 2503.13544
Authors: Juhyeong Kim, Sungyoon Choi, Youngbin Lee, Yejin Kim, Yongmin Choi, Yongjae Lee
Abstract
We propose Decision by Supervised Learning (DSL), a practical framework for robust portfolio optimization. DSL reframes portfolio construction as a supervised learning problem: models are trained to predict optimal portfolio weights, using cross-entropy loss and portfolios constructed by maximizing the Sharpe or Sortino ratio. To further enhance stability and reliability, DSL employs Deep Ensemble methods, substantially reducing variance in portfolio allocations. Through comprehensive backtesting across diverse market universes and neural architectures, shows superior performance compared to both traditional strategies and leading machine learning-based methods, including Prediction-Focused Learning and End-to-End Learning. We show that increasing the ensemble size leads to higher median returns and more stable risk-adjusted performance. The code is available at https://github.com/DSLwDE/DSLwDE.
Complexity vs Empirical Score
- Math Complexity: 3.0/10
- Empirical Rigor: 9.0/10
- Quadrant: Street Traders — empirical and applied, lower math complexity
Why this score: The paper is mathematically straightforward, using standard supervised learning frameworks and ensemble methods without advanced derivations, but demonstrates high empirical rigor through comprehensive backtests, diverse datasets, and an open-source GitHub repository.
Research Flowchart
flowchart TD
A["Research Goal: Robust Portfolio Optimization via Supervised Learning"] --> B["Methodology: Decision by Supervised Learning DSL Deep Ensemble"]
B --> C{"Input: Historical Market Data"}
C --> D["Compute Target Portfolios<br>Maximizing Sharpe/Sortino Ratio"]
D --> E["Train Ensemble Models<br>Cross-Entropy Loss on Weights"]
E --> F["Ensemble Aggregation<br>Reduce Variance & Improve Stability"]
F --> G["Backtest on Diverse Universes"]
G --> H["Key Findings: Superior Risk-Adjusted Returns<br>Stability Increases with Ensemble Size"]
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 Decision by Supervised Learning (DSL), a practical framework that uses supervised learning and deep ensembles for robust portfolio optimization.
Source sentence
We propose Decision by Supervised Learning (DSL), a practical framework for robust portfolio optimization.
Assumes 6
- Risk‑free rate is set to zero in backtesting. market
Source sentence
we set the risk-free rate to zero within our backtesting environment for simplicity and transparency.
- A two‑day trading lag is applied to all trades. execution
Source sentence
our framework incorporates a two-day trading lag
- Transaction costs of 0.4% per trade (both buy and sell) are applied. execution
Source sentence
applies realistic transaction costs of 0.4% per trade, both for buying and selling.
- The model is restricted to long‑only portfolios. model
Source sentence
Our framework currently employs cross-entropy loss and is designed for long-only portfolios.
Show 2 more
- Cross‑entropy loss is used as a convex surrogate for portfolio weight prediction. model
Source sentence
training models to predict optimal portfolio weights directly using convex surrogate losses—such as cross-entropy—rather than non-convex financial metrics.
- Ensembles of independently trained neural networks reduce predictive variance. statistical
Source sentence
Deep Ensemble methods, substantially reducing variance in portfolio allocations.
Data
- Eight investment universes (Large Cap, Range‑bounded, High Dividend, Utility Sector, NASDA market-prices unknown
Source sentence
We evaluate our framework and all baseline methods across eight distinct investment universes
Costs
Modeled
0.4% per trade, both for buying and selling
Source sentence
applies realistic transaction costs of 0.4% per trade, both for buying and selling.
Finds 3
- DSL consistently outperforms machine‑learning baselines (PFL and E2E) and classical strategies, achieving higher cumulative returns, Sharpe and Sortino ratios across most universes. performance
Source sentence
DSL outperforms both Machine Learning baselines (PFL and E2E) as well as classical strategies (equal-weight, value-weight and mSSRM), achieving higher cumulative returns and better Sharpe and Sortino ratios.
- Larger ensemble sizes improve performance: cumulative returns, Sharpe and Sortino ratios all increase as ensemble size grows from 1 to 64. performance
Source sentence
As ensemble size increases from 1 to 64, cumulative returns steadily improve
- DSL achieves the best performance across all tested objectives in stable universes (Large‑Cap, High‑Dividend) and remains competitive in dynamic rolling universes, especially S&P 500. performance
Source sentence
DSL achieves the best performance across all tested objectives, demonstrating its effectiveness when the stock universe is stable
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 13 related papers). Coverage is partial, so this is neither replication nor refutation.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (35)
- Advances in Financial Machine Learning (2018)
- Decision-Focused Learning: Foundations, State of the Art, Benchmark and Future Opportunities (2024) — link
- Anatomy of Machines for Markowitz: Decision-Focused Learning for Mean-Variance Portfolio Optimization — → 2409.09684
- DF2: Distribution-Free Decision-Focused Learning (2023) — link
- End-to-End Risk Budgeting Portfolio Optimization with Neural Networks (2021) — link
- Mean-variance optimization for asset allocation (2021)
- A Globally Optimal Portfolio for m-Sparse Sharpe Ratio Maximization (2024) — link
- Decision-Focused Learning without Decision-Making: Learning Locally Optimized Decision Losses (2022) — link
and 27 more
- Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles (2017) — link
- Long Short-Term Memory (1997) — link
- Attention is All you Need (2017) — link
- Mamba: Linear-Time Sequence Modeling with Selective State Spaces (2024) — link
- Deep Learning for Portfolio Optimization (2020) — link
- Deep deterministic portfolio optimization (2020) — link
- Deep reinforcement learning for portfolio management of markets with a dynamic number of assets (2021) — link
- Market sentiment-aware deep reinforcement learning approach for stock portfolio allocation (2021) — link
- Model-based Deep Reinforcement Learning for Dynamic Portfolio Optimization (2019) — link
- Deep reinforcement learning for portfolio selection (2024) — link
- Deep reinforcement learning for stock portfolio optimization by connecting with modern portfolio theory (2023) — link
- Deep Reinforcement Learning and Mean-Variance Strategies for Responsible Portfolio Optimization (2024) — link
- Deep Reinforcement Learning for Optimal Portfolio Allocation: A Comparative Study with Mean-Variance Optimization — → 2602.17098
- Multi-task Predict-then-Optimize (2023)
- Ensembles of Classifiers: a Bias-Variance Perspective (2022) — link
- Sophia: A Scalable Stochastic Second-order Optimizer for Language Model Pre-training (2024) — link
- On Nonlinear Fractional Programming (1967) — link
- Mutual Fund Performance (1965) — link
- Performance Measurement in a Downside Risk Framework (1994) — link
- A Universal End-to-End Approach to Portfolio Optimization via Deep Learning (2021) — link
- Pattern Recognition and Machine Learning (Information Science and Statistics) (2006)
- An Overview of Machine Learning for Asset Management. (2023)
- An Overview of Machine Learning for Portfolio Optimization. (2024)
- The effects of errors in means, variances, and correlations on the mean-variance framework (2022)
- Sophia: A Scalable Stochastic Second-order Optimizer for Language Model Pre-training (2024) — link
- GuruAgents: Emulating Wise Investors with Prompt-Guided LLM Agents (2025) — link
- LLM-Enhanced Black-Litterman Portfolio Optimization — → 2504.14345
3 of 35 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.