Paper: arXiv 2601.00593
Authors: Yan Liu, Ye Luo, Zigan Wang, Xiaowei Zhang
Abstract
Machine learning is central to empirical asset pricing, but portfolio construction still relies on point predictions and largely ignores asset-specific estimation uncertainty. We propose a simple change: sort assets using uncertainty-adjusted prediction bounds instead of point predictions alone. Across a broad set of ML models and a U.S. equity panel, this approach improves portfolio performance relative to point-prediction sorting. These gains persist even when bounds are built from partial or misspecified uncertainty information. They arise mainly from reduced volatility and are strongest for flexible machine learning models. Identification and robustness exercises show that these improvements are driven by asset-level rather than time or aggregate predictive uncertainty.
Complexity vs Empirical Score
- Math Complexity: 7.0/10
- Empirical Rigor: 9.0/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper employs advanced statistical learning concepts like prediction intervals, quantile estimation, and cross-sectional uncertainty attribution, which requires a high level of mathematical sophistication. However, it is grounded in extensive empirical implementation, testing across a broad set of ML models on a large U.S. equity panel with rigorous robustness checks, transaction cost analysis, and detailed performance metrics.
Research Flowchart
flowchart TD A["Research Goal:<br>Can uncertainty-adjusted sorting<br>improve ML portfolio performance?"] B["Data & Inputs<br>US Equity Panel<br>ML Model Predictions"] C["Methodology:<br>Construct Uncertainty-Adjusted<br>Prediction Bounds"] D["Compute Portfolios<br>Sort assets via:<br>1. Point Predictions<br>2. Uncertainty-Adjusted Bounds"] E["Key Findings<br>1. Improved Risk-Adjusted Returns<br>2. Driven by Asset-Level Uncertainty<br>3. Strongest with Flexible ML Models"] A --> B B --> C C --> D D --> E
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 sorting assets using uncertainty‑adjusted prediction bounds instead of point predictions alone.
Source sentence
We propose a simple change: sort assets using uncertainty‑adjusted prediction bounds instead of point predictions alone.
Assumes 5
- Exchangeability of observations for conformal prediction guarantees. statistical
Source sentence
Under an exchangeability assumption, these procedures deliver exact finite‑sample marginal coverage, regardless of the complexity of the underlying prediction algorithm.
- Predictive intervals are constructed without imposing parametric assumptions on the residual distribution. model
Source sentence
These bounds are obtained using rolling historical prediction residuals, without imposing parametric assumptions on the full residual distribution.
- Long‑short portfolios are formed by sorting stocks into deciles each month and weighting equally within each leg. execution
Source sentence
Each month, stocks are sorted into deciles based on predicted returns; the portfolio goes long the top decile and short the bottom decile, with equal weighting within each leg.
- Transaction costs are modeled as a proportional cost of 20 basis points per trade. execution
Source sentence
Following Avramov et al. (2023), we impose a proportional trading cost of 20 basis points per trade.
Show 1 more
- The sample covers the full CRSP common‑stock universe from January 1967 to December 2016. data
Source sentence
The monthly sample covers January 1967 to December 2016 and includes the full CRSP common‑stock universe.
Data
1992 – 2016 · US stocks · equities · monthly
- CRSP stock returns market-prices public
Source sentence
Stock returns CRSP Monthly 1967–2016 Compute next‑month excess returns using the 1M T‑bill rate.
- Firm characteristics fundamentals public
Source sentence
46 Firm characteristics CRSP/Compustat; CRSP Monthly / Annually 1967–2016 Annual variables updated at end of June; monthly variables updated at month‑end; cross‑sectional quantile transform each month.
- Macroeconomic variables macro public
Source sentence
178 Macroeconomic variables FRED‑MD; medians of characteristics; Welch and Goyal (2008) Monthly 1967–2016 Apply standard transformations; expand to panel by assigning the same macro vector to all stocks each month.
Costs
Modeled
A proportional cost of 20 basis points per trade is applied to all portfolio returns.
Source sentence
Following Avramov et al. (2023), we impose a proportional trading cost of 20 basis points per trade.
Finds 5
Sharpe 2.14 · annual return 39.5%
- Uncertainty‑adjusted sorting yields portfolios that systematically outperform point‑prediction sorting in Sharpe ratio and volatility control. performance
Source sentence
Empirically, the resulting portfolios systematically outperform their point‑prediction counterparts in terms of Sharpe ratios and volatility control.
- Uncertainty‑adjusted portfolios reduce raw return spreads but substantially lower volatility, improving Sharpe ratios for most models. performance
Source sentence
while uncertainty‑adjusted portfolio construction typically reduces raw return spreads relative to point‑prediction sorting, it substantially lowers portfolio volatility. As a result, uncertainty‑adjusted performance—as measured by Sharpe ratios and statistical significance—improves for the majority of models.
- XGBoost point‑prediction long‑short portfolio delivers 39.45% annual return, 18.40% volatility, Sharpe 2.14; after 20 bps costs, return is 32.41% with Sharpe 1.76. performance
Source sentence
The XGBoost long‑short portfolio delivers an annualized return of 39.45% with annualized volatility of 18.40%, corresponding to a Sharpe ratio of 2.14
- Elastic Net and PLS remain robust after transaction costs, maintaining Sharpe ratios around 1.29 and significant t‑statistics. performance
Source sentence
Both models continue to deliver economically meaningful returns, with Sharpe ratios around 1.29 and Newey–West t‑statistic above 6.
- Deeper neural‑network architectures (NN4, NN5) perform poorly and become marginal after transaction costs. negative
Source sentence
deeper neural‑network architectures struggle to generate robust benchmark‑relative performance once trading frictions are introduced.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 40 related papers). Coverage is partial, so this is neither replication nor refutation.
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.