Paper: arXiv 2509.25456

Authors: Mehmet Caner Qingliang Fan

Abstract

In this review, we provide practical guidance on some of the main machine learning tools used in portfolio weight formation. This is not an exhaustive list, but a fraction of the ones used and have some statistical analysis behind it. All this research is essentially tied to precision matrix of excess asset returns. Our main point is that the techniques should be used in conjunction with outlined objective functions. In other words, there should be joint analysis of Machine Learning (ML) technique with the possible portfolio choice-objective functions in terms of test period Sharpe Ratio or returns. The ML method with the best objective function should provide the weight for portfolio formation. Empirically we analyze five time periods of interest, that are out-sample and show performance of some ML-Artificial Intelligence (AI) methods. We see that nodewise regression with Global Minimum Variance portfolio based weights deliver very good Sharpe Ratio and returns across five time periods in this century we analyze. We cover three downturns, and 2 long term investment spans.

Complexity vs Empirical Score

  • Math Complexity: 7.5/10
  • Empirical Rigor: 6.5/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper uses advanced econometrics (precision matrix estimation, shrinkage estimators, nodewise regression) with heavy formulas and derivations, while also presenting an empirical analysis across multiple time periods with performance metrics like Sharpe Ratio.

Research Flowchart

  flowchart TD
  A["Research Goal: Guide for using ML/AI in portfolio weight formation via Precision Matrix of excess asset returns"] --> B["Key Methodology: <br>1. Select ML Technique<br>2. Choose Objective Function<br>3. Joint Analysis"]
  B --> C["Data: Excess Asset Returns"]
  C --> D["Computational Process: <br>Construct Precision Matrix using <br>Nodewise Regression"]
  D --> E["Portfolio Formation: <br>Apply weights to Global Minimum Variance Portfolio"]
  E --> F["Evaluation: <br>Test Period Sharpe Ratio & Returns"]
  F --> G["Key Findings: <br>Nodewise Regression + GMVP delivers <br>superior Sharpe & Returns across 5 time periods (3 downturns, 2 long spans)"]

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 joint analysis of precision matrix estimation techniques with three portfolio weight objective functions (GMV, MV, MSR).

Source sentence
The main contribution and review of this article is joint analysis of precision matrix estimation techniques paired with portfolio weight formation of these three objective functions.

Assumes 4

  • The error terms u_t are stationary, ergodic, zero‑mean with bounded eigenvalues of their covariance matrix. statistical
    Source sentence
    Assumption 4.1. (i). ut is stationary and ergodic such that ut has zero mean and covariance of Σu.
  • The factor process f_t is stationary, ergodic and strong mixing, and independent of u_t. model
    Source sentence
    Assumption 4.3. (i). ft is stationary and ergodic and strong mixing with the same strong mixing coefficient as ut. (ii). ut and ft are independent.
  • The number of factors K grows slower than the cross‑section dimension p and sample size T. model
    Source sentence
    Assumption 4.6. Suppose K = o(p) and K4(logp)2 = o(T), and (logp)2/γ2 −1 = o(T).
  • Transaction costs are a flat 50 basis points per trade. execution
    Source sentence
    use the transaction costs as in Caner et al. (2023), with the transaction costs set to 50 bps.

Data

Backtest sample: 2000 – 2019 · S&P 500 stocks · equities · monthly

  • CRSP S&P 500 stock returns and constituents · CRSP market-prices public
    1985-01 to 2024-12 · S&P 500 constituents
    Source sentence
    The stock returns and historical constituents of S&P 500 stocks are from CRSP in the period covering 1985-01 to 2024-12.
  • Fama‑French 5‑factor data and risk‑free rates · Kenneth R. French – Data Library macro public
    1985-01 to 2024-12
    Source sentence
    The time-series data of Fama-French 5 factors and risk‑free rates are from Kenneth R. French - Data Library.

Costs

Modeled Costs are part of the reported headline results.

Flat 50 basis points per trade (0.005) applied to rebalancing weights.

Source sentence
use the transaction costs as in Caner et al. (2023), with the transaction costs set to 50 bps.

Finds 3

Reported performance (Nodewise regression with Global Minimum Variance (GMV) weighting): Sharpe 0.07

  • Nodewise regression combined with the Global Minimum Variance portfolio delivers very good Sharpe ratios and outperforms the S&P 500 across the five out‑of‑sample periods. performance
    Source sentence
    We see that nodewise regression with Global Minimum Variance portfolio based weights deliver very good Sharpe Ratio and returns across five time periods in this century we analyze.
  • All methods report net returns after subtracting the 50 bps transaction cost, and the reported Sharpe ratios are computed from these net returns. statistical
    Source sentence
    The Sharpe Ratio (SR) is calculated from the average net return and the variance of the net‑of‑cost portfolio returns in the out‑of‑sample period.
  • The joint analysis framework (seven ML‑AI methods × three objectives) identifies the winner as the combination of Nodewise regression with the Maximum Sharpe Ratio portfolio in some periods, but overall Nodewise‑GMV is most robust. performance
    Source sentence
    Nodewise regression of Callot et al. (2021) produces better results generally than other ML‑AI and can beat S&P500.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 23 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.