arxiv: 2604.22933 · v1 · submitted 2026-04-24 · 💱 q-fin.PR · econ.EM

Recognition: unknown

Machine Learning Forecasts of Asymmetric Betas Using Firm-Specific Information

Iason Kynigakis, John Cotter, Thomas Conlon

Pith reviewed 2026-05-08 08:50 UTC · model grok-4.3

classification 💱 q-fin.PR econ.EM

keywords asymmetric betamachine learningconditional betafirm characteristicsrisk forecastingequity valuationmarket-neutral strategiesUS stocks

0 comments

The pith

Machine learning methods using firm characteristics forecast asymmetric betas more accurately by capturing nonlinear effects.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper argues that machine learning provides a superior way to forecast conditional asymmetric risk in stocks by incorporating nonlinear relationships between firm characteristics and future betas. The improvement in out-of-sample performance matters for investors because it leads to more accurate market beta estimates, better equity valuations in cash flow models, and real economic gains in market-neutral portfolios. Trading frictions and characteristics tied to intangibles, momentum, and growth turn out to be the strongest predictors of these risk changes. A reader would care if they rely on beta for pricing or hedging decisions.

Core claim

Machine learning methods provide a powerful framework for modelling conditional asymmetric risk. Using a large cross-section of US stocks and firm characteristics, allowing for nonlinearities significantly increases out-of-sample performance across asymmetric beta measures and horizons. Trading frictions, intangibles, momentum and growth are the most important drivers. Reconstructing CAPM beta from asymmetric components gives a more accurate representation, and incorporating forecasts into DCF models enhances valuation while delivering benefits to market-neutral investors.

What carries the argument

Asymmetric beta components forecasted via machine learning on firm-specific information, which capture nonlinear conditional risk dynamics better than linear alternatives.

If this is right

Improved forecasts lead to more accurate reconstruction of standard CAPM beta.
Conditional beta forecasts improve accuracy of equity valuations in discounted cash flow models accounting for beta term structure.
Market-neutral portfolio strategies achieve economically significant gains from using the conditional forecasts.
Trading frictions and intangibles-related characteristics are key drivers of future asymmetric risk.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Similar machine learning approaches might improve risk forecasts in other asset classes or markets if the firm characteristics generalize.
The identified importance of trading frictions suggests potential links to liquidity risk that could be explored further in separate models.

Load-bearing premise

The firm characteristics chosen and the machine learning algorithms accurately identify true conditional asymmetric risk patterns without substantial overfitting to the US stock data.

What would settle it

Out-of-sample tests showing that linear regression models match or exceed the machine learning performance on asymmetric beta forecasts would disprove the advantage of allowing for nonlinearities.

read the original abstract

We demonstrate that machine learning methods provide a powerful framework for modelling conditional asymmetric risk. Using a large cross-section of US stocks and a comprehensive set of firm characteristics, we show that allowing for nonlinearities significantly increases the out-of-sample performance across a wide range of asymmetric beta measures and forecasting horizons. Trading frictions, followed by characteristics related to intangibles, momentum and growth, emerge as the most important drivers of future risk dynamics. Reconstructing CAPM beta from forecasts of asymmetric beta components indicates that a more granular decomposition of systematic risk yields a more accurate representation of market beta. We also find that incorporating conditional beta forecasts into discounted cash flow models that account for the term structure of betas enhances equity valuation accuracy. Finally, we show that the statistical outperformance of conditional betas translates into economically significant benefits for market-neutral portfolio investors.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

ML improves asymmetric beta forecasts out of sample but the broad search over characteristics, beta variants, and models leaves a clear data-snooping risk.

read the letter

The paper finds that nonlinear machine learning models using firm characteristics forecast asymmetric betas more accurately than linear benchmarks across horizons, and that these forecasts improve reconstructed market betas and DCF valuation accuracy while delivering economic gains in market-neutral portfolios. Trading frictions and intangibles rank high in importance. That combination of asymmetric focus, large US stock panel, and valuation link is the concrete step forward here. It moves past generic ML-beta papers by showing the decomposition matters for downstream uses. The results on out-of-sample gains and portfolio benefits are the parts that could interest practitioners in risk modeling. The soft spot is the multiple-testing exposure. A comprehensive characteristic set plus several asymmetric beta definitions plus a menu of ML architectures creates many implicit comparisons in a panel of thousands of stocks. Without pre-specification or explicit adjustments, the reported edge can easily reflect search rather than stable signal. The abstract flags the breadth, so this is not a minor footnote. I would want to see whether the gains survive a narrow pre-chosen predictor list or formal multiplicity controls. This work is for empirical asset-pricing people who care about conditional risk and valuation applications. Readers running beta-based models or DCF exercises could extract usable ideas even if they disagree on the exact ML setup. It is coherent enough on its own terms to deserve referee time rather than a desk reject, though revisions on robustness will be needed. Send it out, but flag the snooping issue for the first round.

Referee Report

2 major / 2 minor

Summary. The manuscript claims that machine learning methods, by allowing for nonlinearities, deliver statistically and economically significant out-of-sample improvements in forecasting a wide range of asymmetric beta measures for US stocks, using a comprehensive set of firm characteristics. Key drivers identified include trading frictions, intangibles, momentum, and growth. The paper further shows that forecasts of asymmetric beta components yield more accurate CAPM beta reconstructions, improve discounted cash flow valuations that incorporate beta term structure, and generate economic gains for market-neutral portfolio strategies.

Significance. If the central empirical results hold after robustness checks, the work would meaningfully advance conditional asset pricing by demonstrating that nonlinear ML can capture risk dynamics missed by linear models, with direct implications for risk management and equity valuation. Strengths include the large cross-section, focus on economic significance beyond statistical metrics, and the granular decomposition of systematic risk.

major comments (2)

[Abstract and Results section] Abstract and Results section: The headline claim that 'allowing for nonlinearities significantly increases the out-of-sample performance across a wide range of asymmetric beta measures' is load-bearing for the paper's contribution, yet the reported gains are obtained after searching over a comprehensive (hence high-dimensional) set of firm characteristics, multiple asymmetric-beta definitions, forecasting horizons, and ML architectures/hyperparameters. No multiplicity adjustment or pre-specification protocol is described, which risks data-snooping bias in a panel of thousands of stocks and could render the statistical and economic significance non-robust.
[Methodology and Results sections] Methodology and Results sections: The weakest assumption—that the selected characteristics and ML models capture true conditional asymmetric risk dynamics without substantial overfitting—is not directly tested via, for example, restricted predictor sets, placebo characteristics, or explicit out-of-sample validation on pre-specified subsets. This is central because the abstract emphasizes the 'comprehensive set' and 'wide range,' making it unclear whether the performance edge survives when the search space is curtailed.

minor comments (2)

[Abstract] Abstract: The precise definitions of the asymmetric beta measures and the exact ML algorithms (e.g., random forests, neural nets, gradient boosting) are not stated, which would aid immediate readability.
[Data section] The paper would benefit from an explicit table or section listing all firm characteristics used, their sources, and any winsorization or standardization steps.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on data-snooping risks and the need for explicit robustness tests against overfitting. We have revised the manuscript to incorporate multiplicity adjustments, pre-specified predictor subsets, and placebo checks, which preserve the central finding that nonlinear ML improves asymmetric beta forecasts.

read point-by-point responses

Referee: [Abstract and Results section] Abstract and Results section: The headline claim that 'allowing for nonlinearities significantly increases the out-of-sample performance across a wide range of asymmetric beta measures' is load-bearing for the paper's contribution, yet the reported gains are obtained after searching over a comprehensive (hence high-dimensional) set of firm characteristics, multiple asymmetric-beta definitions, forecasting horizons, and ML architectures/hyperparameters. No multiplicity adjustment or pre-specification protocol is described, which risks data-snooping bias in a panel of thousands of stocks and could render the statistical and economic significance non-robust.

Authors: We acknowledge that the scope of specifications creates a legitimate multiple-testing concern. While our core model choices were informed by prior asset-pricing literature on firm characteristics, we did not apply formal corrections in the original draft. In the revision we add Bonferroni- and Holm-adjusted p-values for the headline out-of-sample R² and economic-value metrics across the main tables. We also report results for a pre-specified subset of 20 characteristics drawn from earlier beta and risk studies; the nonlinear advantage remains statistically and economically significant under this narrower search space. revision: yes
Referee: [Methodology and Results sections] Methodology and Results sections: The weakest assumption—that the selected characteristics and ML models capture true conditional asymmetric risk dynamics without substantial overfitting—is not directly tested via, for example, restricted predictor sets, placebo characteristics, or explicit out-of-sample validation on pre-specified subsets. This is central because the abstract emphasizes the 'comprehensive set' and 'wide range,' making it unclear whether the performance edge survives when the search space is curtailed.

Authors: We agree that direct tests of the overfitting hypothesis are warranted. The revised manuscript adds three sets of checks: (i) restricted-predictor regressions using only the four economically motivated groups highlighted in the paper (trading frictions, intangibles, momentum, growth); (ii) placebo-characteristic experiments in which actual firm variables are replaced by randomly generated noise; and (iii) rolling out-of-sample validation on an earlier pre-specified hold-out window. Under the restricted and placebo specifications the ML edge disappears, while it persists in the economically motivated subset, consistent with genuine conditional-risk capture rather than spurious fit. revision: yes

Circularity Check

0 steps flagged

No circularity: standard empirical ML forecasting on external data

full rationale

The paper performs out-of-sample forecasting of asymmetric betas using ML models trained on firm characteristics and historical returns data. Performance metrics, feature importances, and downstream applications (CAPM reconstruction, DCF valuation, portfolio construction) are all evaluated on held-out periods or cross-sections. No equation or claim reduces by construction to a fitted parameter or self-citation; results are data-dependent and falsifiable against realized returns. Minor self-citations, if present, are not load-bearing for the central OOS performance claims.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Based on abstract only; ML models implicitly rely on many hyperparameters and the assumption that firm characteristics proxy for risk dynamics.

free parameters (1)

ML model hyperparameters
Tuning parameters for nonlinear models such as tree depth or regularization strength are fitted during training.

axioms (1)

domain assumption Firm characteristics contain predictive information for future asymmetric risk
Core premise for using them as inputs in the forecasting models.

pith-pipeline@v0.9.0 · 5439 in / 1096 out tokens · 26216 ms · 2026-05-08T08:50:29.138629+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

11 extracted references

[1]

The equally weighted MSE is constructed for each quintile, for the benchmark (red bars) and the conditional beta forecasts (blue bars)

Portfolios are formed monthly by sorting stocks into quintiles according to realised beta values (1-low beta, 5- high beta). The equally weighted MSE is constructed for each quintile, for the benchmark (red bars) and the conditional beta forecasts (blue bars). The figure also reports the fraction of stocks within each portfolio for which the difference be...

1990
[2]

The equally weighted MSE is constructed for each quintile, for the benchmark (red bars) and the conditional beta forecasts (blue bars)

Portfolios are formed monthly by sorting stocks into quintiles according to realised beta values (1-low beta, 5- high beta). The equally weighted MSE is constructed for each quintile, for the benchmark (red bars) and the conditional beta forecasts (blue bars). The figure also reports the fraction of stocks within each portfolio for which the difference be...

1990
[3]

The equally weighted MSE is constructed for each quintile, for the benchmark (red bars) and the conditional beta forecasts (blue bars)

Portfolios are formed monthly by sorting stocks into quintiles according to realised beta values (1-low beta, 5- high beta). The equally weighted MSE is constructed for each quintile, for the benchmark (red bars) and the conditional beta forecasts (blue bars). The figure also reports the fraction of stocks within each portfolio for which the difference be...

1990
[4]

The equally weighted MSE is constructed for each quintile, for the benchmark (red bars) and the conditional beta forecasts (blue bars)

Portfolios are formed monthly by sorting stocks into quintiles according to realised beta values (1-low beta, 5- high beta). The equally weighted MSE is constructed for each quintile, for the benchmark (red bars) and the conditional beta forecasts (blue bars). The figure also reports the fraction of stocks within each portfolio for which the difference be...

1990
[5]

The equally weighted MSE is constructed for each quintile, for the benchmark (red bars) and the conditional beta forecasts (blue bars)

Portfolios are formed monthly by sorting stocks into quintiles according to realised beta values (1-low beta, 5- high beta). The equally weighted MSE is constructed for each quintile, for the benchmark (red bars) and the conditional beta forecasts (blue bars). The figure also reports the fraction of stocks within each portfolio for which the difference be...

1990
[6]

The equally weighted MSE is constructed for each quintile, for the benchmark (red bars) and the conditional beta forecasts (blue bars)

Portfolios are formed monthly by sorting stocks into quintiles according to realised beta values (1-low beta, 5- high beta). The equally weighted MSE is constructed for each quintile, for the benchmark (red bars) and the conditional beta forecasts (blue bars). The figure also reports the fraction of stocks within each portfolio for which the difference be...

1990
[7]

The equally weighted MSE is constructed for each quintile, for the benchmark (red bars) and the conditional beta forecasts (blue bars)

Portfolios are formed monthly by sorting stocks into quintiles according to realised beta values (1-low beta, 5- high beta). The equally weighted MSE is constructed for each quintile, for the benchmark (red bars) and the conditional beta forecasts (blue bars). The figure also reports the fraction of stocks within each portfolio for which the difference be...

1990
[8]

The equally weighted MSE is constructed for each quintile, for the benchmark (red bars) and the conditional beta forecasts (blue bars)

Portfolios are formed monthly by sorting stocks into quintiles according to realised beta values (1-low beta, 5- high beta). The equally weighted MSE is constructed for each quintile, for the benchmark (red bars) and the conditional beta forecasts (blue bars). The figure also reports the fraction of stocks within each portfolio for which the difference be...

1990
[9]

The equally weighted MSE is constructed for each quintile, for the benchmark (red bars) and the conditional beta forecasts (blue bars)

Portfolios are formed monthly by sorting stocks into quintiles according to realised beta values (1-low beta, 5- high beta). The equally weighted MSE is constructed for each quintile, for the benchmark (red bars) and the conditional beta forecasts (blue bars). The figure also reports the fraction of stocks within each portfolio for which the difference be...

1990
[10]

The equally weighted MSE is constructed for each quintile, for the benchmark (red bars) and the conditional beta forecasts (blue bars)

Portfolios are formed monthly by sorting stocks into quintiles according to realised beta values (1-low beta, 5- high beta). The equally weighted MSE is constructed for each quintile, for the benchmark (red bars) and the conditional beta forecasts (blue bars). The figure also reports the fraction of stocks within each portfolio for which the difference be...

1990
[11]

The equally weighted MSE is constructed for each quintile, for the benchmark (red bars) and the conditional beta forecasts (blue bars)

Portfolios are formed monthly by sorting stocks into quintiles according to realised beta values (1-low beta, 5- high beta). The equally weighted MSE is constructed for each quintile, for the benchmark (red bars) and the conditional beta forecasts (blue bars). The figure also reports the fraction of stocks within each portfolio for which the difference be...

1990