Paper: arXiv 2408.07497

Authors: Jozef Barunik, Martin Hronec, Ondrej Tobek

Abstract

We propose a novel machine learning approach for forecasting the distribution of stock returns using a rich set of firm-level and market predictors. Our method combines a two-stage quantile neural network with spline interpolation to construct smooth, flexible cumulative distribution functions without relying on restrictive parametric assumptions. This allows accurate modelling of non-Gaussian features such as fat tails and asymmetries. Furthermore, we show how to derive other statistics from the forecasted return distribution such as mean, variance, skewness, and kurtosis. The derived mean and variance forecasts offer significantly improved out-of-sample performance compared to standard models. We demonstrate the robustness of the method in US and international markets.

Complexity vs Empirical Score

  • Math Complexity: 8.5/10
  • Empirical Rigor: 6.0/10 (capped from 7.5: reported Sharpe ratio of 4.34 is implausibly high for daily or lower-frequency data; see penalty flags)
  • Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 7.5)

Why this score: The paper employs advanced machine learning methods (quantile neural networks, spline interpolation) and detailed mathematical derivations for distribution forecasting, resulting in high math complexity. It demonstrates strong empirical rigor through extensive out-of-sample testing across multiple markets, comparison to benchmarks, and derivation of practical statistics like mean and variance.

Research Flowchart

  flowchart TD
  A["Research Goal:<br>Forecast Stock Return Distributions"] --> B["Data & Predictors"]
  B --> C["Two-Stage Quantile Neural Network"]
  C --> D["Spline Interpolation<br>to CDFs"]
  D --> E["Compute Statistics<br>Mean, Variance, Skewness, Kurtosis"]
  E --> F["Key Outcomes<br>Accurate Non-Gaussian Modelling<br>Superior Mean/Variance Forecasts<br>Robust in US & Int'l Markets"]
  
  subgraph Inputs
      B
  end

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 two‑stage quantile neural network combined with spline interpolation to forecast full return distributions.

Source sentence
Our method combines a two-stage quantile neural network with spline interpolation to construct smooth, flexible cumulative distribution functions without relying on restrictive parametric assumptions.

Assumes 4

  • The rescaling mechanism constrains the modeled return distributions to the location‑scale family. model
    Source sentence
    Specifically, the rescaling mechanism inherent in the architecture constrains the modeled return distributions to the location-scale family.
  • Missing characteristic values are imputed using cross‑sectional medians. data
    Source sentence
    Missing characteristic values are imputed using cross-sectional medians (Gu et al., 2020b).
  • Fundamental signals are updated weekly with financial statement information from at least six months earlier. data
    Source sentence
    This means that the fundamental signals are updated every week with financial statement information from financial years ending at least 6 months earlier.
  • Linear interpolation is used to ensure numerical stability and enforce monotonicity when the estimated density becomes ill‑behaved. model
    Source sentence
    To ensure numerical stability and enforce monotonicity, linear interpolation is used in cases where the estimated density becomes ill‑behaved.

Data

Backtest sample: 1995 – 2018 · US stocks · equities · monthly

  • CRSP/Compustat merged database · Wharton Research Data Service (WRDS) fundamentals commercial
    1963 to 2018 · U.S. equity market
    Source sentence
    We use the merged CRSP/Compustat database from the Wharton Research Data Service (WRDS) for the US sample.

Recognised sources: Refinitiv / LSEG (Datastream, Eikon, Tick History) Compustat CRSP Kenneth French Data Library WRDS

Costs

Not modeled A trading or execution result is reported without costs.

Finds 3

Reported performance (two-stage quantile neural network long‑short portfolio): Sharpe 4.34

  • The two‑stage model achieves the lowest continuous ranked probability score (CRPS) across all regions and both full and liquid samples, outperforming GARCH and other neural network benchmarks. predictive-accuracy
    Source sentence
    The two-stage neural network achieves the lowest CRPS in all regions and for both the full and liquid samples.
  • Mean and variance forecasts derived from the distributional forecasts significantly improve out‑of‑sample performance compared to standard models. performance
    Source sentence
    The derived mean and variance forecasts offer significantly improved out-of-sample performance compared to standard models.
  • Only the central quantiles are priced in the cross‑section, with long‑short portfolios based on these predictions delivering significant profitability; other moments are not priced. pricing-calibration
    Source sentence
    We find that only the central τ-quantiles are priced, with significant profitability of long‑short portfolios based on these predictions.

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.

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.