Regime-Conditional Distributional Comparison of Trading Strategies: A GAMLSS/ZAGA Framework Applied to the S&P 500

Krzysztof Ozimek

arxiv: 2606.31251 · v1 · pith:BY4QA546new · submitted 2026-06-30 · 💱 q-fin.ST

Regime-Conditional Distributional Comparison of Trading Strategies: A GAMLSS/ZAGA Framework Applied to the S&P 500

Krzysztof Ozimek This is my paper

Pith reviewed 2026-07-01 03:08 UTC · model grok-4.3

classification 💱 q-fin.ST

keywords trading strategiesGAMLSSzero-adjusted gammamarket regimesS&P 500adjusted information ratiowalk-forward backtestdistributional comparison

0 comments

The pith

The dominance between a support vector machine trading strategy and buy-and-hold on the S&P 500 varies with market regime.

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

Conventional single-number performance summaries lose information about how trading strategies behave under different market conditions. This paper models sequences of adjusted information ratios from a walk-forward backtest using a GAMLSS model with zero-adjusted gamma response, where the parameters depend on realised volatility and cumulative market momentum. Regime-specific differences in expected value and variance are derived directly from the fitted model and tested via parametric bootstrap at representative points. A reader would care because the results show that which strategy performs better is not fixed but depends on the prevailing regime.

Core claim

The paper establishes that the relationship between a polynomial support vector machine strategy and a buy-and-hold benchmark, measured by sequences of adjusted information ratios, is regime-dependent when the sequences are modelled jointly with a GAMLSS/ZAGA specification conditioned on realised volatility and cumulative market momentum; regime-specific differences in expectation and variance are obtained analytically and assessed with bootstrap tests of three null hypotheses at six representative regimes.

What carries the argument

The GAMLSS model with zero-adjusted gamma response distribution whose location, scale, and shape parameters are additive functions of realised volatility and cumulative market momentum, allowing analytical extraction of regime-specific differences in expected adjusted information ratio and its variance.

If this is right

Strategy evaluation must be performed conditionally on market regime rather than through aggregate statistics over the full sample.
Differences in expected adjusted information ratio and its variance between the two strategies can be computed directly from the fitted parameters at any chosen combination of volatility and momentum.
Parametric bootstrap provides regime-specific p-values for hypotheses about expectation, variance, and their ratio.
The framework replaces single-number rankings with a surface of performance differences across market conditions.

Where Pith is reading between the lines

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

Portfolio construction could incorporate real-time regime indicators to switch between strategies dynamically.
The same distributional approach could be applied to other performance metrics or asset classes to test whether regime dependence is widespread.
Adding interaction terms or additional covariates such as interest-rate levels might refine the location of regime boundaries.

Load-bearing premise

The zero-adjusted gamma distribution accurately describes the adjusted information ratio sequences and the two chosen market covariates capture the conditions that drive differences in strategy performance.

What would settle it

A new set of out-of-sample folds in which the parametric bootstrap tests fail to reject the null hypotheses of equal expectation, equal variance, or equal ratio at every tested regime point would falsify the claim of conditional dominance.

read the original abstract

Conventional comparisons of algorithmic trading strategies reduce each performance metric to a single number over the full backtest horizon, thereby discarding information about how performance varies with market conditions. This paper proposes a distributional framework that addresses this shortcoming. A walk-forward backtest of 146 out-of-sample folds on the S&P 500 (2002--2025) is used to compute the Adjusted Information Ratio ($IR^{\ast}$) for a polynomial Support Vector Machine strategy (SVMP) and a buy-and-hold benchmark (BH) in each fold. The resulting $IR^{\ast}$ sequences are modelled jointly via a Generalised Additive Model for Location, Scale and Shape (GAMLSS) with a Zero-Adjusted Gamma (ZAGA) response, with distributional parameters conditioned on market regime covariates: realised volatility and cumulative market momentum. Strategy comparison is conducted through (i) regime-specific differences in expected $IR^{\ast}$ ($\Delta E$) and its variance ($\Delta Var$), derived analytically from the fitted ZAGA parameters, and (ii) parametric bootstrap tests of three null hypotheses concerning $E(IR^{\ast})$, $Var(IR^{\ast})$, and their ratio, evaluated at six representative market regimes. The results demonstrate that the dominance relationship between SVMP and BH is conditional on market regime. The proposed GAMLSS/ZAGA framework constitutes a methodologically rigorous and practically interpretable alternative to conventional strategy evaluation.

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

ZAGA can't handle negative IR* values, which undercuts the regime-conditional comparisons.

read the letter

The main point is that the paper's central modeling choice doesn't match the data. Adjusted Information Ratio can be negative, yet the ZAGA response in GAMLSS is restricted to [0, ∞) with a point mass at zero. That mismatch makes the fitted parameters, the analytically derived ΔE and ΔVar, and the parametric bootstrap tests invalid for the regimes where underperformance occurs.

The new element is the joint GAMLSS/ZAGA setup that models the full distribution of IR* sequences from the 146-fold walk-forward backtest, conditioned on realised volatility and cumulative momentum. It moves past single-number summaries by comparing expectations, variances, and their ratio across six representative regimes. The walk-forward design itself is straightforward and the analytical extraction of differences from the fitted parameters is efficient.

The backtest length and the choice of covariates are reasonable for the S&P 500 setting. The bootstrap tests for the three null hypotheses are a clean way to quantify regime-specific dominance.

The distribution support problem is not minor. It directly affects whether the claimed conditional dominance between SVMP and buy-and-hold holds. The abstract gives no sign that negatives were shifted, truncated, or handled with a different zero-adjusted family that allows negatives. Without that fix, the results rest on an unsupported assumption.

This is aimed at quant finance researchers who already use distributional or regime-aware evaluation. A reader looking for a template on GAMLSS in strategy comparison could extract the high-level structure, but the execution gap limits its immediate value.

I would not send it for peer review until the response distribution is corrected or justified with evidence that negatives do not appear in the folds.

Referee Report

1 major / 1 minor

Summary. The paper proposes modeling sequences of Adjusted Information Ratios (IR*) from 146 walk-forward backtest folds (S&P 500, 2002–2025) for an SVMP strategy and buy-and-hold benchmark via GAMLSS with ZAGA response distribution, conditioned on realized volatility and cumulative market momentum. Regime-specific comparisons are performed via analytically derived differences in expectation (ΔE) and variance (ΔVar) plus parametric bootstrap tests of three null hypotheses, with the central claim that dominance between SVMP and BH is conditional on market regime.

Significance. If the distributional model were correctly specified, the framework would supply a more informative alternative to single-summary-statistic strategy evaluation by enabling regime-conditional inference on both location and scale parameters.

major comments (1)

[Abstract / model specification] Abstract and GAMLSS/ZAGA model specification: ZAGA is a zero-adjusted gamma distribution whose support is [0, ∞). The Adjusted Information Ratio is defined as excess return divided by tracking error and therefore takes negative values whenever a strategy underperforms its benchmark. Negative-momentum regimes among the 146 folds will therefore contain negative IR* observations for at least one strategy, violating the support of the chosen response distribution and invalidating all fitted parameters, the derived ΔE and ΔVar quantities, and the parametric bootstrap tests that rest on them.

minor comments (1)

No model diagnostics, goodness-of-fit statistics, or residual plots for the GAMLSS fit are referenced, which would be required to evaluate whether the ZAGA specification is even approximately appropriate once the support issue is addressed.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for identifying a critical flaw in the distributional assumption. We address the point directly below and will revise the manuscript accordingly.

read point-by-point responses

Referee: Abstract and GAMLSS/ZAGA model specification: ZAGA is a zero-adjusted gamma distribution whose support is [0, ∞). The Adjusted Information Ratio is defined as excess return divided by tracking error and therefore takes negative values whenever a strategy underperforms its benchmark. Negative-momentum regimes among the 146 folds will therefore contain negative IR* observations for at least one strategy, violating the support of the chosen response distribution and invalidating all fitted parameters, the derived ΔE and ΔVar quantities, and the parametric bootstrap tests that rest on them.

Authors: We agree that the referee's observation is correct and represents a fundamental specification error. The ZAGA distribution cannot accommodate negative values, yet IR* (excess return over tracking error) is negative whenever either strategy underperforms. Our dataset does contain such negative observations in negative-momentum regimes. We will therefore replace the ZAGA response with a GAMLSS distribution that supports the full real line (e.g., normal or Student's t). The model will be re-estimated, all regime-specific ΔE and ΔVar quantities recomputed, and the parametric bootstrap tests repeated. These changes will be documented in the abstract, Section 3 (model specification), Section 4 (results), and the discussion. revision: yes

Circularity Check

0 steps flagged

No circularity: standard GAMLSS fitting applied to independently computed backtest metrics

full rationale

The derivation chain begins with walk-forward computation of IR* sequences from out-of-sample folds, followed by fitting a GAMLSS model with ZAGA response using external covariates (realised volatility, cumulative momentum). Regime-specific ΔE and ΔVar are then obtained analytically from the fitted parameters, and bootstrap tests are performed on those quantities. None of these steps reduces by the paper's own equations to a fitted parameter renamed as a prediction, a self-definition, or a load-bearing self-citation chain. The modeling choices (distribution family, covariates) are external statistical decisions whose validity is independent of the target comparisons; the paper therefore remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the distributional assumption that IR* follows a zero-adjusted gamma conditional on the two regime covariates, plus the modeling choice that these covariates adequately proxy relevant market conditions.

free parameters (1)

GAMLSS distributional parameters
Parameters of the ZAGA response (location, scale, shape, and zero-inflation) are estimated from the 146 IR* observations.

axioms (1)

domain assumption The Adjusted Information Ratio sequences are adequately described by a zero-adjusted gamma distribution whose parameters vary smoothly with realised volatility and cumulative market momentum.
This is the core modeling assumption invoked when the paper states the GAMLSS/ZAGA response is used to derive regime-specific expectations and variances.

pith-pipeline@v0.9.1-grok · 5784 in / 1336 out tokens · 52244 ms · 2026-07-01T03:08:40.801096+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

2 extracted references · 2 canonical work pages

[1]

https://doi.org/10.1111/j.1467-9876.2005.00510.x. 20 K. Ozimek Rigby, R. A., Stasinopoulos, D. M., Heller, G. Z., & De Bastiani, F. (2020). Distributions for modeling location, scale, and shape: Using GAMLSS in R . CRC Press. https://doi.org/10.1201/9780429298547. Ryan, J. A., & Ulrich, J. M. (2025). quantmod: Quantitative Financial Modelling Framework (R...

work page doi:10.1111/j.1467-9876.2005.00510.x 2005
[2]

Smeeton, N., Spencer, N

https://doi.org/10.1016/j.eneco.2011.05.001. Smeeton, N., Spencer, N. H., & Sprent, P. (2025). Applied nonparametric statistical methods (5th ed.). CRC Press. https://doi.org/10.1201/9780429326172. Stasinopoulos, D. M., & Rigby, R. A. (2017). Flexible regression and smoothing using GAMLSS in R. CRC Press. https://doi.org/10.1201/b21973. Stasinopoulos, M.,...

work page doi:10.1016/j.eneco.2011.05.001 2011

[1] [1]

https://doi.org/10.1111/j.1467-9876.2005.00510.x. 20 K. Ozimek Rigby, R. A., Stasinopoulos, D. M., Heller, G. Z., & De Bastiani, F. (2020). Distributions for modeling location, scale, and shape: Using GAMLSS in R . CRC Press. https://doi.org/10.1201/9780429298547. Ryan, J. A., & Ulrich, J. M. (2025). quantmod: Quantitative Financial Modelling Framework (R...

work page doi:10.1111/j.1467-9876.2005.00510.x 2005

[2] [2]

Smeeton, N., Spencer, N

https://doi.org/10.1016/j.eneco.2011.05.001. Smeeton, N., Spencer, N. H., & Sprent, P. (2025). Applied nonparametric statistical methods (5th ed.). CRC Press. https://doi.org/10.1201/9780429326172. Stasinopoulos, D. M., & Rigby, R. A. (2017). Flexible regression and smoothing using GAMLSS in R. CRC Press. https://doi.org/10.1201/b21973. Stasinopoulos, M.,...

work page doi:10.1016/j.eneco.2011.05.001 2011