acceptodds
Under review as a conference paper at ICLR 2027

Out-of-Distribution-Gated Momentum: Controlling Exposure via Feature-Based Wasserstein Distance

Abstract

Standard trend-following strategies rely on volatility scaling to size positions, inherently ignoring the knowledge about uncertainty when deployed in out-of-distribution (OOD) market conditions. We propose a distribution-aware familiarity score to explicitly quantify this uncertainty without relying on learned representations. Our method measures the optimal-transport distance between current market return distributions and their closest analogues in the training data, acting as a smooth, continuous throttle on position exposure. By isolating this sizing mechanism from the directional signal, we guarantee that performance differences stem purely from OOD exposure control. Across five asset classes and multiple validation schemes, our optimal-transport throttle consistently reduces portfolio drawdowns, with ablations confirming the mechanism captures OOD dynamics distinct from standard volatility. We also report critical negative results: jointly learning direction and size fails to converge, and enforcing strict statistical calibration actively degrades performance-highlighting a fundamental tension in OOD detection. Finally, we demonstrate that distribution-aware familiarity serves as a robust, interpretable principle for controlling exposure under domain shift.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Out of Distribution, Portfolio Management, Deep Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_dpt3OpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper asks whether an out-of-distribution (OOD) signal that uses no learned representation can improve position sizing (not direction) in trend-following. It constructs T1, a k-nearest-neighbour Wasserstein-2 distance between the current day's trailing return window and its k nearest training analogues in a four-dimensional, train-only-standardised condition-feature space, and passes it through a train-calibrated sigmoid gate that can only shrink a standard volatility-scaled magnitude. The mechanism is evaluated on 15 liquid instruments in five asset classes under walk-forward validation, cluster-robust bootstrap, hyperparameter / direction-signal / evaluation-window / transaction-cost sweeps, and against eleven competing gating scores, all under a perturbation-verified causal-timing protocol. The headline claim is a drawdown reduction (13/15 instruments, mean −6.42 pp; 68/75 walk-forward folds) with explicitly no Sharpe improvement, plus two reported tensions (calibration coverage vs. drawdown benefit; portfolio decorrelation vs. benefit) and a Lean 4 formalisation of the paper's elementary mathematics.

Critical strengths
  • Freezing a non-learned direction signal across every magnitude comparison, and pre-registering ΔMDD as the single primary test, makes the ablation family genuinely interpretable.
  • The eleven competing gates share identical machinery and the paper reports where each one wins (Mahalanobis 14/15, energy distance best MDD, CUSUM the only gate with non-negative mean Sharpe).
  • Verification discipline exceeds the norm: every signal passes a perturbation-based causal-timing test, the Lean scope is stated honestly, and the six-variant DMN failure is reported as a negative result.
Critical weaknesses
  1. The entire economic case is a drawdown-only effect with no Sharpe benefit (0/15 survive the Deflated Sharpe Ratio) that largely disappears once the gate correlation producing it is removed (+0.307 vs +0.157; 9.95% → 15.14%).
  2. T1 rests on a stationarity assumption its own coverage check refutes (33.1% exceedance vs 10% nominal), leaving it an arbitrary member of a gate class in which CUSUM, Mahalanobis, energy distance and MMD are equally competitive.
  3. Execution realism is only partly handled: turnover rises 1.52×, the benefit decays from 13/15 at 5 bps to 8/15 at 50 bps, and no spread, slippage or market-impact model is used.
  4. The Lean formalisation verifies only elementary statements on which none of the empirical claims depend, while generative AI was used to "create or modify" figures, so the plotted numbers cannot be verified from the paper.
  5. The evidence base is narrow (15 yfinance instruments, 2–3 per asset class, crypto only from 2020) and the headline negative result rests on an unverified DMN reimplementation at a single data scale.
Questions for authors
  1. The entire economic case is a drawdown-only effect with no Sharpe benefit (0/15 survive the Deflated Sharpe Ratio) that largely disappears once the gate correlation producing it is removed (+0.307 vs +0.157; 9.95% → 15.14%).
  2. T1 rests on a stationarity assumption its own coverage check refutes (33.1% exceedance vs 10% nominal), leaving it an arbitrary member of a gate class in which CUSUM, Mahalanobis, energy distance and MMD are equally competitive.
  3. Execution realism is only partly handled: turnover rises 1.52×, the benefit decays from 13/15 at 5 bps to 8/15 at 50 bps, and no spread, slippage or market-impact model is used.
  4. The Lean formalisation verifies only elementary statements on which none of the empirical claims depend, while generative AI was used to "create or modify" figures, so the plotted numbers cannot be verified from the paper.
  5. The evidence base is narrow (15 yfinance instruments, 2–3 per asset class, crypto only from 2020) and the headline negative result rests on an unverified DMN reimplementation at a single data scale.
1 comment

Dear Reviewer dpt3,

Because of the discussion periods will close soon. Dont know if my comments have solved all your concerns ?

Thank you again for your time and thoughtful feedback.

Reviewer_jyzuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Problem. Position sizing for single-asset trend following. Direction comes from a fixed MACD-combo momentum signal. The paper asks whether an out-of-distribution score can throttle the volatility-scaled position size.

Method.

  • Condition features. (σ^20,σ^60,dt,r5)(\hat\sigma_{20}, \hat\sigma_{60}, d_t, r_5) through day t−1t-1, z-scored on training data only.
  • T1 score. Softmax-weighted average of 1-D Wasserstein-2 distances between the trailing 20-day return window and the windows of its k=20k=20 nearest training-period neighbours.
  • Gate. A sigmoid on the standardised T1 score, thresholded at the training 90th percentile, multiplies the volatility-targeted magnitude mvol=clip(0.15/σ^60,0,2)m_{vol}=\mathrm{clip}(0.15/\hat\sigma_{60},0,2). The gate can only shrink exposure.

Evaluation. 15 instruments in five asset classes, daily data 2014–2024, single 60/40 chronological split for the headline and all competitor comparisons, 5-fold expanding-window walk-forward for the T1 vs. baseline test only. Metrics are ARR, AVol, ASR, MDD, Calmar. Sixteen alternative scores are plugged into the same gate.

Claims.

  • MDD falls on 13/15 instruments (mean −6.42 pp) and on 68/75 walk-forward folds, with no Sharpe benefit.
  • A naive volatility-ratio gate improves only 7/15, so T1 is argued not to be volatility timing in disguise.
  • Negative results on learned sizing, nominal calibration coverage, and portfolio decorrelation are reported.
  • Mathematical statements are machine-checked in Lean 4.
Critical strengths

1. Clean experimental isolation of the sizing decision. Direction is fixed and non-learned, every competing gate reuses the identical sigmoid machinery, and the one-day timing shift is verified by a perturbation test. Any difference between gates is attributable to the score alone. This is the right way to study a sizing signal.

2. Complete and candid reporting. Sixteen gates are reported in full, including those that beat T1. The paper states that it buys no Sharpe, that the Deflated Sharpe Ratio clears the bar on 0/15 instruments, that nominal calibration coverage destroys the benefit, and that portfolio decorrelation removes it.

Critical weaknesses

1. The headline drawdown reduction is not separated from a plain reduction in average exposure, and the one scale-free drawdown metric worsens.

Corollary 1 shows that Sharpe is invariant to constant scaling. MDD is not. Shrinking exposure by a constant reduces MDD almost proportionally. Table 5 shows the gate lowers realised volatility from 0.155 to 0.116 (about 25%) while MDD falls from 31.13% to 24.72% (about 21%). The two reductions are nearly equal. The Calmar ratio, which is approximately scale-free, falls from 0.076 to 0.059.

MetricBaselineT1 gateChange
AVol0.1550.116−25%
MDD (%)31.1324.72−21%
Calmar0.0760.059worse

A control that matches the gate's mean exposure or its realised volatility with a constant multiplier is absent. Without it, the primary claim cannot be attributed to timing. Any risk-adjusted metric the paper reports (Sharpe, DSR, Calmar) is flat or worse.

2. The proposed score is dominated by a two-parameter classical baseline in the paper's own table, and the statistical tests do not cover that comparison.

In Table 7, CUSUM on ∣rt∣|r_t| has higher mean ASR (0.001 vs. −0.029), lower mean MDD (23.39% vs. 24.72%), and more instruments improved (14/15 vs. 13/15). Appendix T nevertheless states that "both T1 and CUSUM are non-dominated points in Table 7". By the table's own numbers T1 is dominated. The walk-forward and cluster-robust tests of Section 4.3 are run only for T1 vs. the ungated baseline, never for T1 vs. CUSUM or any other gate. What is statistically supported is that gating beats no gating. The specific contribution, the W2-kNN score, is not shown to add value over simpler alternatives.

3. The leakage-prevention and calibration procedures cannot be verified, and one of them has a structural train/test asymmetry.

No code or data are provided (Appendix W). The paper describes a perturbation test and train-only standardisation, but a reviewer must take this on trust. Independently of that, training-period calibration of (μ,σ,zthresh)(\mu,\sigma,z_{thresh}) runs the kNN search over training windows that overlap the query by up to 19 of 20 returns. These near-duplicate neighbours have W2 distance close to zero, so training T1 is biased downward and the threshold is set too low. Test queries have no such neighbours. This alone would produce test exceedance far above nominal, which is exactly what Section 4.8 reports (33% vs. 10%) and interprets as evidence of regime shift.

Questions for authors

Q1. Constant-exposure control. Please add a baseline that multiplies mvolm_{vol} by a constant equal to the gate's mean value per instrument, and another that matches the gate's realised volatility. Report ΔMDD and Calmar against these controls, not only against the ungated baseline.

Q2. Statistical comparison against competing gates. Please run the Section 4.3 walk-forward and cluster-robust tests for T1 vs. CUSUM, T1 vs. Mahalanobis turbulence, and T1 vs. the drawdown rule. Please also correct the Appendix T statement that T1 is non-dominated in Table 7.

Q3. kNN calibration bias. For training-period queries, exclude neighbours whose windows overlap the query (a purge of at least ww days on each side) and recompute (μ,σ,zthresh)(\mu,\sigma,z_{thresh}). Report the resulting test exceedance rate and ΔMDD. Also confirm that in the walk-forward evaluation the kNN reference set is restricted to each fold's own training window.

Q4. Point-in-time verification. The paper states that standardisation and gate parameters use only training data. Please release the code, or at minimum the perturbation-test script and the per-fold calibration values, so this can be checked. Please also state whether adjusted-close prices (Appendix W) introduce any dependence on post-period dividend adjustments in the MACD inputs.

Q5. Multiple non-overlapping evaluation periods. The competitor comparison and all ablations use one 60/40 split whose test period is about 855 trading days dominated by 2022. Please report the full gate comparison on at least two further non-overlapping test periods with independently calibrated models, and include the 2008 period for non-crypto instruments.

Q6. Volatility target for crypto. A 15% target with σ^60\hat\sigma_{60} near 60–80% leaves crypto positions at roughly a quarter of nominal size before the gate is applied. Please report results across a grid of σtarget\sigma_{target} values, or per-asset-class targets, so that the near-zero crypto effect can be separated from this design choice.

Q7. Gate functional form. Only a sigmoid gate is used. Please justify the functional form against alternatives (linear ramp, step, piecewise, or a validation-selected shape) through a validation-based selection, and state how the 90th percentile and s=2.5s=2.5 were fixed before any test result was seen.

Reviewer_6Qi3OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies the distribution-aware exposure control for trend-following strategies. It keeps a conventional momentum type trick and adjusts volatility-scaled positions using the Wasserstein distances between recent return windows and historical windows retrieved through knn. Both theoretical and numerical analyses are conducted.

Critical strengths
  1. This paper addresses the important problem of controlling trading exposure under distribution shift.
  2. The authors clearly state the method’s limitations and report unfavorable findings, including the lack of a significant Sharpe-ratio improvement and the competitiveness of simpler alternatives.
Critical weaknesses
  1. The writing could be significantly improved. For example, the abbreviations, such as OHLCV, MACD, and EWMA are directly used with explanations. Moreover, various important numerical results are deferred to the appendix. It is a little bit hard for me to justify the effectiveness of the proposed method solely based on the main draft.

  2. It seems that the authors don't provide the sample code to help the reviewer verify the correctness of numerical experiments independently.

Questions for authors

Please see the weaknesses part.

At the current stage, I don't think the paper meets the standard for acceptance at a top-tier ML conference, and I tend toward rejection. However, I am open to revisiting my assessment based on the authors’ response and subsequent discussion.

1 comment

Dear Reviewer 6Qi3,

As the discussion phase is coming to a close, we would like to ask whether our responses have fully addressed your follow-up concerns. If any questions remain, please let us know. We would be happy to provide further clarification.

Thank you again for your time and thoughtful feedback.

Comments on the paper

2 comments

Dear Reviewers,

We have submitted our rebuttal and revised paper, addressing your comments and suggestions.

We hope the revisions help clarify our contributions and address your concerns.

Thank you for your time and consideration.

Thank you for the detailed follow-up. We appreciate your comments that the additional experiments resolve several concerns. We would like to clear the contributions supported by the revised evidence.

A controlled study of distribution-aware exposure management. The paper isolates magnitude control by holding trading direction fixed and comparing eleven alternative gating scores under shared execution and gate machinery. This design allows us to examine which findings depend on Wasserstein distance, which extend to simpler alternatives, and which reflect exposure reduction. Sections 4.5 and 4.10 report these distinctions explicitly. The contribution is the systematic characterization of this mechanism class, including its boundaries.

Evidence across evaluation schemes and instruments. Drawdown reduction relative to conventional volatility scaling persists under walk-forward evaluation, calendar blocks, alternative direction signals, transaction-cost sweeps and an extension to fifteen additional instruments. After excluding overlapping calibration windows, the reduction remains on 70/75 walk-forward folds and 70/75 calendar-block pairs. These results establish that the observed behavior extends beyond the original split and calibration procedure.

A substantive calibration finding. The additional analysis identifies why the original calibration–drawdown tension arose and demonstrates that it need not persist. With purged calibration, adaptive thresholding reaches 9.6% empirical exceedance against a nominal 10%, while retaining a mean drawdown reduction of 3.94 percentage points on 14/15 instruments. This provides a concrete methodological lesson for nearest-neighbour distributional scores built from overlapping time-series windows.

A clearer account of when gating helps. The constant-exposure controls distinguish full-period exposure reduction from crisis-conditional behavior. Within the GFC and COVID episodes, lower drawdowns relative to the full-period volatility-matched control persist after purging and across additional instruments. We present this as exploratory evidence within those episodes. Together with the mixed 2022 results, it identifies a specific empirical boundary for future investigation.

Evidence beyond the default gate configuration. Selecting gate shape, percentile and slope through nested validation retains the drawdown reduction against the ungated baseline and increases its magnitude from 3.28 to 5.39 percentage points. Thus, that finding does not depend on retaining the particular default gate configuration.

We believe these results support a useful contribution: a reproducible experimental design for studying distribution-aware exposure control, evidence across multiple evaluation settings, and a calibration analysis that materially changes how the mechanism should be understood. We respectfully ask that these supported contributions be weighed alongside the limitations identified by the controls.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.