acceptodds
Under review as a conference paper at ICLR 2027

Adapt the Mode, Not the Model: Test-Time Operating-Point Selection for Hierarchical Portfolio Agents

Abstract

Trained hierarchical decision system often exposes several operating points, or modes, which modules to consult and in what combination allowing test-time adaptation without retraining or fine-tuning any module. Existing work on hierarchical and multi-agent systems typically argues that a specific configuration is superior on average, leaving a critical question unanswered: for a given environment, which mode should be deployed? We address this in portfolio construction using RASP (Regime-Aware Selection of operating Points), a three-tier agent whose operating points vary solely in their intermediate routing decision paths while sharing identical execution architectures. We show that the optimal operating point varies significantly across market regimes and universes, demonstrating that no single mode uniformly dominates across different environments or market drawdowns. To exploit this structure at test time, we propose an online selector based on a Hedge/fixed-share mixture over modes that observes only realized utility. Our framework effectively adapts to shifting market conditions without prior regime knowledge, outperforming both static mode assignments and standard follow-the-leader strategies. For hierarchical decision systems broadly, these results demonstrate the value of test-time mode selection over model retraining and emphasize the importance of identifying the precise conditions under which routing mechanisms provide an advantage.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Quant Investment

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_DxnPOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2(+1)
Soundness
2
Read review
Paper summary

This paper proposes RASP, a framework for adapting hierarchical portfolio agents to changing market conditions without retraining their models. The framework uses three portfolio strategies: a basic return prediction model, an expert-based consensus model, and a model that incorporates additional market features. Rather than selecting a single fixed strategy, RASP dynamically adjusts the weights of these strategies based on recent investment performance using the Hedge and fixed-share algorithms. It also includes a validation-based gate that selects a fixed strategy when the adaptive mixture does not achieve better performance. The authors report that RASP generally achieves higher risk-adjusted returns and lower trading costs than several existing portfolio methods.

Critical strengths

One strength of this study is that it addresses a practical and meaningful problem in portfolio management. I find the motivation generally reasonable, as the best investment strategy may change depending on market conditions, and selecting an appropriate strategy is important for real-world applications. The idea of adapting an existing portfolio system without retraining its models also seems reasonable from a practical perspective. In addition, the study compares multiple investment strategies using a shared portfolio allocator and evaluates their performance using long-term historical market data. The experiments consider actual portfolio performance, including risk-adjusted returns and transaction costs, rather than focusing only on prediction accuracy. The proposed method also reports generally better performance than existing approaches in long-horizon experiments. These aspects make the motivation of this study reasonable from an application-oriented research perspective.

Critical weaknesses

Despite the well-motivated research problem, I found it somewhat difficult to follow the details of the paper. In particular, the presentation of the methodology seems more complicated than necessary given the relatively straightforward main idea of the study.

This study is closer to an application-oriented study that uses existing methods rather than proposing a new machine learning methodology. Therefore, I do not think that proposing a new machine learning algorithm is necessarily required. However, even for an application-oriented study, the authors should provide sufficient justification for why existing methods were selected and combined, and why the resulting design is appropriate for addressing the research problem. In particular, Hedge and fixed-share are well-established algorithms designed to track the best experts over time, and applying them to the selection and combination of multiple portfolio strategies is also similar to existing approaches. Furthermore, the authors do not sufficiently explain the rationale for selecting the three operating points or the additional benefits of the hierarchical structure compared to conventional portfolio ensembles. For these reasons, I find it difficult to see what substantial advantages or practical differences RASP offers over existing expert selection and portfolio ensemble methods.

This concern is also reflected in the ablation results, which do not clearly demonstrate the contribution of test-time adaptation itself. In Table 5, removing the validation gate reduces the Sharpe ratio from 0.954 to 0.750, whereas removing fixed-share from the ungated Hedge strategy only reduces it from 0.750 to 0.718. This suggests that the validation gate plays a substantial role in the reported performance improvement. However, as far as I understand, the validation gate is determined before the evaluation period and does not perform dynamic adaptation during testing. Please correct me if I have misunderstood this. Therefore, I think it is reasonable to question whether the performance improvement comes from online mode adaptation or from selecting a fixed strategy based on validation data.

In addition, the word "Equation" is repeated in lines 297 and 303. I recommend that the authors carefully review the manuscript for grammatical and typographical errors to improve its overall readability.

Questions for authors
  1. The performance difference between RASP and Hedge seems to reflect the effect of adding the validation gate rather than the superiority of a new online selection algorithm. What are the authors' thoughts on this interpretation?

  2. RASP appears to show a larger performance advantage over existing methods in Table 2 than in Table 3. Could the authors provide a possible explanation for this difference?

  3. In Section 4.4, why is each operating point's utility normalized using its own historical range? Since this may not preserve relative performance differences across modes, could the authors explain the rationale and whether a common normalization scale would affect the results?

3 comments
AuthorsOpenReview ↗

Response to Reviewer DxnP (1/2): baseline protocol, deployed configuration, W3/Q1, Q2

Thank you for the review. Your reading of the gate (W3) is correct, and Q2 bears on which numbers are being compared, so we give the baseline protocol and the deployed configuration first. Baseline protocol. The five recent systems in Table 3 — FinCon (Yu et al., 2024), TradingAgents (Xiao et al., 2024), MetaTrader (Niu et al., 2022), MacroHFT (Zong et al., 2024) and FINSABER (Li et al., 2026) — are reproduced and run by us from each paper's method description and released code, mapped onto our protocol rather than quoted: the same universes, the same daily OHLCV-derived information set (≤126 trading days), the same weekly Friday calendar, the same risk model, the same cap-constrained ℓ1-cost allocator and the same 25 bps one-way cost. Their published numbers use different assets, frequencies and horizons and are not comparable, so we do not use them. Under this protocol each reproduced system's book reduces to the static rule to which its allocation step corresponds (App. C.3), and the systems are reported together with those rules as one baseline family. Deployed configuration. The RASP row of the submitted Table 3 reports the gate-free mixture at fixed λ_alloc. The revised paper reports the deployed policy: the gate is fitted once on F1–F2, before any test fold, and kept O1, so the deployed policy is the fixed mode O1, at 0.754 / 9.4 / 0.462. The gate-free mixture appears beside it at 0.750 / 10.4 / 0.439. Table 3. 25 bps one-way. US: 44 cells, F3–F6, 2014–2026. JP+EU: 69 held-out cells, evaluated once.

PolicyUS SRUS U_CVaRUS MDD%US Turn.JP+EU SRJP+EU MDD%
Deployed (gate→O1)0.754−0.486−20.79.40.462−26.6
Gate-free mixture0.750−0.486−20.910.40.439−26.6
Follow-the-leader0.732−0.495*−21.311.20.427−26.7
O2 routed consensus0.657−0.523**−21.914.20.327−28.3
O3 alpha-augmented0.752−0.491*−20.810.60.440−26.6
Min-Variance (FinCon repro.)0.784−0.495−21.39.90.480−28.1
HRP (TradingAgents repro.)0.760−0.502*−21.58.70.484−27.9
Momentum (MetaTrader repro.)0.707−0.641**−24.512.70.348−31.6
Equal-Weight (MacroHFT repro.)0.914−0.533**−23.01.40.573−29.8
Risk-Parity (FINSABER repro.)0.895−0.511*−22.11.80.583−28.7
*/**: Deployed ahead at p < 0.05 / 0.01 (one-sided paired stationary bootstrap, mean block 4, 5,000 resamples).
W3/Q1. Online adaptation or validation selection? Validation selection. The gate is decided once on F1–F2 (2008–2013), before any test fold, and never revisited; it chose O1 because O1 0.0457 > the mixture's 0.0417. The deployed policy is a fixed mode. Its +0.004 SR over the gate-free mixture is not significant (ΔU +0.0003, p = 0.43), and removing fixed share leaves SR nearly unchanged (0.750 → 0.749). The mixture alone does not beat the best fixed mode (US ΔU −0.0003, p = 0.57). It is above follow-the-leader at lower turnover (US +0.0013, p = 0.24; JP+EU +0.0017, p = 0.20; random pools +0.0035, p = 0.031); only the last is significant. Contribution 2 is rewritten to state this.
Q2. Table 2 vs Table 3. Table 2 reports a single 2024–2025 window, so it is not on the same protocol as Table 3. For direct comparability, the revised paper reports the four universes on the Table 3 protocol (F3–F6, 25 bps), net SR:
MethodBigTech-6Consumer-20High-Vol-15Defensive-15
---------------
Equal-Weight1.290.960.740.70
Risk-Parity1.230.960.690.69
HRP1.220.820.630.62
Min-Variance1.190.780.580.47
Deployed (gate→O1)1.060.820.520.55
The two tables now agree: EW and RP lead on Sharpe in all four universes; Deployed is below both. Deployed has the shallowest drawdown on Consumer-20 (−15.8%) and pooled.
AuthorsOpenReview ↗

Response to Reviewer DxnP (2/2): Q3, W2, W1, revision

Q3. Normalization. The code uses one common range (meta_policy._causal_scale: min/max over all modes' past utilities), and Sec. 4.4 is aligned with it. We also ran the per-mode variant you describe (E2), mean U_MV per quarter:

SampleO1Hedge, sharedHedge, per-mode
Selection F1–F20.04570.04170.0399 (shared ahead, p = 0.045)
US test0.08320.08290.0831 (p = 0.34)
JP+EU0.04170.03860.0370 (shared ahead, p = 0.049)
Per-mode scaling discards level differences and is worse on the selection folds and on JP+EU. On the US test folds the two versions do not differ. Neither version changes the gate.
W2. Why these components and three operating points.
  • Hedge/fixed share. Only delayed quarterly net utility is observed, mode leadership is not predictable from state, and the leader switches. Exponential weights gives O(√(T ln|Π|)) regret without a loss model; fixed share handles switching leaders (Herbster & Warmuth 1998); FTL is its η → ∞, β = 0 limit. A contextual bandit/RL selector would need state features and tuning on 22 validation cells.
  • Router (Eq. 4), consensus (Eq. 6), allocator (Eq. 7) are the cost-net certainty equivalent, the Gaussian posterior mean, and the exact ℓ1-cost problem with one solver for every mode, so mode differences are not unit or solver artifacts.
  • Three points: base O1, plus coordination (O2) or more information (O3): the minimal set that separates the two. No point dominates (best mode: 11 US universes O1 4 / O2 2 / O3 5; JP+EU 8 / 1 / 3). O2 is the weakest mode in every sample, with the highest turnover (14.2×/yr in the US). O3 matches the base on Sharpe (0.752 vs 0.754). W2. Versus portfolio ensembles (new, E5, post hoc, so exploratory). On mean–variance our modes have no advantage over a conventional ensemble; the advantage is limited to CVaR. Same Hedge rule over the five classical allocators, and over those five plus O1–O3: | Policy | US SR | US U_MV | US U_CVaR | JP+EU U_MV | JP+EU U_CVaR | |---|---|---|---|---|---| | Deployed | 0.754 | 0.0832 | −0.486 | 0.0417 | −0.570 | | Ensemble, 5 classical | 0.851 | 0.1103 | −0.529 | 0.0531 | −0.636 | | Ensemble, 5 + O1–O3 | 0.834 | 0.1037 | −0.511 | 0.0512 | −0.600 | | Equal-Weight | 0.914 | 0.1219 | −0.533 | 0.0623 | −0.663 | U averaged over quarters (Table 3's JP+EU CVaR averages per cell: Deployed −0.577).
  1. On mean–variance the classical ensemble beats our modes and EW beats the ensemble; the gate rule would deploy 1/N for that menu.
  2. On CVaR-95 (Table 3; all portfolios fully invested) Deployed beats both ensembles in the US and JP+EU (p ≤ 0.004). Adding O1–O3 to the classical menu improves its CVaR but lowers its MV utility (−0.0066, p = 0.008). W1. Presentation. In the revised PDF, Sec. 4 opens with a five-line summary; derivations and Defs. 4.1–4.2 move to the appendix; the repeated "Equation" at l. 258, 297, 303 and the Fig. 1 caption are fixed. Table 4 moves to the appendix with the evaluated grid {50, 100, 200, 400} plus per-cell validation selection (Deployed SR 0.628 / 0.714 / 0.764 / 0.785 at fixed λ; 0.754 as deployed). Revision. Tables 2–5 report the deployed configuration and the reproduced baselines; the abstract, Sec. 5 and conclusion state three findings: no operating point dominates; selection from realized utility is limited by signal-to-noise, so the gate deploys O1; the deployed policy has a CVaR advantage over 1/N (US +0.047, p = 0.004; JP+EU +0.096, p = 0.0002), while 1/N leads on Sharpe and mean–variance, now stated as the main limitation.
AuthorsOpenReview ↗

summary of rebuttal revision

  • Q3 normalisation → B7 (the formula now matches the code) and A2 (shared range ahead on F1–F2, p ≈ 0.045, and on JP+EU, p ≈ 0.049; tied on U.S., p = 0.34).
  • W2 vs portfolio ensembles → A1, E5 rows (CVaR: Deployed ahead of Ens-classical and Ens-all in U.S. and JP+EU, p ≤ 0.004).
  • W2 static mixture → A1, E4 row (JP+EU +0.0045, p = 0.032).
  • W1 presentation → B5, B6, B8.

Hope the revision and clarify in rebuttal can solve your concern and consider increase the rating of the paper.

Reviewer_cLunOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies mode selection in hierarchical portfolio construction: which operating mode should be deployed as market conditions change? It proposes an online mixture over three fixed modes, updating their weights from realized utility without retraining the underlying modules. Backtests report improvements over selected static configurations and follow-the-leader strategies. The paper argues that expert routing has context-dependent value and should be evaluated jointly with allocation, risk, and trading costs.

Critical strengths
  • The paper provides a practical framework for deployment-time adaptation. Selecting among existing experts can be simpler than retraining modules. Routing value is universe-dependent. And RASP dynamic selector updates mode allocations purely from realized utility via Hedge/fixed-share.
Critical weaknesses
  • The deployment gate seems to account for a substantial share of the improvement. Removing it reduces the reported U.S. Sharpe ratio from 0.954 to 0.750 and increases turnover from 1.4 to 10.4 times NAV annually. These differences make it difficult to attribute the overall gain specifically to adaptive mode weighting, the routing itself is not stable.

  • The selector relies on established Hedge/fixed-share updates, combined with portfolio optimization. The contribution is primarily a deployment framework and empirical study; the algorithm itself is not new.

Questions for authors
  • Actually the experiments are done on several highlighted universes, which already contains priors. Does the method perform consistently on randomly sampled stock pools?
  • How strongly do the results depend on the hand-designed expert set? Say the advantage over MacroHFT is not that strong in the long-horizon test. Does the performace gain comes from a more complicated agent framwork or say the adaptive mode weighting, the deployment gate?
  • What will happen with nonzero quadratic market impact?
3 comments
AuthorsOpenReview ↗

**Response to Reviewer cLun (1/2): baseline protocol, deployed configuration, W1**

Thank you for the review. W1 and Q2 both turn on how the baselines and the deployed policy are defined, so we give the two definitions first and then answer. Baseline protocol. The five recent systems in Tables 2–3 — FinCon (Yu et al., 2024), TradingAgents (Xiao et al., 2024), MetaTrader (Niu et al., 2022), MacroHFT (Zong et al., 2024) and FINSABER (Li et al., 2026) — are reproduced and run by us, from each paper's method description and released code, and mapped onto our protocol rather than quoted from their papers: the same universe construction and the same eleven universes, the same daily OHLCV-derived information set (trailing windows of up to 126 trading days), the same weekly Friday calendar, the same risk model (Σ over 126 days, strictly prior), the same long-only cap-constrained ℓ1-cost allocator, and the same 25 bps one-way cost. Their published numbers come from different assets, frequencies and horizons, so they are not comparable to ours and we do not use them. Under this protocol a reproduced system's decision layer acts on our information set and is passed through our allocator, so its realized book reduces to the static rule to which its allocation step corresponds; App. C.3 gives the reduction for each of the five. The systems and those five rules are reported as one baseline family, and the values in Tables 3 and 5 are runs of the reproduction. Deployed configuration. The RASP row of the submitted Table 3 reports the meta-policy before the gate's deployment decision is applied — the gate-free mixture at fixed λ_alloc. The revised paper reports the deployed policy, which is what the gate chooses: the gate is fitted once on F1–F2, before any test fold, and it kept O1 (validation utility 0.0457 against 0.0417 for the mixture), so the deployed policy is the fixed mode O1 at SR 0.754, turnover 9.4, JP+EU 0.462. The gate-free mixture is reported beside it at 0.750 / 10.4 / 0.439, so the reader can see exactly what the gate contributes. This is also why the deployed policy does not lead on Sharpe: the reproduced systems' books are ahead of it on that metric. Tables 3 and 5. 25 bps one-way. US: 44 cells, folds F3–F6, 2014–2026. JP+EU: 69 held-out cells, evaluated once.

PolicyUS SRUS U_CVaRUS Turn.JP+EU SR
Deployed (gate→O1)0.754−0.4869.40.462
Gate-free mixture (RASP-full, β = 0.05)0.750−0.48610.40.439
Hedge β = 00.749—10.40.441
Static equal mixture0.750—10.30.432
Follow-the-leader (FTL)0.732−0.495*11.20.427
O3 alpha-augmented0.752−0.491*10.60.440
O2 routed consensus0.657−0.523**14.20.327
Min-Variance (FinCon repro.)0.784−0.4959.90.480
HRP (TradingAgents repro.)0.760−0.502*8.70.484
Momentum (MetaTrader repro.)0.707−0.641**12.70.348
Equal-Weight (MacroHFT repro.)0.914−0.533**1.40.573
Risk-Parity (FINSABER repro.)0.895−0.511*1.80.583
*/**: Deployed ahead at p < 0.05 / 0.01 (one-sided paired stationary bootstrap, mean block 4, 5,000 resamples). —: not computed.
W1. Gate vs. online mixture. We agree on the substance: the gate, not adaptive weighting, determines what is deployed, and routing is not stable. The pair in your quotation (0.954 → 0.750) is the submitted table's; the released protocol gives 0.754 → 0.750 for the same on/off comparison, with turnover 9.4 → 10.4. The gate adds ΔU = +0.0003 per quarter (p = 0.43), which is not significant. It kept O1 because the mixture's validation utility (0.0417) was below O1's (0.0457). The gate-free mixture on its own (ΔU per quarter, one-sided p):
SampleHedge − O1Hedge − FTL
---------
US (44 cells)−0.0003 (p = 0.57)+0.0013 (p = 0.24)
JP+EU (69)−0.0031 (p = 0.95)+0.0017 (p = 0.20)
Random pools (40)−0.0038 (p = 0.91)+0.0035 (p = 0.031)
The mixture does not beat O1 in any sample. It is numerically above FTL in all three; the difference is significant only on random pools. The cause is measured: the median per-quarter gap between modes is about 20× smaller than its s.d. (0.0016 vs 0.032), which limits any selector that reads realized utility. Routing instability has the same cause: O2 wins the most quarters in High-Vol (45%) but has the lowest full-sample utility in every universe. Contribution 2 is rewritten to state the deployed configuration and this bound.
AuthorsOpenReview ↗

**Response to Reviewer cLun (2/2): W2, Q1–Q3, revision**

W2. Novelty. Agreed: the algorithm is not new and we do not claim one (cf. Das & Banerjee, KDD 2011; Li & Hoi, CSUR 2014). The contribution is a controlled protocol and a negative result on adaptive mode weighting: (i) the three modes share features, risk model, allocator, universe and calendar, so differences isolate the Tier-2 path; (ii) the mixture is one portfolio that pays for every weight change; (iii) gate and rates are fixed on F1–F2; Deflated Sharpe 0.998 at N = 194 passes, but PBO is 0.88 over the 8 meta-policy configurations and fails the < 0.2 rule (their validation ranking does not carry to test, consistent with the gate rejecting the mixture); (iv) 18.75 years, US plus two held-out markets, net of cost. No module is retrained; all 44 cells run in ~14 min on one CPU. For hierarchical trading systems it gives a direct check: compare against the best fixed configuration and 1/N, net of switching cost. Q1. Random pools (new, E3). Setup: 10 random 20-name pools per fold from the 200 most liquid stocks; unchanged pipeline; gate not revisited. Columns 4–5: Deployed minus comparator (p).

40 poolsSRTurn.ΔU_MVΔU_CVaR
Deployed0.7918.8——
Hedge0.7739.4+0.0038 (0.087)—
FTL0.7429.7+0.0073 (0.005)—
O20.66813.6+0.0194 (0.014)+0.0553 (<0.001)
EW0.8841.6−0.0278+0.0859 (0.001)
RP0.8861.9−0.0210+0.0447 (0.009)
As in the main samples: no dominant mode (best mode O3 19 / O1 16 / O2 5); Deployed above FTL; CVaR advantage over EW and RP; EW and RP ahead on Sharpe and mean–variance. Deployed vs Hedge is not significant at 0.05.
Q2. MacroHFT, experts, source of the gain.
  • MacroHFT. Under the common protocol the reproduced MacroHFT book reduces to the equal-weight rule (App. C.3), SR 0.914; the deployed policy is at 0.754. Your observation therefore reads: EW beats Deployed on Sharpe. We state this as the main limitation in the main text.
  • Expert set. The deployed policy (O1) uses no experts, so the result does not depend on the expert set. Doubling O2's panel from 4 to 8 does not change its SR significantly (0.657 → 0.641, p = 0.13) and raises turnover. The experts are fixed trailing rules (Table 6); learned/LLM experts are named as future work.
  • Source. No component yields a gain over the best fixed mode (Table 3 above). Gate: 0.750 → 0.754, not significant (W1). Adaptive weighting: same US SR as the static equal mixture (0.750) and as β = 0 (0.749). Smooth weighting is above FTL. O2, the most complex mode, is the weakest. Q3. Quadratic impact (new, E1). Λ_ii = 2ησ_i·AUM/ADV_i, η = 0.142 (Almgren et al. 2005), on top of 25 bps; allocator not re-tuned. Net SR: | AUM | Deployed | Hedge | FTL | O2 | EW | |---|---|---|---|---|---| | 0 | 0.754 | 0.750 | 0.732 | 0.657 | 0.914 | | $100M | 0.752 | 0.748 | 0.731 | 0.652 | 0.914 | | $1B | 0.731 | 0.731 | 0.716 | 0.609 | 0.913 | Impact is small up to $100M (liquid large caps). At $1B O2 loses most: the Deployed − O2 utility gap grows from +0.0119 to +0.0160 (p < 0.001). The order is unchanged except that Deployed and Hedge tie at $1B. EW is almost unaffected and remains ahead. Impact inside the allocator (Gârleanu & Pedersen 2013) is listed as an extension. Revision. Tables 2–5 report the deployed configuration and the reproduced baselines as described in 1/2; the five reproduced systems appear with the static rules their books reduce to, as one baseline family. The abstract, Sec. 5 and conclusion state three findings: (1) no operating point dominates; (2) selecting from realized utility is limited by signal-to-noise, so the gate deploys O1 and the mixture does not beat it; (3) the deployed policy has a CVaR advantage over 1/N (US +0.047, p = 0.004; JP+EU +0.096, p = 0.0002), while 1/N leads on Sharpe and mean–variance. Contribution 2 becomes "the validation gate deploys the fixed mode O1; the deployed policy exceeds follow-the-leader with lower turnover and holds a CVaR advantage over 1/N that replicates in two held-out markets".
AuthorsOpenReview ↗

summary of rebuttal revision

  • Q1 random pools → A1, E3 rows. Deployed is ahead of O2 (+0.0194, p = 0.014) and Argmax (+0.0073, p = 0.005), and ahead of Hedge (+0.0038, p = 0.087). Under CVaR it is ahead of EW, RP and HRP (p ≤ 0.009).
  • Q3 quadratic impact → A1, E1 rows. Deployed is ahead of O2 at $0 (+0.0119, p = 0.003) and at $1B (+0.0160, p = 0.001). A5 adds the citation.
  • W2 novelty → no PDF change.
  • Presentation → B1–B4.

Hope the revision and clarify in rebuttal can solve your concern and consider increase the rating of the paper.

Reviewer_uYuVOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies whether a portfolio strategy can adapt to changing markets by reweighting three fixed forecasting modes without retraining them. The modes differ in how they forecast returns, but all use the same turnover-aware mean–variance optimizer. Each quarter, the method gives more weight to portfolios that delivered higher realized utility. A validation rule determines before testing whether to deploy the adaptive blend or use a fixed mode.

Critical strengths

None

Critical weaknesses

The main weakness is that the paper does not establish what the proposed setup adds beyond standard adaptive weighting. It uses established Hedge and fixed-share updates, and its headline results combine cases that adapt with cases that use a fixed strategy.

Questions for authors

Why does the method blend portfolio positions rather than expected-return forecasts? Since all three modes share the same turnover-aware mean–variance optimizer, a direct comparison would apply the same quarterly weights to the forecasts, then optimize once using the portfolio’s actual holdings.

4 comments
AuthorsOpenReview ↗

Response to Reviewer uYuV (1/2): baseline protocol, deployed configuration, W1

Thank you for the review. Both halves of your weakness are correct, and we answer them directly. Two definitions bear on them, so we give those first. Baseline protocol. The five recent systems — FinCon (Yu et al., 2024), TradingAgents (Xiao et al., 2024), MetaTrader (Niu et al., 2022), MacroHFT (Zong et al., 2024) and FINSABER (Li et al., 2026) — are reproduced and run by us from each paper's description and released code, mapped onto one common protocol: the same universes, the same daily OHLCV-derived information set (≤126 trading days), the same weekly Friday calendar, the same risk model, the same cap-constrained ℓ1-cost allocator, the same 25 bps one-way cost. We do not quote their published numbers, which come from different assets, frequencies and horizons. Under this protocol each reproduced system's book reduces to the static rule to which its allocation step corresponds (App. C.3), so the systems and those rules are reported as one baseline family. Deployed configuration. The RASP row of the submitted Table 3 reports the gate-free mixture at fixed λ_alloc. The revised paper reports the deployed policy: the gate is fitted on F1–F2 and kept O1, so the deployed policy is the fixed mode O1 at 0.754 with turnover 9.4; the gate-free mixture is 0.750 at 10.4. Adaptive and fixed separated. 25 bps one-way. US: 44 cells, F3–F6, 2014–2026. JP+EU: 69 held-out cells, evaluated once.

PolicyTypeUS SRUS U_MVUS U_CVaRUS Turn.JP+EU SR
Deployed (gate→O1)fixed0.7540.0832−0.4869.40.462
Gate-free mixtureadaptive0.7500.0829−0.48610.40.439
Hedge β = 0adaptive0.749——10.40.441
Static 1/3 mixturefixed mix0.750——10.30.432
Follow-the-leaderadaptive0.7320.0816−0.49511.20.427
O3 alpha-augmentedfixed0.7520.0833−0.49110.60.440
O2 routed consensusfixed0.6570.0713−0.52314.20.327
Equal-Weight (MacroHFT repro.)fixed0.9140.1219−0.5331.40.573
W1. (a) The algorithm is standard; we do not claim a new one. Hedge with fixed share, and the gate, are established components, and the paper says so. (b) The two cases are now reported separately. Separated, the adaptive policy does not beat the best fixed mode in any sample: ΔU per quarter, mixture − O1, is US −0.0003 (p = 0.57), JP+EU −0.0031 (p = 0.95), 40 random pools −0.0038 (p = 0.91). The mixture is above follow-the-leader in all three (+0.0013, p = 0.24; +0.0017, p = 0.20; +0.0035, p = 0.031), significant only on the random pools. The gate adds +0.0003 (p = 0.43); with β = 0 the Sharpe is 0.749 and with a static 1/3 mixture 0.750, against 0.750 for the gate-free mixture: adapting is indistinguishable from not adapting.
The cause is measured, not asserted. The median per-quarter gap between the best and second-best mode is 0.0016 against a gap standard deviation of 0.032 — about 20× below its own noise. No selector that reads realized quarterly utility, standard or otherwise, can separate the modes at that ratio. Consistent with this, O2 wins the most quarters in high volatility (45%) and has the lowest full-sample utility in every universe.
What the setup adds, stated without overclaiming: (i) a comparison object in which the three operating points share features, risk model, allocator, universe and calendar, so differences isolate the Tier-2 path (routing versus more information) — an expert or strategy menu, as in standard adaptive weighting, differs in many things at once; (ii) an explicit, irreversible fallback decided on F1–F2 before any test fold is read, so the failure case is reported rather than tuned away; (iii) a result that transfers: the deployed policy beats 1/N on CVaR-95 in the US (+0.047, p = 0.004) and in both held-out markets (+0.096, p = 0.0002), while 1/N leads on Sharpe (0.914 / 0.573 against 0.754 / 0.462) and on mean–variance utility, now the main stated limitation. Contribution 2 is rewritten to report the fixed deployed mode, not an adaptive gain.
AuthorsOpenReview ↗

Response to Reviewer uYuV (2/2): blending positions vs. blending forecasts

Your question is the right control and we ran it (new, E6, in the appendix): the same quarterly Hedge weights applied to the modes' forecasts, then one allocator call per rebalance against the portfolio's own drifted holdings. Otherwise identical — allocator, Σ (126 days, strictly prior), cap, κ, λ, weekly Friday calendar, 25 bps. Why the two are not equivalent. With long-only, per-asset caps, Σw = 1 and an ℓ1 turnover charge against the book's own drifted holdings, the mean–variance solution is not linear in μ, so blending positions and blending forecasts do not commute. They coincide only in the unconstrained, κ = 0 case, where w ∝ (1/λ)Σ⁻¹μ. Second, each mode's w solves against that mode's own holdings while the blended forecast solves once against the blended book's holdings, so the turnover term differs as well. Third, the quarterly utility that drives the Hedge update is computed on the traded book, not on the forecasts. Why we blend positions. The Hedge weights are learned from each mode's realized net utility, and that utility belongs to the mode's own traded book, including its own costs. Blending those books keeps the quantity that is weighted equal to the quantity that is scored: the mixture's return is the P-weighted return of the books that earned the utilities, up to the trades needed to rebalance between them. A forecast blend applies the same weights to an object that was never scored; its book is a new portfolio, whose utility the weights did not see. That is a design argument, not evidence that positions are better, so we tested it. Results (new, E6). U_MV per quarter; Hedge_fc = forecast blend, Hedge_pos = position blend.

SampleDeployedHedge_posHedge_fcStatic_posStatic_fcFTL
F1–F2 (22)0.04570.04170.04140.04090.03980.0403
US (44)0.08320.08290.08240.08280.08130.0816
JP+EU (69)0.04170.03860.04000.03750.03830.0367
One-sided paired stationary bootstrap (mean block 4, 5,000 resamples), Hedge_fc minus comparator, U_MV: F1–F2, position blend −0.0003 (p = 0.58); US, position blend −0.0004 (p = 0.64), static position blend −0.0004 (p = 0.60), follow-the-leader +0.0009 (p = 0.34); JP+EU, position blend +0.0013 (p = 0.14), static position blend +0.0026 (p = 0.040), follow-the-leader +0.0030 (p = 0.11). Against the deployed fixed mode: US −0.0008 (p = 0.39), JP+EU −0.0019 (p = 0.25). Turnover is slightly lower for the forecast blend (10.3 vs 10.4 in the US; 11.3 vs 11.5 in JP+EU). All 135 cells reproduce the stored operating-point targets to ≤ 1.1e-11, so the forecasts come from the stored validation configurations; nothing was re-selected.
Reading. Forecast blending does not change the conclusion. Against the position blend the difference is not significant in any sample, and its sign changes between samples. Neither blend beats the fixed deployed mode. Forecast blending also does not pass the gate: on F1–F2 it scores 0.0414 against O1's 0.0457 (Deployed ahead, p = 0.042), so the deployment decision is the same under either blend. Its significant results are against the static position blend in JP+EU (above) and against follow-the-leader on CVaR-95 in all three samples (+0.0080, p = 0.044 on F1–F2; +0.0062, p = 0.047 US; +0.0084, p = 0.0012 JP+EU). The blended object is therefore not what limits the method; the limit is the gap-to-noise ratio of the mode utilities (median gap 0.0016 against a gap s.d. of 0.032).
Revision. Sec. 4.4 gains a paragraph on why the mixture is over positions, states the equivalence condition (no constraints, κ = 0), and names forecast blending as the alternative; the E6 table and tests go to the appendix; Table 5 gains the forecast-blend row.
AuthorsOpenReview ↗

summarize of rebuttal revision

  • 2/2 positions vs forecasts → A3 (deployed ahead of the forecast blend in all three samples: +0.0043, +0.0008, +0.0017).
  • W1(a) standard algorithm → no PDF change.

Hope the revision and clarify in rebuttal can solve your concern and consider increase the rating of the paper.

AuthorsOpenReview ↗

Addtional results and friendly reminder

Dear reviewer uYuV

What does the setup add? We use standard Hedge/fixed-share updates to test whether reweighting three frozen forecasting modes helps when all three share an allocator and face the same trading costs. The validation gate selected the optimizer-only mode (O1), which ran throughout testing.

Why blend positions? Hedge scores the net utility of each mode’s traded portfolio. Blending positions weights the portfolios that earned those scores. Your control applies the same quarterly weights to forecasts, then optimizes once against the portfolio’s actual holdings. Asset caps and turnover costs can make the two allocations differ.

PolicyValidationU.S. testJapan–Europe test (pooled)
Fixed O1 (deployed)0.04570.08320.0417
Optimize each mode, then blend positions0.04170.08290.0386
Blend forecasts, then optimize once0.04140.08240.0400

Mean quarterly mean–variance utility from the 11 November rebuttal, net of 25 bps one-way trading costs. Higher is better.

Both blends fell below O1 on validation. Of the two blends, positions scored higher in the U.S. and forecasts scored higher in pooled Japan–Europe. Neither test established a forecast-blend advantage (one-sided p = 0.64 and 0.14, respectively). we would be grateful for any further comments, and we would ask you to consider whether the revised evidence solved your concerns also increase the overall score.

Best

The authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

overall summary of rebuttal revision

#LocationChange
B1Abstract, first sentence"Trained … system often exposes" → "A trained …"; the "which modules to consult …" clause put in parentheses
B2Intro, para. 2Comma and parenthesis fixes in the "we also call modes … Figure 1 …" sentence
B3Intro, contributions"Stress" → "stress"; the fragment "Shift the focus …" → "We shift the focus …"
B4Fig. 1 caption"deploying it via check regime shifts" → "deploying it behind a validation gate"
B5Sec. 3 (background), after the return equation"Costs depend …" → "Because costs depend …"
B6Sec. 4, router, allocator"Equation equation (n)" (repeated word) fixed at 3 places
B7Sec. 4.4 meta-policyLoss normaliser corrected to the code: one range shared by all modes (meta_policy._causal_scale)
B8Sec. 4.4"meta-policys" → "meta-policy's"
A1New App. C.2 "Additional checks requested in review", Table tab:rebuttalMarket impact (E1), random pools (E3), static mixture (E4), ensembles (E5), all as deployed − comparator ΔU with p values
A2App. C.2, paragraph "Loss normaliser"Shared vs per-mode range (E2)
A3App. C.2, paragraph "Blending positions versus forecasts"Why positions are blended, plus the forecast-blend result (E6)
A4Reproduction map (Table tab:repro)New row for the rb_e*.py scripts
A5main.bibAdded almgren2005direct

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.