acceptodds
Under review as a conference paper at ICLR 2027

PortfolioRouter: Consequence-Aware Multi-Task Routing for Portfolio Agents

Abstract

Portfolio agents can recommend materially different allocations for the same trading date, but standard best-expert classification does not weight errors by their financial cost. The cost of a routing error depends on the return forgone by choosing one proposal over another. We introduce PortfolioRouter, a consequence-aware portfolio router that learns when agent choice matters and evaluates the value of switching between proposals. Its multi-task routing module uses market-state representations and current portfolio proposals to jointly learn a general selector, a critical-date detector, a specialist trained for critical-date selection, and an auxiliary prediction of volatility-normalized relative asset returns. A default-relative value model then uses realized historical feedback to estimate whether changing the default is worthwhile. Its objective penalizes harmful replacements, and estimates across history lengths are combined conservatively. Selective deployment uses a gate calibrated on the preceding quarter to decide whether to add a switch. In U.S. and China backtests covering 2025 through June 2026, with same-close execution and zero transaction costs, PortfolioRouter achieves mean returns of 34.86% and 45.92%. These are 8.69 and 13.59 percentage points higher than the same models’ general paths. Retaining only the switches on retrospectively identified critical dates still adds 3.99 and 11.10 points. Code and artifacts are avaliable in https://anonymous.4open.science/r/PortfolioRouter.

Primary area:
reinforcement learning
Keywords:
LLM Agent, Router, Financial Market

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_qCFaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PortfolioRouter, a method for selecting among completed portfolio proposals when routing errors have different economic consequences. The paper separates date-level criticality, defined retrospectively using the realized gap between the best and second-best proposals, from candidate-specific switching value relative to a default selector. The method combines a multi-task routing model, a default-relative value model trained on benefit/harm targets, and a deployment gate calibrated on the preceding quarter.

The evaluation uses 23 U.S. ETFs and 25 China ETFs from 2025 through 2026 H1. Candidate allocations are generated by four language models, and the router is evaluated with quarterly walk-forward splits. The full method reports mean cumulative return / Sharpe of 34.86% / 1.51 in the U.S. and 45.92% / 1.87 in China. The paper also reports fixed-agent controls, component ablations, matched objective and predictor comparisons, and diagnostics on when switching contributes value.

Critical strengths
  1. The paper identifies a meaningful mismatch between routing accuracy and downstream portfolio value. Treating date-level criticality separately from the value of a particular replacement is a sensible formulation: an incorrect choice on a high-gap date can matter much more than several mistakes on low-gap dates. The switching analysis is consistent with this point; only slightly more than half of the overrides beat General, while the aggregate contribution is still positive.
  2. The experimental section makes a serious effort to separate routing effects from the quality of the proposal pool. In addition to broad baseline comparisons, the paper includes fixed-agent controls, task-removal ablations, alternative predictors under a common switching framework, and attribution/timing diagnostics. These controls are useful because they go beyond simply reporting a higher terminal return.
Critical weaknesses
  1. Mixed evidence for consequence-aware supervision. Table 12 shows that Economic Switching Value outperforms Ordinary Correctness in China but not in the U.S. Since the full model also benefits from estimator selection, the independent contribution of consequence-aware supervision remains unclear. A more controlled comparison is needed.
  2. Concerns about backtest validity. The evaluation assumes same-close execution with zero latency and transaction costs. Moreover, historical proposals generated by recent LLMs may contain future information. Delayed-execution tests, transaction-cost analysis, and stronger point-in-time validation would improve credibility.
  3. Limited methodological novelty. The framework combines established techniques, including multi-task learning, specialist selection, value estimation, and calibrated switching. Its contributions beyond existing cost-sensitive routing and decision-focused learning methods should be clarified.
  4. Insufficient explanation of performance gains. Although ablations support the effectiveness of individual components, they do not fully explain the substantial improvements over baselines. A clearer analysis of how critical-date detection, value estimation, and selective switching jointly contribute would strengthen the paper.
  5. Limited robustness evidence. Experiments cover only two markets over a relatively short period, with three runs per method. The substantial variance in China further raises concerns about stability. Additional evaluation across market periods and agent configurations would strengthen generalizability.
Questions for authors
  1. In the matched objective study (Table 12), Economic Switching Value improves over correctness-based objectives in China but not in the U.S. Which component explains the gap between these matched neural results and the much stronger full PortfolioRouter result? Can the authors provide a comparison that changes the supervision objective while keeping the selected estimator family and the rest of the switching pipeline fixed?
  2. Can the authors report the main results with an implementable execution delay and with transaction costs tied to realized turnover (for example, several reasonable per-trade or per-turnover cost levels)? I would like to know whether the gain over General remains after accounting for the extra switching.
  3. How were the four language-model proposal streams made point-in-time? In particular, what information was supplied to each model and what are the relevant training or knowledge-cutoff dates relative to the 2025-2026 evaluation period? If a fully point-in-time proposal pool is not available, please clarify how the reported results should be interpreted given the possible post-date knowledge noted in the paper.
5 comments
AuthorsOpenReview ↗

Clarification 1/2

We thank the reviewer for their valuable advice. We would like to address these concerns through controlled comparisons of supervision objectives, delayed-execution and robustness experiments, and further explanation of the method's design, switching gains and proposal-generation process.

Regarding whether estimator selection explains the gains, we ran two supervision controls and found economic value ahead of both correctness objectives in both markets. These controls extend the neural comparison in Appendix D.6.1, Table 12. In the first, every objective receives the same estimator families, search space and budget, with selection confined to prior-quarter validation. In the second, we fix the quarterly estimator family, architecture/training budget, protection weight, history aggregation and gate-score type to those selected by the original economic-value pipeline. Both controls share features, chronological training splits, base decisions and target normalization. Numerical gate thresholds are recalibrated for each objective on the preceding quarter, following Sec. 4.3. Three-seed mean cumulative returns are:

Control / marketOrdinary correctnessCritical-weighted correctnessEconomic value
Equal search space, U.S.32.98%28.13%34.86%
Equal search space, China44.09%40.28%45.78%
Original quarterly settings, U.S.32.48%27.05%34.86%
Original quarterly settings, China41.19%44.22%45.78%

Economic value exceeds both correctness objectives in both markets. The advantage remains when correctness objectives receive the same search budget, and when we compare objectives within the original selected models and switching settings.

The full system selects among estimator families (Appendix A, "Value-model search"), whereas Table 12 uses a dual MLP throughout. It selects ExtraTrees in four U.S. quarters and a direct MLP in one, while China uses dual MLP, random forest and ExtraTrees.

Regarding methodological novelty, our contribution is a portfolio-routing formulation that gives date-level criticality and candidate-specific switching value distinct roles in learning and deployment. The distinction matters because a date can have a large top-two return gap while the default already selects the best proposal. Criticality therefore supervises the detector and specialist (Secs. 3 and 4.1), while the value model learns the benefits and harms of each replacement, with an additional penalty for replacing an optimal reference choice (Sec. 4.2, Eq. (8)). At deployment, the router retains accepted specialist interventions and adds further switches only when predicted value relative to General is positive and the gate calibrated on the preceding quarter is satisfied (Sec. 4.3, Eq. (10)). Together, these targets and rules direct specialized learning toward economically important dates and assess each additional replacement on its own economic value. This design relates to multi-expert deferral [1], cost-sensitive deferral [2] and decision-focused prediction [3] through their focus on expert selection and downstream decision quality (Sec. 2).

References

[1] Rajeev Verma, Daniel Barrejon, and Eric Nalisnick. Learning to defer to multiple experts: Consistent surrogate losses, confidence calibration, and conformal ensembles. In Proceedings of the 26th International Conference on Artificial Intelligence and Statistics, volume 206 of Proceedings of Machine Learning Research, pp. 11415–11434. PMLR, 2023.

[2] Jean V. Alves, Diogo Leitão, Sérgio Jesus, Marco O. P. Sampaio, Javier Liébana, Pedro Saleiro, Mário A. T. Figueiredo, and Pedro Bizarro. Cost-sensitive learning to defer to multiple experts with workload constraints. arXiv preprint arXiv:2403.06906, 2024.

[3] Adam N. Elmachtoub and Paul Grigas. Smart “predict, then optimize”. Management Science, 68(1):9–26, 2022.

AuthorsOpenReview ↗

Clarification 2/2

Regarding the source of the switching gains, we decomposed the value branch's replacements into benefits and harms and found that beneficial switches contributed more than harmful switches lost in both markets. This extends the switching analysis in Sec. 5.3 and Appendix D.5.2. Adding the value branch to the base policy, formed by the specialist and forest filter, raises return from 27.80% to 34.86% in the U.S. and from 35.76% to 45.92% in China. About 92 and 96 decisions change per run; 55% and 53% improve on the base action. Positive log-return contributions total 1,361 versus 822 bps of harm in the U.S., and 2,006 versus 1,294 bps in China. Successful replacements contribute more than failed ones lose, explaining the aggregate gain even though only slightly more than half of the switches improve on the base action.

Regarding whether extra switching remains worthwhile after trading delay and costs, we replayed the signals at the next open with turnover-based fees and found positive net gains over General in both markets. We kept the original signals fixed (Sec. 5.1) and charged fees on ETF purchases plus sales after price drift. All policies use the same calendar and carry holdings continuously across quarter boundaries. At 5 bps, PortfolioRouter returns 21.27% versus General's 13.49% in the U.S., and 28.64% versus 22.67% in China: gains of 7.78 and 5.97 percentage points. Sharpe ratios are 0.969 versus 0.683 and 1.223 versus 1.008. PortfolioRouter also exceeds all compared routers and fixed LLM experts within the shared proposal pool at this setting, with lower traded notional than General. The next-open advantage over General persists across 0/1/5/10/20 bps, so the gain remains after charging for the ETF trades required by the selected allocations.

Regarding proposal provenance, we examined the prompt template and generation logs, which document date-filtered inputs and retrospective generation on September 20, 2026. All four generators in Sec. 5.1 received the decision date, ETF universe, current and previous seven sessions of OHLCV, current prices and date-filtered macro observations. All four used the same balanced role, temperature 0.3, and long-only allocations with optional cash. The requested model identifiers in the logs are gpt-5.4, qwen3.6-plus, deepseek-v4-flash and gemini-3.5-flash. The logs record the requested aliases but do not establish exact provider model versions or training cutoffs, so possible post-date model knowledge remains a limitation (Sec. 6). The original fixed-agent controls compare policies on the same cached proposals (Appendix D.3).

Regarding dependence on the LLM proposal pool, we retrained the complete pipeline with four price-only experts and found gains over General in both markets over 2025–2026 H1. The experts use equal weight, 20-session momentum, 20-session inverse volatility and a 60-session trend rule. Each uses inputs available through the signal date. Under the same next-open/5-bps setting, the results are:

Price-only poolGeneral returnPortfolioRouter returnGain (percentage points)
U.S.40.35%42.82%+2.47
China32.30%39.16%+6.86

Both markets retain positive routing gains when every proposal is constructed from historical numerical inputs. We also evaluated the previously unused 2026Q3 quarter, where return differences from General are +0.32 percentage points in the U.S. and −3.84 percentage points in China.

Regarding uncertainty from a shared market history, we used paired block bootstrapping and found positive 95% intervals for the U.S. gain over General at all tested block lengths. The three-seed results share one market history (Appendix C.2). We used 2,000 paired block-bootstrap replicates at lengths of 5/10/20 trading days under the original scoring convention. Both the U.S. return and Sharpe improvements over General have positive intervals at all three lengths.

Regarding China's seed dispersion, we retrained with a 70th-percentile criticality threshold and found a positive routing gain with lower variation across seeds. We held the architecture, seeds and training budgets fixed when examining the dispersion reported in Appendix D.1, Table 5. The retrained variant gives 39.66% ± 1.78%, a gain of 5.33 percentage points over its own General head.

AuthorsOpenReview ↗

Follow-up on supervision controls and backtest validity

Dear Reviewer qCFa,

Thank you again for your detailed review. We are following up on our November 10 responses, “Clarification 1/2” and “Clarification 2/2.” Your questions motivated several additional experiments. For convenience, we summarize the evidence most directly relevant to Q1–Q3:

  1. Contribution of economic supervision beyond estimator selection (Q1). With the quarterly estimator families fixed to those selected by the original pipeline, economic supervision achieves returns of 34.86% versus 32.48% for the strongest correctness objective in the U.S., and 45.78% versus 44.22% in China. Each objective’s numerical gate threshold is calibrated on the preceding quarter. The equal-search-budget comparison also favors economic supervision in both markets. Together, these controls provide evidence that the objective contributes beyond estimator-family selection.

  2. Gains under delayed execution and turnover-based costs (Q2). With next-open execution and 5-bps costs on actual ETF purchases and sales, PortfolioRouter retains net return gains over General of 7.78 and 5.97 percentage points in the U.S. and China, respectively, alongside higher Sharpe ratios. Under this setting, it also exceeds all compared routing baselines and fixed LLM experts within the shared proposal pool. Its traded notional is lower than General’s in both markets.

  3. Evidence without pretrained LLM proposals (Q3). We retrained the complete pipeline using four price-only experts constructed from numerical observations available through the signal date. Under next-open execution and 5-bps costs, the gains over General are 2.47 and 6.86 percentage points. This provides complementary evidence for the routing mechanism without reliance on pretrained LLM knowledge. Possible post-date knowledge remains a limitation of the original LLM backtests.

Our central methodological contribution is the separation of date-level criticality for specialized learning from candidate-specific, default-relative switching value. The additional controls strengthen the empirical evidence for this formulation.

We would be grateful if you could reconsider your overall assessment in light of these additional results. If any concerns remain decisive for your recommendation, could you please identify them so that we can address them during the remaining discussion period?

Thank you for your time and consideration.

Reviewer_2uP1OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces PortfolioRouter, a consequence-aware framework for selecting among portfolio agents. Rather than treating all selection errors equally, the method considers when agent choice matters and whether switching from a default selection is worthwhile. It combines multi-task learning for general selection and critical-date specialization with a value model estimating the benefits and harms of switching, followed by a calibrated deployment mechanism. The authors evaluate PortfolioRouter through walk-forward backtests on U.S. and Chinese ETF markets, reporting favorable performance in terms of cumulative returns and Sharpe ratios compared with the baselines. They also conduct ablation studies and switching analyses to examine the roles of key components.

Critical strengths
  1. PortfolioRouter achieves higher cumulative returns than the best fixed-agent policy identified retrospectively in both markets. Since all policies use the same portfolio proposals, evaluation dates, and return calculations, the improvement cannot be attributed to differences in proposal generation. This provides evidence that selecting agents over time can add value beyond committing to a single agent throughout the evaluation period.

  2. Beyond standard ablation studies, the paper also examines two specific questions. Critical-date attribution shows how gains are distributed between critical and ordinary dates, while randomized switching-time experiments show that the selected dates for additional switches outperform conditional random alternatives with the same switching frequency. Together, these analyses support the paper's central hypothesis that routing decisions have different economic consequences across dates and that switching timing matters beyond switching frequency.

Critical weaknesses
  1. The backtests use market information available through the close of day tt, while assuming execution at the same closing price. This raises a timing concern, as the final closing information may not be available before trading orders must be submitted. Although the authors acknowledge this limitation, the lack of delayed-execution analysis makes it difficult to assess whether the reported advantages over the baselines are robust to a feasible trading schedule.

  2. The paper defines critical dates using the 80th percentile of the top-two expert-return gap, but does not report comparisons with alternative thresholds, such as the 70th or 90th percentile. Since this threshold determines which dates are used to supervise the specialist, its choice directly affects the training process. This leaves open how much of the reported gain from critical-date specialization depends on the choice of threshold.

Questions for authors
  1. The results are reported without transaction costs, so it is unclear how differences in portfolio turnover affect the comparison. Do PortfolioRouter’s return and Sharpe advantages over the strongest routing and fixed-agent baselines hold when the same transaction-cost assumptions are applied to all policies?

  2. The reported standard deviations capture variation across seeds, but all runs share the same market history. How strong is the evidence for the return and Sharpe advantages over the strongest routing baselines when temporal dependence is taken into account? The changes in the U.S. return rankings in 2026 H1 make this particularly relevant.

3 comments

We thank the reviewer for their valuable advice. We would like to address these concerns through experiments with delayed execution, turnover-based costs and alternative critical-date thresholds, paired block-bootstrap analysis, and further explanation of the evaluation protocol and statistical comparisons.

Regarding the execution-timing and transaction-cost concerns, we replayed signals at the next open with trading fees and found higher net return and Sharpe than General in both markets. We kept the original signals fixed and delayed trades from the same close (Sec. 5.1) to the next trading open. Fees apply to actual ETF purchases plus sales, calculated from holdings after price drift. Every policy uses the same execution calendar and cost assumptions. At 5 bps, the three-seed mean results are:

Next-open executionPortfolioRouter return / SharpeGeneral return / Sharpe
U.S.21.27% / 0.96913.49% / 0.683
China28.64% / 1.22322.67% / 1.008

The net return gains over General are 7.78 and 5.97 percentage points. PortfolioRouter also exceeds every compared routing baseline and fixed LLM expert within the shared proposal pool in both return and Sharpe. It trades less than General: mean daily traded notional is 0.513 versus 0.555 in the U.S. and 0.625 versus 0.655 in China. The advantages therefore survive delayed execution and costs tied to realized ETF turnover.

The next-open advantage over General persists at every tested fee level: 0/1/5/10/20 bps. At the next close with 5 bps, the return differences are +3.89 percentage points in China and −1.32 percentage points in the U.S. The execution analysis carries holdings continuously across quarter boundaries; the threshold and bootstrap analyses below follow the paper's original scoring convention.

Regarding dependence on the critical-date threshold, we retrained at alternative quantiles and found that the original 80th-percentile setting gave the highest return and Sharpe in both markets. We define critical dates using the 80th percentile of the top-two return gap (Sec. 3, Eq. (3)). The new variants retrain the joint backbone (Sec. 4.1) at the 70th and 90th percentiles, keeping the architecture, seeds and training budgets fixed, and recalibrate the downstream branches. This changes the detector labels and the dates used to train the specialist. The 80th-percentile row reproduces Sec. 5.2, Table 1; the other rows are newly retrained variants. The three-seed means are:

Training quantileU.S. return / SharpeChina return / Sharpe
70th percentile24.20% / 1.10939.66% / 1.617
80th percentile34.86% / 1.51445.92% / 1.867
90th percentile20.03% / 0.92936.13% / 1.495

The original 80th-percentile setting gives the highest return and Sharpe in both markets. China retains gains over the corresponding retrained General head at all three thresholds; at the 70th percentile, the return is 39.66% ± 1.78% across seeds, a gain of 5.33 percentage points over its own General.

Regarding evidence beyond seed variation, we used paired block bootstrapping and found positive 95% intervals for the U.S. gain over General at all tested block lengths. The three-seed results (Appendix C.2) share one market history. We use 2,000 paired block-bootstrap replicates with block lengths of 5/10/20 trading days to measure uncertainty across time. Shared blocks preserve short-range temporal dependence and pair every policy on the same sampled dates. Under the original scoring convention, the estimated return advantages over the strongest full-period routing comparators are +2.35 percentage points versus U.S. Ridge [−7.81, 11.38] and +5.43 percentage points versus China A2C [−7.50, 17.35], using 10-day blocks. The corresponding Sharpe differences are also positive; the 95% intervals for these return and Sharpe comparisons include zero.

Against General, the U.S. return and Sharpe improvements have positive 95% intervals at all three block lengths. At length 10, the return gain is 8.69 percentage points [1.08, 18.47] and the Sharpe gain is 0.274 [0.032, 0.543]. The U.S. gain over General is therefore supported by resampling the market history as well as by the seed comparisons.

Thank you for the detailed rebuttal. While the additional experiments address my concerns regarding execution timing and critical-date threshold sensitivity, the block-bootstrap results remain inconclusive about whether PortfolioRouter significantly outperforms the strongest routing baselines, with the 95% confidence intervals including zero. I therefore maintain my original positive rating of 3.

Dear Reviewer 2uP1,

Thank you for reviewing the additional experiments and clarifying your assessment. We would also like to highlight the supervision controls reported in “Clarification 1/2,” which may be relevant to evaluating the method’s contribution.

With the quarterly estimator families fixed to those selected by the original pipeline, economic supervision achieves returns of 34.86% versus 32.48% for the strongest correctness objective in the U.S., and 45.78% versus 44.22% in China. Each objective’s gate threshold is calibrated on the preceding quarter. The equal-search-budget controls also favor economic supervision in both markets.

These comparisons help isolate the contribution of economic supervision beyond estimator-family selection, complementing the full-system baseline comparisons. Together with the delayed-execution and transaction-cost results, they provide additional evidence for the paper’s central formulation.

We would appreciate your thoughts on whether these controlled comparisons strengthen the evidence for the proposed economic supervision.

We would be happy to clarify any remaining questions.

Reviewer_7akrOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PortfolioRouter for making daily selections among candidate asset allocations provided by multiple portfolio agents, which jointly trains general selection, key date detection, expert selection on key dates, and auxiliary asset return forecasting. It conducts rolling backtests covering the period from 2025 to the first half of 2026 on ETF markets in the United States and China, with comparision on the complete method with a fixed agent, a supervised learning router, a reinforcement learning router, and existing financial systems, and analyzed the contributions of each component and switching behavior to returns.

Critical strengths
  • It models the profit gap between each set of candidate configurations for a given date, as well as the potential gains and losses resulting from replacing the default configuration. This enables routing decisions to account for the economic consequences of different options and better align with the actual decision-making objectives of portfolio routing.

  • The experiment covers a wide range of baseline including fixed agents, return forecasting, and reinforcement learning across two markets and provides results from matched controls, component ablation tests, and multiple random seed runs.

    • It achieves the highest average Sharpe ratio among the benchmarks in both markets, with cumulative returns over the entire period increasing by 8.69 and 13.59 percentage points, respectively, compared to the general strategy.
Critical weaknesses
  • The inputs for the candidate LLMs were filtered by date, but they may still contain prior knowledge from after the simulated decision date. The paper does not fully disclose the complete prompt, model version, and actual generation time when the candidate configurations were generated. And potential information leaks in the historical backtesting have not been fully ruled out.

  • The preceding quarter’s validation data are used to select the model configuration, training settings, protection weight, history aggregation scheme, and gating threshold, while each quarter contains only approximately 60 trading days. Although the authors apply checks for consistent incremental gains and fallback rules, it remains unclear how sensitive the selected configurations and switching thresholds are to the choice of validation window. Would modest changes to this window lead to materially different configurations or performance on the same test periods?

Questions for authors

See weaknesses above.

3 comments

We thank the reviewer for their valuable advice. We would like to address these concerns through validation-window sensitivity experiments and an evaluation with a price-only expert pool, together with further explanation of the proposal inputs, model identifiers and generation timing.

Regarding the validation-window concern, we repeated calibration under six window definitions and found higher return and Sharpe than General in both markets under every definition. The original calibration uses the preceding quarter's validation data (Sec. 4.3; Appendix B, "Calibration"). For each alternative window, we reselected the value estimator, protection weight, history aggregation and gate. We held the trained predictors, LLM proposals, scoring convention and test dates fixed. The table reports three-seed mean cumulative returns; the full-quarter and General rows reproduce Appendix D.1, Tables 4–5, while the other rows are new comparisons:

Validation windowU.S. returnChina return
Full preceding quarter34.86%45.92%
Latest 40 dates32.62%42.29%
Latest 50 dates32.05%43.61%
Latest 60 dates32.46%46.08%
Drop first 5 dates31.94%43.95%
Drop last 5 dates29.17%45.45%
General reference26.17%32.33%

All six windows yield higher return and Sharpe than General in both markets. The calibration gains therefore persist across the tested changes in window length and boundaries.

Regarding the window's effect on training selection, we reran the full search and retained gains in three of four market–window comparisons, with the remaining U.S. result close to General. After dropping the first or last five validation dates, we reran the ten-candidate search, checkpoint selection and downstream calibration (Sec. 4.1; Appendix A, "Backbone training and selection"). Each retrained router is compared with its own General head to measure the additional routing gain with the new backbone. In the U.S., dropping the first five dates gives a gain of 2.53 percentage points; dropping the last five gives a return slightly below General, 26.71% versus 27.05%. China retains gains of 6.55 and 11.84 percentage points, with seed standard deviations of 2.28 and 2.69 percentage points, respectively.

Regarding proposal provenance, we examined the prompt template and generation logs, which document date-filtered inputs and retrospective generation on September 20, 2026. For the four generators in Sec. 5.1, we provide the source prompt template, output schema and generation logs. Each model received the decision date, ETF universe, current and previous seven sessions of OHLCV, current prices and date-filtered macro observations. All four used the same balanced role, temperature 0.3, and long-only allocations with optional cash. The requested model identifiers in the logs are gpt-5.4, qwen3.6-plus, deepseek-v4-flash and gemini-3.5-flash. The logs record the requested aliases but do not establish exact provider model versions or training cutoffs, so possible post-date model knowledge remains a limitation (Sec. 6). The original fixed-agent controls compare policies on the same cached proposals (Appendix D.3).

Regarding possible dependence on pretrained LLM knowledge, we retrained the complete pipeline with four price-only experts and found gains over General in both markets over 2025–2026 H1. The four experts use explicit rules: equal weight, 20-session momentum, 20-session inverse volatility and a 60-session trend rule. Each uses only numerical observations available through the signal date. We evaluated the retrained pipeline with next-open execution and 5 bps charged on ETF purchases plus sales:

Price-only poolGeneral returnPortfolioRouter returnGain (percentage points)
U.S.40.35%42.82%+2.47
China32.30%39.16%+6.86

The gains of 2.47 and 6.86 percentage points show that PortfolioRouter adds value over General without LLM-generated proposals, while accounting for delayed execution and transaction costs.

AuthorsOpenReview ↗

Gentle follow-up on our response

Dear Reviewer 7akr,

we would like to follow up on our response, including the six validation-window comparisons, the full training-selection sensitivity tests, and the price-only expert evaluation.

Could you let us know whether any concerns about validation sensitivity or proposal provenance remain unresolved? We would be happy to provide further clarification.

Best regards,

The Authors

I thank the authors for the additional results and the detailed rebuttal. I will maintain my original score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.