acceptodds
Under review as a conference paper at ICLR 2027

Learning When and How to Trade on Government Policies

Abstract

Government policies influence financial markets, but supportive language does not necessarily imply profitable trading opportunities. We study when policy information adds value to existing trading strategies and how this value evolves across trading days. We construct PolicyValue-Bench, a point-in-time benchmark aligning 5,125 government policy documents from 2021-2025 with policy histories and market data for five Chinese A-share sector ETFs. We present PolicyState, a stateful gating framework that assesses policy text and LLM-generated judgments in the context of market conditions and retained policy effects. It learns policy admission, opportunity retention, and daily risk release through separate opportunity and risk states that guide exposure increases and cash exits. Training combines matched-control return supervision with sequential trading optimization. In a full-year 2025 out-of-sample evaluation across four base strategies, our method achieves an average net cumulative return of 16.25%, representing relative improvements of 140.4% over matched bases and 54.0% over the strongest comparison method. It improves returns in 19 of 20 sector-strategy combinations.

Primary area:
learning on time series and dynamical systems
Keywords:
Policy-driven trading, Financial language models, Financial benchmarks, Stateful decision-making, Incremental trading value

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_VKWeOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper introduces PolicyValue-Bench, a dataset of 5,125 Chinese government policy documents covering 2021–2025, and PolicyState, a framework designed to incorporate policy information into existing trading strategies. The model combines policy text, LLM-generated assessments, market features, and recurrent opportunity and risk states to determine exposure adjustments and cash exits. The authors evaluate the approach across five Chinese sector ETFs and four underlying trading strategies. The main results cover 2025, with an additional expanding-window evaluation for 2024. PolicyState reportedly achieves an average net cumulative return of 16.25%, compared with 6.76% for the underlying strategies, alongside improvements in Sharpe ratio and maximum drawdown.

Critical strengths

Relevant problem and potentially useful dataset: The paper addresses an interesting question, namely, whether government policy announcements contain incremental trading information beyond what is already reflected in market prices. The construction of a dataset linking policy documents, historical announcements, and sector ETF returns is a useful contribution.

Reasonably comprehensive model comparisons and ablations: The authors compare their approach against several text-based models, LLMs, and financial tradnig agents, as well as a more directly comparable GRU architecture. The ablation studies also help assess the contributions of the different components. These are useful experiments, although they do not fully resolve the concerns about the financial evaluation.

Critical weaknesses

Lack of volatility targeting and risk-normalized performance comparisons: My main concern is the absence of a proper risk-normalized evaluation. The authors report substantial improvements in cumulative returns and maximum drawdowns, but do not report the realized annualized volatility of the individual strategies or compare their performance under a common volatility target. This is a significant omission for a paper evaluating financial trading strategies. Without knowing the volatility and risk exposure of each method, it is difficult to interpret the differences in cumulative returns and maximum drawdowns. A strategy generating higher returns may simply be taking more risk, while lower drawdowns may reflect lower exposure rather than better investment decisions. The reported Sharpe and Sortino ratios provide some risk-adjusted information, but their statistical uncertainty is not reported. In particular, there are no confidence intervals or significance tests for the differences in Sharpe, Sortino, or maximum drawdown against the competing methods. Given that the paper's main claims are based on improvements in trading performance and risk management, I consider this a fundamental weakness of the experimental design. The authors should report realized annualized volatility, average exposure, and performance under comparable risk budgets, rather than relying primarily on raw cumulative returns and drawdowns.

Limited out-of-sample evaluation and insufficient evidence of statistical robustness: The main experimental results are based on a single out-of-sample year, 2025. Although Appendix D includes an expanding-window evaluation for 2024, this amounts to only two annual test periods, which is limited for a financial strategy intended to operate under changing market conditions. A more extensive walk-forward evaluation would be needed to assess whether the reported improvements are consistent across different market regimes, particularly periods of elevated volatility, market stress, and changing policy conditions.

Furthermore, the 20 sector-strategy combinations should not be interpreted as 20 independent tests. The four strategies within each sector trade the same underlying ETF and are exposed to the same market movements and policy announcements. The authors provide a bootstrap confidence interval for the cumulative return improvement over the matched base strategies, which is useful. However, the reported p-value of 0.0382 is relatively close to the conventional significance threshold, and comparable uncertainty estimates are missing for the risk-adjusted metrics and the performance difference against the strongest architectural baseline. Taken together, these limitations make it difficult to assess how robust the reported improvements really are.

Questions for authors

Volatility Targeting: Why have the authors not reported realized annualized volatility for each strategy? Could they provide a comparison of all methods under a common volatility target or risk budget, including returns, Sharpe ratios, and maximum drawdowns?

Confidence Intervals: Could the authors provide confidence intervals or statistical significance tests for teh differences in Sharpe ratios, Sortino ratios, and maximum drawdowns, particularly against the full-input GRU baseline?

Out-of-sample performance: Why is the out-of-sample evaluation limited to 2024 and 2025? Could the authors provide a more extensive walk-forward analysis, or at least a breakdown of performance across different market and volatility regimes?

Impact of policy signals: Could the authors clarify how much of the reported improvement comes from policy-driven trading signals rather than changes in average market exposure or the model's ability to exit positions during market declines?

Reviewer_u62qOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces PolicyValue-Bench and PolicyState, a policy-conditioned overlay for existing ETF trading strategies. Frozen text and judgment features feed separate opportunity and risk states, trained first with matched-control return targets and then through sequential trading optimization. The selected ensemble improves mean 2025 account return from 6.76% to 16.25% across five sectors and four base strategies.

Critical strengths

The task distinguishes favorable policy language from incremental trading value. The explicit admission, retention, and risk-release structure is interpretable as a modeling design, and the execution protocol includes next-open trading and concrete account constraints. The appendix contains important controls: a parameter-matched full-input GRU, market-only and semantic controls, past-policy replacement, synchronized block-bootstrap inference, two expanding-window test years, and substantial cost sensitivity. These controls make the evidence considerably stronger than the main return table alone.

Critical weaknesses
  1. Architecture and ensemble effects are not sufficiently separated. Table 8 reports single-model mean returns of 12.60% for PolicyState and 11.53% for the full-input GRU, whereas the selected ensembles achieve 16.25% and 11.03%. Much of the headline gap therefore emerges through aggregation and thresholding. This is not evidence of invalid selection, but the full seed/ensemble/search budgets, matched ensemble-selection controls, and a paired confidence interval against the GRU are necessary to establish the architectural contribution. Uncertainty relative to the original base strategies does not resolve this comparison.

  2. The point-in-time evidence remains narrower than the intended practical claim. Appendices A.2 and H state that policy availability is recorded by date without first-seen timestamps or revision histories, and that pretrained knowledge boundaries are not established. I am not asserting that leakage occurred. A one-session delayed-policy control and auditable document/model versions would quantify whether the reported advantage depends materially on the availability convention. This matters because the paper's contribution is actionable incremental policy value, rather than retrospective association.

Questions for authors
  1. What are the matched seed/ensemble/search budgets and paired uncertainty for PolicyState versus the full-input GRU? How much advantage remains under identical ensemble-selection rules?
  2. Does the advantage persist when policy availability is delayed by one trading session?
Reviewer_QxBHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper examines when a government policy announcement should prompt a change to an existing trading position, and how long its influence should persist. It introduces PolicyValue-Bench, which comprises 5,125 policy documents linked to five Chinese sector ETFs, and PolicyState, which integrates policy text, LLM judgments, and market information to maintain separate opportunity and risk states. These states are used to guide additional exposure and exits to cash. Training combines matched-return supervision with direct optimization of trading performance. In the main 2025 evaluation, average net returns are reported at 16.25%, compared with 6.76% for the base strategies, with supporting evidence from component ablations, matched neural controls, and an additional 2024 evaluation.

Critical strengths

1/ The model addresses a meaningful aspect of policy-driven trading. By separating opportunities from risks, the model can retain a useful policy signal while temporarily exiting a position. Together with the benchmark's focus on improvements over an existing strategy, this gives the work a clear purpose beyond merely predicting policy sentiment.

2/ The controlled comparisons strengthen the empirical contribution. The full-input GRU is closely matched to the proposed model in inputs, training procedure, and parameter count. The market-input and policy-replacement controls also help establish that policy content contributes to the gains. Positive results in both evaluation years, along with the transaction-cost sensitivity analysis, further support the practical relevance of the results.

Critical weaknesses

1/ My major concern is that the contribution appears incremental compared to existing work. Using policy or financial text to guide trading, retaining information over time, and optimizing trading decisions have already been studied in the literature discussed in Section 2. The distinctive contribution here seems to lie mainly in the industry-policy dataset and the particular opportunity/risk recurrence built on frozen LLM representations. While these are useful additions, the paper does not clearly establish what broader methodological insight they provide beyond adapting existing ideas to this setting.

2/ The advantage over the matched GRU depends substantially on ensembling. In Table 8, individually executed models average 12.60% for PolicyState and 11.53% for the GRU, whereas their ensembles achieve 16.25% and 11.03%, respectively. The uncertainty analysis compares PolicyState with the base strategies, but not with this stronger architectural control. Without a fuller account of seed and ensemble selection, I am less confident that the headline margin represents a consistent architectural advantage.

3/ The evidence for generalization remains limited. While the additional 2024 evaluation is useful, both test years cover the same five sectors, which were selected for high policy density. The four base strategies reuse each sector's market history, and the sector-omission analysis changes only the aggregation rather than testing transfer to an unseen sector. This leaves uncertainty about performance outside the particular assets and periods studied.

Questions for authors

1/ How many random seeds and ensemble combinations were evaluated on validation data for PolicyState and the full-input GRU? Were the candidate sets and selection budgets comparable for the two?

2/ Will the benchmarking, leaderboard or training/evaluation dataset be open-sourced?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.