acceptodds
Under review as a conference paper at ICLR 2027

Trading Strategy Optimization via Textual Gradient

Abstract

Quantitative trading strategy design aims to discover trading programs from historical data that remain effective in future markets, which can be viewed as a black-box program optimization problem. LLM-based textual gradients offer a promising approach by providing explicit optimization directions for iterative strategy refinement. However, directly applying textual gradients faces two challenges: (1) optimization is myopic, underutilizing experience from previous evaluations; and (2) aggregate backtest feedback overlooks temporal robustness, potentially favoring strategies that perform well only in specific market periods. To address these challenges, we propose TradeGrad, an experience-guided textual-gradient framework for robust trading strategy optimization. TradeGrad leverages accumulated optimization experience to estimate textual gradients and employs multi-scale revisions for both strategy exploration and refinement. It further introduces the Cross-Period Robust Objective (CPRO), which emphasizes performance in unfavorable historical periods to promote temporal robustness. Experiments on cross-sectional and time-series strategy design in Chinese A-share and U.S. equity markets show that TradeGrad achieves the best in-sample and out-of-sample performance across all four settings. Notably, its Chinese cross-sectional strategy achieves 27.99% annualized return, 12.19% maximum drawdown, and a Sharpe ratio of 1.63, approximately 68% higher than the CSI 300 benchmark. Further analyses validate the proposed components and show consistent improvements in both in-sample and out-of-sample performance throughout optimization. The code is available at https://anonymous.4open.science/r/TradeGrad-7FC6.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Quantitative Trading, Textual Gradient, Program Optimization

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

This paper proposes TradeGrad, which treats trading strategy design as black-box program optimization and iteratively revises strategies through experience-guided textual gradients. It combines global optimization memory, archived-strategy sampling, and multi-scale program revisions. It also introduces CPRO, a cross-period objective that emphasizes poorly performing historical periods. Experiments cover cross-sectional and time-series strategy design on CSI 300 and S&P 500, with out-of-sample backtests, component ablations, and optimization-trajectory analyses.

Critical strengths

S1. TradeGrad combines optimization memory, comparisons across strategies, and multi-scale revisions so that textual gradients can draw on previous search experience. The ablations in Table 2 support these components under the same objective and search budget.

S2. The evaluation covers two markets and two strategy settings, with transaction costs included in backtesting. The out-of-sample results and component ablations provide preliminary support for the method’s practical value.

Critical weaknesses

W1. The main comparisons use different optimization objectives. TradeGrad searches and selects strategies using CPRO, while most baselines use in-sample Sharpe ratio. Although Table 3 evaluates CPRO within TradeGrad, it does not establish how much of the advantage over other methods comes from the optimizer rather than the objective.

W2. The out-of-sample evidence for temporal robustness is indirect. CPRO emphasizes poorly performing historical periods, but the out-of-sample evaluation mainly reports metrics aggregated over the entire period. It remains unclear whether the method improves performance during unfavorable or worst-performing periods.

W3. Some advantages are small relative to run-to-run variability. For example, the CSI 300 cross-sectional OOS Score is 65.65, compared with 60.69 for the Volatility baseline, while TradeGrad’s standard deviation is 12.15. Three runs with means and standard deviations provide limited evidence that these smaller advantages are reliable.

Questions for authors

Q1. Could the authors compare TradeGrad with the main automated-search baselines under the same optimization and strategy-selection objective, such as CPRO, while keeping the search budget fixed? This would help separate the benefit of the textual-gradient optimizer from that of changing the objective.

Q2. How does TradeGrad perform across individual out-of-sample periods, particularly unfavorable ones? Could the authors report results by year or other predefined time segments, including worst-period or lower-tail performance, and compare them with objectives that do not use lower-tail aggregation?

Q3. Could the authors increase the number of independent runs and report performance differences with confidence intervals against the main competing baselines, particularly Volatility in the CSI 300 cross-sectional setting? If the analysis uses daily returns, please account for temporal dependence.

Reviewer_nqe7OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose in this paper an experienced-guided textual gradient method, named TradeGrad, to generate trading strategies by LLMs in an evolution manner. Basically, TradeGrad introduces the following major components.

  • To take advantage the successful and failure experience from previous trading strategies and backtest results, TradeGrad proposes to periodically employ LLMs (S_local and S_global function) to extract past experience into the global optimization memory.
  • To avoid trading strategies overemphasize on some short and favorable periods, TradeGrad proposes the concept of CPRO to make trading strategies maximize the average score of the bottom 30% of periods, thereby facilitating temporal robustness.

Experiments show that TradeGrad is able to generate strategies better than other evolutional approaches in CSI 300 and S&P 500.

Critical strengths

None

Critical weaknesses

None

Questions for authors

It seems that SglobalS_{global} and SlocalS_{local} are executed by LLMs. If so, please explicitly mention this in the revision.

As mentioned in Section 4.1 and Section 5, in general-purpose program evolution methods, AlphaEvolve (2025) and ShinkaEvolve (2026) represent more recent developments compared to the traditional Island algorithm. In addition, in domain-specific program evolution, ProFiT (2026) is more recent than QuantEvolve (2025). The authors are encouraged to conduct experiments to compare their performance.

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

The paper proposes TradeGrad, an LLM-based framework for optimizing executable quantitative trading strategies using textual gradients. It addresses two limitations of directly applying textual-gradient optimization to trading: individual updates can be myopic and fail to reuse experience accumulated during search, and optimization based on aggregate historical performance may favor strategies whose success is concentrated in favorable market periods.

TradeGrad addresses the first issue using an archive of evaluated strategies, periodically consolidated optimization memory, performance-biased sampling of reference strategies, and multi-scale updates that alternate between structural rewrites and localized edits. It addresses the second issue through the Cross-Period Robust Objective (CPRO), which evaluates strategies across historical windows and optimizes the lower tail of the window-level scores.

The method is evaluated on both cross-sectional and time-series strategy design in Chinese A-share and U.S. equity markets. TradeGrad obtains the highest reported IS and OOS scores among the evaluated strategy-design methods in all four settings. Ablations study optimization memory, reference-strategy sampling, multi-scale revision, and CPRO, together with hyperparameter and optimization-trajectory analyses.

Critical strengths
  1. The paper presents a well-motivated adaptation of textual-gradient optimization to a setting with temporal distribution shift. TradeGrad does more than repeatedly ask an LLM to improve a program: it accumulates global optimization experience, compares sampled strategies and their outcomes when constructing textual gradients, and combines structural exploration with localized refinement. These components form a coherent optimization framework rather than an ad hoc trading-agent pipeline.

  2. The empirical evaluation is relatively broad and provides consistent evidence in favor of the proposed method. TradeGrad is evaluated on Chinese A-share and U.S. equity markets under both cross-sectional and time-series strategy-design settings and achieves the highest reported IS and OOS scores among the evaluated strategy-design methods in all four settings. The paper also reports three-run results and ablations showing that optimization memory, sampling, multi-scale revisions, and CPRO each contribute to the final performance.

Critical weaknesses
  1. The evidence that CPRO specifically improves temporal generalization is suggestive but not fully conclusive. The available historical data are divided into one chronological IS/OOS split, and strategy optimization uses the IS period while the single held-out OOS period is used for final evaluation. Consequently, the OOS improvement provides evidence for generalization to that particular future period, but does not establish that CPRO consistently improves robustness across different temporal shifts or market regimes. This matters because temporal robustness is a central motivation and contribution of the paper. Multiple rolling or walk-forward IS/OOS splits would provide substantially stronger evidence for this claim.

  2. The contribution of CPRO is intertwined with a fairly engineered objective and several design choices. CPRO combines sigmoid-normalized annualized return and maximum drawdown with reference levels and sensitivity parameters, a trading-frequency gate, a weighting coefficient, and a fixed 30% lower-tail aggregation. Although the paper provides ablations and some sensitivity analysis, it remains unclear whether the gain comes primarily from the proposed lower-tail cross-period principle or from favorable objective engineering for this benchmark. A comparison against simpler robust alternatives using otherwise matched scoring would make the methodological contribution of CPRO more convincing.

Questions for authors
  1. Temporal generalization: Since temporal robustness is a central motivation for CPRO, do the authors have results under multiple rolling or walk-forward IS/OOS splits rather than the single chronological split used in the main experiments? I would be particularly interested in whether CPRO consistently improves OOS performance when the optimization and held-out periods encounter different market conditions. Such evidence would materially increase my confidence in the paper's central robustness claim.

  2. What aspect of CPRO drives the improvement? CPRO combines cross-period evaluation, lower-tail aggregation, sigmoid normalization of return/drawdown, and a trading-frequency gate. Could the authors further isolate the lower-tail component, for example by comparing lower-tail versus mean aggregation while keeping the per-period score and frequency treatment identical? This would clarify whether the improvement is specifically attributable to lower-tail robust optimization.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.