acceptodds
Under review as a conference paper at ICLR 2027

From Bars to Trades: Unleashing the Potential of LLMs in Tick-Level Alpha Mining

Abstract

Alpha mining plays a central role in quantitative research by discovering predictive signals from market data. Recent LLM-based approaches have demonstrated the potential of automating alpha mining by generating and refining factor expressions over predefined market features. However, their reliance on predefined features constrains the search space and thereby limits the discovery of predictive patterns that require constructing new variables from fine-grained market events. To address this limitation, we study alpha mining directly from tick-level trade data, formulating it as a hierarchical search over variable construction and factor composition. The enlarged search space makes it difficult to identify which changes improve predictive performance. To this end, we propose CARTs, an LLM-driven framework for tick-level alpha mining. CARTs integrates a tick-level DSL that separates variable construction from factor composition, evidence-guided search with controlled comparisons and research memory for fine-grained credit assignment, and multi-view evaluation beyond predictive performance. Together, these components enable efficient and evidence-driven factor exploration. Experiments show that directly adapting conventional alpha miners to tick data can degrade performance, while CARTs outperforms the strongest tick-level baseline by 88.6% in IC at the same 100-factor library size and generalizes across markets.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Alpha mining, LLM Agents, Symbolic Regression

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_EJC6OpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper argues that LLM alpha miners, which recombine predefined bar features, cannot express hypotheses that depend on how trades are timed and sized inside a bar. It formulates tick-level mining as a joint search over event aggregation and factor composition, and proposes CARTS, a two-level domain language plus an agent that revises measurement or expression according to stored evidence and a multi-view screen. On the Chinese A-share market the system is competitive as a bar-level miner, and a one-day study reports a large correlation gain over external miners that lose accuracy when they are handed tick data. A costed long-short-constrained portfolio, a growing library, and a direct transfer of mined factors to two other equity markets are used as further support.

Critical strengths
  1. The problem is the critical one for this area. Automated miners that only recombine open, high, low, close, and volume cannot represent trade clustering, signed flow, or order lifetime, and those are exactly the measurements a tick tape is for.

  2. The two-level language is a sensible response to that task. Event filters, measurements, and reductions are separated from the later expression, so a revision can in principle touch one without rewriting the other. The work also shows, under a shared factor budget, that two existing agentic miners get worse when their inputs are switched to tick data. That negative result is informative and is reported clearly.

  3. The bar-level study is broader than a single correlation. It includes a constrained portfolio with stated buy and sell costs, a comparison against both standard libraries and learning-based predictors, and ablations of the search guidance and the multi-view screen.

Critical weaknesses
  1. The economic evidence does not attach to the tick claim. The costed portfolio, with its high annualized return, is the ten-day bar library on one Chinese test window, 2024 through 2025. That window contains a violent regime break, no second split is reported, and no uncertainty across mining runs is given. The tick experiments stop at correlation. Individual discovered tick factors sit near an information coefficient of 0.02. Nothing in the work shows that the event-level variables, rather than the combiner and the bar features, are what a portfolio would pay for.

  2. The search credit and the acceptance decision are not identified. The agent decides whether a failure came from the event definition, the expression, or the hypothesis, and it designs its own follow-up comparisons. There may need to check that those attributions are correct. Separately, an external model judges whether the program matches the stated hypothesis, and that binary judgment is a hard gate on acceptance. The ablation that removes the multi-view screen removes this semantic veto together with the redundancy filter, so the large drop in correlation cannot be assigned to either one.

  3. The generalization claim may outrun the evidence. Transferred factors produce a rank correlation near 0.02 on the US universe. On the Japanese universe a static library has higher rank correlation than CARTS. The paper does not show that trade sign, cancellations, and passive-order time exist in comparable form on those tapes, so it is unclear whether any tick-specific variable survived the transfer.

Questions for authors
  1. Under one horizon, one universe, and one label, what is CARTS's own tick-minus-bar difference in correlation and in the costed portfolio, across independent mining runs?

  2. Do the factors transferred to the other two markets still depend on event-level variables that have comparable definitions there?

  3. If acceptance keeps only the leakage and redundancy checks, what happens to test correlation and the costed portfolio?

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

The paper argues that alpha mining over daily bars cannot express hypotheses about how trades cluster inside a bar. It formulates tick-level mining as a joint search over event aggregation and factor composition, and implements that search in CARTS with a two-level DSL, evidence-guided revision, and multi-view checks. On Chinese A-shares, existing miners lose IC when their inputs are switched to tick data, while CARTS does not. The headline gain is an 88.6% IC improvement over the strongest tick-level baseline at a fixed library size of 100, with additional results on library growth and transfer to SP1000 and JPX.

Critical strengths

1.Mining alpha from trade events rather than from finished bars is a meaningful problem. Daily price and volume can be identical while the timing and size of trades differ, and that difference is invisible to OHLCV factors. CARTS makes variable construction part of the search, which is a real step beyond combining a fixed feature list.

2.The representation is concrete. Each variable is a filter, a measurement, and a reduction, and the factor is an expression over those variables. The separation lets the search revise the event definition without rewriting the whole formula, and the leakage check compares a factor with and without future trades.

3.Removing evidence-based credit assignment raises the token cost and lowers all four predictive metrics. Removing the multi-view checks lowers those metrics as well and leaves more correlated clusters in the library.

Critical weaknesses

1.Full A-share tick processing over the recent three-year window is a large and accuracy-sensitive computation, and the paper does not show that this pipeline is under control. It gives neither stage-wise runtime or complexity, nor the CPU and GPU resources used. Those information is needed to judge whether the mining procedure is reproducible at the stated scale.

2.The paper uses DeepSeek-V4 and mines 500 factors in total, 400 of which are built from transaction-level data, but the appendix shows only one factor formula. One formula is not enough to judge the quality of the whole mined library. More formulas from that library, together with statistical metrics of its factors, are needed to support the claimed quality of the factor library.

3.Figure 6 does not specify the backtest. The caption gives a TopK dropout portfolio and a ten-day horizon, and the appendix states commission, but not the execution price, the buy and sell times, limit-up and limit-down handling, or price adjustment. Combined with the missing code, the cumulative excess curve is hard to be interpreted.

4.The tick fields, the treatment of cancellations and timestamps, and the code that computes IC and the portfolio cannot be checked. An anonymous repository and a description of the dataset are necessary before the results can be verified.

Questions for authors

1.In Figure 6, at what price and time are positions opened and closed? Are buys and sells filled at the close, the next open, or some other price? Also, how are limit moves and adjustments handled?

2.How long does the tick processing take at each stage, and was it run on CPUs, GPUs, or both?

3.Does the cumulative curve in Figure 6 use the tick-factor library, the OHLCV library, or both?

4.Which limitations do the authors consider important to state in the manuscript, and which of them are left as future work?

This work addresses a challenging problem: factor mining on tick-level high-frequency data. If the Weaknesses and questions above are answered convincingly and the concerns are resolved, I am willing to raise my score.

Reviewer_SU5aOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies LLM-based alpha mining directly from tick-level transaction data. It extends the search from factor expressions over predefined features to a hierarchical search over both variable construction from trade events and factor composition. CARTs implements this formulation using a tick-level DSL, evidence-guided search with controlled comparisons and research memory, and multi-view evaluation. The paper reports gains over existing alpha-mining methods, component ablations, and transfer to SP1000 and JPX without re-mining.

Critical strengths
  1. Treating variable construction as part of alpha mining is a meaningful extension beyond expression search over fixed features. The separation between variable construction and factor composition is interpretable and permits targeted refinement of information that may be lost when trades are aggregated into OHLCV bars.

  2. The empirical results provide several encouraging signals. Table 2 supports the motivation by showing that direct tick-level adaptations can hurt AlphaAgent and RD-Agent, while CARTs remains effective. The component ablations and transfer to SP1000 and JPX without re-mining offer useful additional evidence. The chronological data split and stated future-information leakage check are also appropriate safeguards.

Critical weaknesses
  1. The main comparison does not establish comparable search effort. Matching the final library size at 100 factors does not match the number of candidate evaluations, LLM calls, validation queries, or tokens used to obtain that library. CARTs performs repeated refinement, controlled comparisons, and additional experiments when evidence is insufficient. Consequently, the reported IC improvement from 0.0431 to 0.0813 may partly reflect differences in search effort. The JEC and MVE ablations are useful, but a matched-budget joint variable/factor search control without evidence-guided credit assignment would more directly isolate the proposed mechanism.

  2. The tick-level baseline adaptations are insufficiently specified for interpreting the central comparison. It is unclear which tick fields and derived variables AlphaAgent and RD-Agent receive, whether they can construct new transaction-level aggregations, and whether they use the same LLM backbone and stopping criteria. A representation advantage could therefore contribute to the observed gains. Together with the search-budget ambiguity, this prevents a convincing attribution of the improvement to the proposed search and credit-assignment strategy. I recommend weak rejection; a reproducible description of the adaptations and a matched-budget comparison could materially change my assessment.

Questions for authors
  1. For CARTs, AlphaAgent, and RD-Agent in the 100-factor comparison, what are the numbers of candidate evaluations, LLM calls, validation evaluations, tokens, and the stopping criteria? Can the authors provide a performance-versus-search-budget comparison?

  2. How are the tick-level AlphaAgent and RD-Agent variants implemented: which raw fields and aggregations are available, can they construct new variables, and do all LLM-based methods use the same backbone?

  3. Does an existing ablation match both the joint variable/factor search space and the search budget while removing evidence-guided credit assignment? If so, please identify it; otherwise, this is the additional control most relevant to my assessment.

  4. Please confirm that the 2024–2025 test period is excluded from all factor mining, refinement, library selection, and stopping decisions.

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

This paper proposes CARTS, an LLM-driven framework for tick-level alpha mining that jointly explores variable construction and factor composition. It combines a structured DSL, evidence-guided search, and multi-view evaluation to discover predictive factors from transaction-level data.

Critical strengths
  • S1. Well-motivated problem. This paper extends automated factor discovery to tick-level data, which is a practically relevant direction with the potential to uncover richer trading patterns.

  • S2. Clear and coherent framework. The separation of variable construction and factor composition is intuitive and well-structured.

  • S3. The structured DSL makes event selection, aggregation, and factor composition explicit and easier to inspect. This is a useful design choice for financial factor research.

Critical weaknesses
  • W1. My biggest concern is, the headline improvement largely comes from comparisons with baselines that degrade on tick-level data, while CARTS itself shows only modest gains over OHLCV inputs. More controlled experiments would help clarify where the improvements come from.

  • W2. Prior works such as AlphaAgent and FactorMiner have explored several related components, including research memory, feedback-driven search, and factor evaluation. The authors could better distinguish their contributions, particularly how controlled comparisons between variable construction and factor composition offer new insights.

  • W3. The current ablations mainly remove JEC and MVE as a whole. Further separating joint search, evidence-based credit assignment, and research memory would help clarify their individual contributions.

  • W4. The existing scalability and token efficiency results are useful. It would be interesting to further examine the relationship between total search cost and factor quality, ideally under comparable budgets across different agents, to better demonstrate the practical efficiency of CARTS.

Questions for authors

Please answer and address the weaknesses listed above. Thank you.

Comments on the paper

5 comments

Thank you for the detailed rebuttal and additional experiments. I appreciate the authors' efforts to address my concerns, particularly the matched-token comparisons, clarification of baseline implementations, and fine-grained ablations.

The new evidence substantially alleviates my original concern about unequal search budgets. However, I remain somewhat unconvinced about the independent contribution of evidence-guided credit assignment. Under the matched 30M-token budget, removing credit assignment changes IC only from 0.0811 to 0.0798. While this suggests a positive contribution. I also note that the cross-market experiments, as clarified in the responses to another reviewer, use only OHLCV-dependent factors. Consequently, these experiments do not directly establish cross-market generalization of the proposed tick-level variable construction.

Thank you for your follow-up response. We further clarify the effectiveness of the credit assignment mechanism and the cross-market generalization ability of the factors mined by CARTs.

First, we conducted ablation studies under two settings: a fixed token budget and a fixed factor library size. The results show that incorporating credit assignment improves the predictive performance of the factor library while reducing the token cost of mining effective factors. These two sets of experiments jointly support the effectiveness of credit assignment in terms of both predictive performance and mining efficiency.

Second, our existing cross-market experiments show that factors based solely on OHLCV data achieve strong predictive performance in both target markets, supporting the cross-market generalization ability of the factors mined by CARTs.

To further examine the performance of factors incorporating tick-level information, we conducted additional experiments in the cryptocurrency market (BTC). We followed the same experimental configuration as FactorMiner, using 2024 data for training and 2025 data for testing, with both the bar interval and prediction horizon set to 10 minutes. The results are as follows:

MethodICICIRRankICRankICIR
Alpha1010.02100.220.02040.24
Alpha1580.02640.270.01960.24
FactorMiner0.01830.180.01630.19
CARTs0.02870.320.02810.35

CARTs outperforms all baselines across the four predictive metrics. Its combined factor set also achieves strong predictive performance in the cryptocurrency market, providing further evidence of the cross-market generalization ability of the factors mined by CARTs.

We hope that these clarifications and additional results address your remaining concerns. In light of this new evidence, we would be grateful if you could reconsider your assessment and overall score.

Dear Reviewers,

Thank you for your constructive feedback and the helpful discussion. We have uploaded a revised manuscript incorporating the additional experiments and clarifications provided in our responses. To facilitate your reassessment, we summarize the main revisions below and indicate the corresponding reviewer concerns and manuscript locations.

1. Matched-budget comparisons, search resource usage, and budget scaling.

Relevant concerns: Reviewer fchq, W4; Reviewer SU5a, W1 and Q1.

Locations: Sections 4.2.1 and 4.2.2; Appendix A.2.

We added comparisons under the same 30M-token budget and LLM backbone in both the OHLCV and tick-level mining settings. CARTs achieves the best in both settings. We also report candidate evaluations, LLM calls, validation evaluations, total token consumption, and stopping criteria, together with results under increasing token budgets. These additions complement the fixed-library-size comparisons and clarify the relationship between search resources and predictive performance.

2. Fine-grained ablations and the contribution of credit assignment.

Relevant concerns: Reviewer fchq, W3; Reviewer SU5a, W1, Q3, and the follow-up comment; Reviewer EJC6, W2 and Q3.

Locations: Section 4.3; Appendix A.2.

We expanded the ablations to separately examine credit assignment, research memory, hypothesis consistency, and redundancy checks, beyond removing JEC and MVE as a whole. Credit assignment is evaluated under both a fixed token budget and a fixed factor library size. Under the same 30M-token budget, it improves three predictive metrics, maintains Rank ICIR, and increases the number of discovered factors from 89 to 97. At a fixed library size, it improves all four predictive metrics while reducing the token cost per 10 alphas from 3.95M to 3.03M. These complementary settings provide evidence for its contribution to both predictive performance and mining efficiency.

3. Cross-market evaluation of factors incorporating tick-level information.

Relevant concerns: Reviewer SU5a, the follow-up comment; Reviewer EJC6, W3 and Q2.

Locations: Section 4.2.3; Appendix A.4.

We clarified that the existing SP1000 and JPX transfer experiments use only OHLCV-dependent factors because of data availability constraints. We also added BTC experiments using a factor set incorporating tick-level information, with both the bar interval and prediction horizon set to 10 minutes. CARTs outperforms all evaluated baselines across the four predictive metrics, providing additional evidence of the predictive value of its tick-informed factor set in another market.

4. Novelty and distinctions from prior alpha-mining systems.

Relevant concern: Reviewer fchq, W2.

Location: Section 3.1.

We strengthened the explanation of CARTs’s contribution. We clarify how controlled comparisons distinguish changes in variable construction from changes in factor composition, and how research memory retains the experimental conditions and comparative evidence needed for subsequent refinement.

5. Implementation details, reproducibility, and scope.

Relevant concerns: Reviewer SU5a, W2 and Q2; Reviewer PCuK, W1–W3, Q1, Q2, and Q4.

Locations: Section 4.1; Appendix A.1; Section 5.

We expanded the description of tick-level baseline adaptations. We also added computational resource and runtime information, factor-acceptance details, and backtesting specifications covering execution timing and prices, transaction costs, price-limit handling, and price adjustment.

6. Independent runs, temporal robustness, and portfolio performance after transaction costs.

Relevant concerns: Reviewer fchq, W1; Reviewer EJC6, W1, Q1, and Q3.

Locations: Appendices A.1 and A.3

We added results across three independent mining runs, including comparisons between CARTs’s Tick and OHLCV inputs under the same one-day prediction target. We further report portfolio performance after transaction costs, examine 100 versus 500 tick-level factors, and evaluate the portfolio impact of removing hypothesis consistency. In addition, we introduce a second chronological split, using January–September 2025 for mining, October–December 2025 for validation, and January–June 2026 for testing. These experiments address different aspects of robustness and extend the evaluation beyond predictive correlations.

We sincerely appreciate your feedback, which has helped us strengthen the experimental evidence and clarify the contributions and scope of our work. We hope that the revised manuscript and our responses address your concerns. In light of these additions, we would be grateful if you could reconsider your assessment and, where the new evidence warrants it, consider raising your score.

Thank you again for your time and thoughtful consideration.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.