acceptodds
Under review as a conference paper at ICLR 2027

Conditional Factors: Discovering When to Use a Formula

Abstract

Formulaic alpha mining produces short, executable predictors, but a formula alone does not say which assets, in which information states, it should rank. We introduce conditional factors: programs that pair an applicability predicate with a ranking formula, discover the two jointly, and execute them as a pair. In U.S. equities, a shared symbolic financial state enters both the predicate and the score; the generator is rewarded by rank correlation measured only on the predicate's support, subject to support-validity constraints. An expected-rank analysis explains why the support matters: changing the comparison set can change the optimal order. We use fixed-formula controls to ask three questions. First, conditional evaluation changes the selected formula for 155 of 192 predicates in a frozen candidate library, although the ranking advantage of the new choice changes sign across years. Second, holding formulas fixed, predicates raise executed returns over always-on use in both markets and, in cryptocurrency, above same-count random support; the U.S. matched-support ranking gain, however, is positive in 2024 and negative in 2025. Third, a signed decomposition of executed returns shows two routes by which predicates help: amplifying favorable use and attenuating unfavorable use. On 2024 cryptocurrency spot data, the 20-day spot-state program earns the highest return among the reported GP, PPO, PS-Tree, and AlphaSAGE-8 comparisons; in a preregistered 2025 extension its gain over always-on use persists while its absolute return trails the external methods. Formula-level applicability is thus an executable and testable part of factor discovery whose value varies across periods.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Formulaic Alpha Mining, Computational Finance, Reinforcement Learning, Symbolic Regression, Conditional Ranking

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_jZkjOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper formulates financial factors as executable programs comprising an applicability predicate, a ranking formula, a fixed direction, and a forecast horizon. A symbolic generator searches for predicate–formula pairs using ranking quality evaluated on each predicate's support, subject to support-validity requirements.

Critical strengths
  1. A clear formulation with a relevant mathematical motivation. Proposition 1 and the three-asset example explain why evaluating a formula on its actual comparison set can matter. Keeping predicate, formula, direction, and horizon together also makes the resulting object inspectable and its execution semantics explicit.

  2. Careful controls and unusually transparent reporting. The study preserves formula directions, uses historical information availability, includes same-count support controls and cryptocurrency exposure-matched controls, and distinguishes reused periods from the preregistered extension. Reporting reversals and unsuccessful variants makes the evidence more interpretable and provides useful information about the limits of conditional factor discovery.

Critical weaknesses
  1. The incremental value of joint discovery and the shared-state construction is not established. The mathematical argument motivates evaluating rankings on their support, but does not establish a benefit from jointly learning the predicate and formula. The directly relevant structure comparison in Appendix F.2, Table 16, favors two-stage discovery under the original conditional-quality objective: +0.170% full return and +3.217 percentage points over always-on, versus −0.536% and +1.225 points for joint discovery. Appendix F.3 also reports almost identical conditional increments for separate-state and shared-state construction (+1.19 versus +1.225 points). These are exploratory controls on a variant, so they do not establish that the main implementation is inferior

  2. The independent evidence for a reliable applicability advantage remains limited. Section 4 explicitly states that 2024 U.S. outcomes entered development and that earlier cryptocurrency versions had already been evaluated on 2024 before the universe and features were expanded. In the later 2025 results, the U.S. within-support selection advantage is negative (Table 1), the U.S. matched-support ranking gain reverses, and the 20-day cryptocurrency spot-state program returns −0.07%, below the listed external methods (Table 4). Its +3.83-point improvement over its own always-on version is useful evidence, but does not by itself establish an advantage over alternative discovery methods. The uncertainty analysis in Appendix E.3 is welcome, yet most reported 2024 fixed-formula bootstrap intervals cross zero.

Questions for authors

Is there a comparison of the main joint method against two-stage predicate/formula discovery using the same grammar, source data, total candidate-evaluation budget, selection rule, and pool capacity?

Does adding the learned financial-state predicate provide a reliable benefit beyond the prespecified visibility clock under this control?

Which precise program-selection rules, predicates, and comparisons were frozen before the 2025 readout?

2 comments
AuthorsOpenReview ↗

Response to Reviewer jZkj: matched discovery and the scope of applicability evidence

Thank you for distinguishing the F.2/F.3 variants from the main method. We have added the requested comparison and clarified what each experiment answers.

Q1. Matched joint versus two-stage comparison. Yes. The main methods now share the grammar, source data, evaluation budget, selection rule and pool capacity:

Measure, averaged across annual origins and seedsJointTwo-stage
Distinct scoring expressions, matched capacity53.3347.32
Effective signal dimension, quality/coverage-matched subset12.5412.47

Joint discovery explores a broader set of expressions; effective signal dimension is similar overall. Our public response to Reviewer adtv, Q3, explains why this supports retaining joint discovery while acknowledging the two-stage return advantage. Preserving exploration breadth while improving returns is the next objective (Section 5.4; Appendix H.3).

F.3 addresses a different choice: sharing the state between condition and score. Its similar returns make sharing a compact, inspectable representation, rather than an established source of economic advantage. The theory motivates support-aware evaluation and allows both joint and sequential discovery.

Q2. Does the state add reliable value beyond the clock? We added a use-time clock-only control that fixes formulas, directions and portfolio rules and removes only the state filter. Separately, industry/size/historical-risk/count-matched supports test local state information. H20 RankIC increments are:

Frozen population20242025
Main clock + state+0.00961-0.00487
No-hard-clock variant+0.01961-0.00507

The evidence identifies a local advantage in 2024 that does not persist in 2025. The no-hard-clock variant's 2024 gain also survives a larger minimum support. Portfolio increments beyond the clock depend on the funding/risk control. Finding a local relationship, establishing transfer and improving a portfolio are therefore distinct questions; our framework makes each testable, while stable superiority remains future work (Section 5.2; Appendix H.4).

Q3. What was frozen before 2025? Training ends in 2022; selection uses fixed rules on 2023. The original registry fixes:

ComponentFrozen before the extension
ProgramsFormula actions, predicates, directions and horizons
Selection and executionSelection rules, universe and trading settings
ComparisonsAlways-on, matched random supports, AlphaSAGE and paired ranking
Primary setting and successH20/5 bp; positive use increment in at least two seeds and the combined account; superiority to same-coverage random support

All selected-pool hashes match the registry; no program was reselected using 2025 outcomes. Appendix H.9 and the code provide the precise rules. Later baseline additions, accounting correction and annual refits remain separately identified.

Thank you for the clarifications. However, my main concerns remain. First, joint discovery shows greater expression diversity but no demonstrated performance advantage over two-stage discovery, which achieves better returns. Second, the learned state provides no robust out-of-sample benefit, with RankIC gains reversing in 2025 and clock-only outperforming clock plus state. These issues undermine the paper's central empirical claims, so I maintain my original rating.

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

The paper treats a factor as a pair: a Boolean predicate that decides which assets are ranked, and a formula that ranks them. The two are searched together and then frozen, including the sign estimated on the source period. An expected-rank argument says that the best ordering can change when the comparison set changes, so a formula should be scored on its own support. Experiments separate three questions. Support-aware selection changes the chosen formula for 155 of 192 predicates, but the new choice is not better in every year. Holding the formula fixed, predicates improve executed returns relative to always-on use. In U.S. equities the absolute return stays negative, and the ranking gain versus same-count random support reverses in 2025. On 2024 cryptocurrency data the 20-day spot-state program has the highest reported return. In the preregistered 2025 extension its advantage over always-on use remains, while its absolute return falls behind the external methods.

Critical strengths

1.The three empirical questions are kept separate. Formula selection, fixed-formula applicability, and complete-program performance are not collapsed into one backtest. Direction is held fixed, so a predicate comparison is not confounded with flipping the sign of the formula.

2.The controls match the claim being tested. Always-on use asks what the predicate changes. Same-count random support asks whether the selected names beat an equally small random set. In cryptocurrency, an exposure-matched control also equalizes deployed capital. These are the right comparisons for an applicability claim.

  1. Each retained object is a frozen predicate, score, source-period direction, and horizon. The program used later is the program scored during search, and the direction is not re-estimated on the test window. U.S. accounting inputs are rebuilt from Compustat records visible at the decision time, so later revisions do not backfill history. A full representative program is shown, not only a schematic.
Critical weaknesses

1.The predicate is chosen inside a menu that is fixed before search. In U.S. equities the clock has four visibility stages and the state test has four templates. The generator learns expressions inside those 16 cells. The predicate is the cell’s clock and threshold conjoined with a symbolic state that also enters the score. Cryptocurrency likewise chooses among fixed price–volume parents and an optional spot-state menu. A learned-clock variant concentrates 386 of 400 programs on one stage. The search therefore does not discover applicability as an open condition. It fills in formulas under author-specified regimes.

2.The U.S. return increment is not a profitable applicability effect. Clock-plus-state loses 0.39% and 2.51% at the two horizons in 2024–2025. The quoted gain of about one percentage point is relative to always-on formulas that lose more. The clock alone has a slightly better mean return than clock-plus-state, so the added financial state does not improve the portfolio result in this comparison. The paper notes that U.S. predicates also change the capital deployed. The ranking comparison with same-count random support is positive in 2024 and negative in 2025. Three seeds share one market path, and about half of the U.S. program-months have no defined within-support RankIC.

3.Support-aware selection changes the formula without reliably improving it. For 155 of 192 predicates the within-support rule and the full-universe rule pick different formulas. The within-support choice ranks better in 2024 and worse in 2023 and 2025, with at most two of three seeds agreeing. Portfolios built from either selection have negative three-seed mean returns in the evaluation window. The result shows that the comparison set can change discovery. It does not show that scoring on the support selects a better formula.

4.The technical increment is the object, not a new discovery principle. Attaching an applicability set to each formula is a useful change, but the search itself is standard and the reported effect is small and period-dependent.

Questions for authors

1.After matching capital or gross exposure, how much of the U.S. return increment over always-on use remains? The cryptocurrency experiments already report an exposure-matched control.

2.What was locked as the primary 2025 comparison before that year was read: the increment over always-on use, or the absolute return against external methods? 3.For the paired RankIC results, what is the distribution of support size, and how many program-months are dropped because the within-support correlation is undefined? Does the sign reversal in 2025 remain on months where both supports are large enough for a stable rank correlation?

4.Can a predicate fall outside the 16 U.S. cells or the cryptocurrency parent menu, or is every reported condition an instance of those templates?

2 comments
AuthorsOpenReview ↗

Response to Reviewer pPQt: what is searched and what the evidence establishes

Thank you for asking us to clarify what is learned and what the evidence establishes. We have revised the argument and added the requested controls.

W3 / W4. Contribution and interpretation. We agree that changing the selected formula does not establish better selection. Its significance is that applicability changes the learning problem: the assets being compared help determine which ranking is appropriate. Our contribution makes the scoring relationship and its intended scope a joint, executable hypothesis that can be discovered and tested. Reliable selection improvement and transfer are further objectives.

The new constraint experiment shows why factor discovery and portfolio returns need separate evaluation. With the same frozen formulas, U.S. long-only cumulative returns in 2024–2025 at 10 bp are:

Allocation ruleConditionalAlways-on: shared weightsAlways-on: own fit
Equal weights, frozen60.26%62.84%62.84%
2023 fit, frozen18.18%46.83%23.35%
Rolling, nonnegative50.07%65.68%19.03%
Rolling, signed70.67%77.24%11.99%

Shared weights isolate conditional use; separate fits compare complete systems. The conclusion changes with that choice even though the formulas do not. We therefore distinguish discovery, use and allocation, and separately identify the non-executable future-direction diagnostic (Table 6; Appendix H.10).

Q4 / W1. Templates versus search. Every reported predicate instantiates the stated templates. What is learned within them determines which assets qualify and how they are ranked:

PrespecifiedLearned by search
Disclosure stages and state-test formsFinancial expression defining the state
Inputs and operator grammarPrice–volume formula and state–score composition

Stage stratification preserves exploration across contexts; freely learning the clock concentrated search in one stage. We updated Appendix G and added 50 representative frozen U.S. factors, with full expressions and actions, to the repository and anonymous supplement so this distinction can be inspected directly.

Q1 / W2. Capital and risk. The positive point increment over always-on remains after exposure or achieved-risk matching; see Reviewer adtv, Q1, for the comparison table and interpretation. This is a relative selective-use effect. The separate state-versus-clock question is addressed in Reviewer jZkj, Q2.

Q3. Support sizes. Appendix H.5 retains all 2,304 possible H20 program-months per year:

Diagnostic20242025
Mature support: median [IQR]31 [12, 232.5]29 [12, 241]
Undefined native RankIC1,1111,157
Paired minimum 30: valid months / increment1,187 / +0.015411,136 / -0.00334
Paired minimum 100: valid months / increment848 / +0.00398850 / -0.01887

Undefined correlations mainly reflect insufficient valid labels. Both native and random supports must meet the minimum, and the 2025 reversal remains after excluding small supports. Seed variation and uncertainty on the shared market path are reported separately.

Q2. Prespecified target. Applicability was a primary target: the original protocol registered the increment over always-on and matched random supports, alongside AlphaSAGE and paired ranking. Our public response to Reviewer jZkj, Q3, gives the freeze and success criteria (Appendix H.9).

Thank you for your response. I checked it out, and It addressed my main concerns. I think the content and code materials are enough to support their claims, and I will update my score, have a good luck.

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

The paper proposes a framework for jointly discovering symbolic alpha formulas and predicates that determine when and on which assets each formula should be applied. The authors argue that evaluating a formula over teh entire asset universe may overlook predictive relationships that hold only for specific subsets of assets. The approach uses policy-based symbolic search and evaluates candidate formulas based on their ranking quality within the selected support. Experiments are conducted on U.S. equities and cryptocurrencies, with comparisons against AlphaSAGE, GP, PPO, and PS-Tree. The authors find that conditional evaluation changes formula selection and can improve returns relative to applying the same formulas unconditionally. However, the results are mixed across markets and evaluation periods.

Critical strengths

Interesting approach to conditional alpha discovery: The paper addresses a relevant problem in alpha mining: a formula may work well for certain assets or market conditions but poorly elsewhere. Explicitly learning when a formula should be applied is a sensible extension of existing symbolic alpha-discovery methods. The theoretical discussion on how optimal rankings depend on the comparison set is also useful.

Good effort to isolate the contribution of conditional predicates: The fixed-formula experiments are useful, as they allow the authors to examine the effect of changing applicability conditions while keeping the underlying formulas unchanged. The use of point-in-time data, transaction costs, and frozen trading directions is also positive. I appreciate that the authors report negative results and acknowledge limitations in their evaluation.

Critical weaknesses

Lack of volatility targeting and risk-normalised performance comparisons: My main concern is with how financial performance is evaluated. The paper compares cumulative returns and maximum drawdowns across strategies without reporting their annualized volatility or targeting a common level of risk. This is particularly problematic because the proposed method changes when factors are active and how much capital is deployed. A strategy that takes less risk will generally exhibit smaller gains, losses, and drawdowns. Without normalising volatility, it is difficult to establish whether the reported improvements reflect better predictive signals or simply differences in risk exposure.. The authors do report Sharpe ratios, which are more informative in this respect as being volatility-invariant, but the main comparisons still rely heavily on raw returns and drawdowns. I consider this a fundamental weakness in the experimental design, rather than just a missing robustness check.

Limited out-of-sample evaluation and potential selection bias: The paper does not adopt a systematic walk-forward validation framework, despite searching a large space of candidate formulas and applicability conditions. Instead, the evaluation relies on a limited number of fixed test periods. The authors also acknowledge that some evaluation periods had already been examined during development. This raises concerns about potential selection bias, particularly given the number of candidate formulas, configurations, and experimental choices involved. The cryptocurrency results illustrate the issue. The 20-day spot-state strategy achieves a 24.37% return in 2024 but approximately zero in 2025, underperforming several comparison methods. This raises questions about whether the discovered relationships are robust across different market regimes or are specific to the chosen evaluation periods. Given the nature of the problem, I would have expected rolling or expanding-window walk-forward validation, with multiple out-of-sample periods, to be a central part of the experimental design.

Unclear incremental benefit of joint discovery: The main methodological contribution is the joint discovery of applicability predicates and ranking formulas. However, I am not convinced that the experiments establish a clear advantage over simpler approaches where formulas and applicability conditions are learned sequentially. In Appendix F.2, the two-stage approach actually outperforms joint discovery under the original conditional-quality objective, both in cumulative returns (+0.170% versus −0.536%) and conditional increments (+3.217 versus +1.225 percentage points). Although this is an exploratory comparison, it raises questions about the benefit of the proposed joint optimization. Similarly, Appendix F.3 reports very similar results for shared-state and separate-state constructions, suggesting that coupling the predicate and scoring representations may not be particularly important. The theoretical argument establishes that the optimal ranking can depend on the comparison set, but this does not necessarily imply that predicates and formulas need to be learned jointly. I would have liked to see a more convincing matched-budget comparison against a sequential approach, particularly given that joint discovery is presented as the main methodological contribution.

Questions for authors

Volatility Targeting: Could the authors report annualized realized volatility for all strategies and provide performance comparisons under a common ex ante volatility target? Given that conditional activation changes portfolio exposure, I believe this is necessary to interpret the return and drawdown comparisons. Out-of-sample performance: Why was a walk-forward validation framework not adopted? Given the size of the symbolic search space and the acknowledged examination of some evaluation periods during development, how can the authors establish that the results are not sensitive to the particular train/test split? Incremental benefits of joint optimization: Could the authors clarify why the two-stage approach outperforms joint discovery in Appendix F.2? What evidence supports the incremental benefit of joint optimization over learning formulas and predicates sequentially? Performance Analysis: What explains the deterioration in cryptocurrency performance between 2024 and 2025? Is it primarily driven by changes in predictive quality, market conditions, or portfolio exposure?

1 comment
AuthorsOpenReview ↗

Response to Reviewer adtv: risk controls, temporal evaluation and discovery

Thank you for identifying the missing risk and temporal comparisons. We agree these are necessary to interpret the results, and have added them to the revision.

Q1. Risk comparability. We agree that raw returns need a common-risk comparison. We added realized volatility and a common 10% ex ante volatility target for all seven U.S. accounts below (2024–2025, 5 bp, equal-funded H5/H20):

U.S. accountOriginal annualized vol.Targeted returnTargeted realized vol.
Clock + state1.708%-4.794%4.256%
Same-score always-on2.008%-8.479%5.554%
State only4.527%-11.680%8.324%
AlphaSAGE adapter4.124%-11.888%7.414%
GP adapter4.348%-11.917%7.117%
PPO adapter4.240%-11.515%7.439%
PS-Tree4.172%-5.153%4.575%

The target is not fully reached under the exposure limits, so we also use a lower common target that both conditional and always-on accounts can attain. At a 1% target, their realized volatilities are 1.004%/0.981%, and the conditional return increment is +0.745 percentage points. Exposure matching gives the same qualitative conclusion: the relative point benefit is not explained solely by investing less. Appendix H.1 gives the full controls and uncertainty, including fixed-formula and random-support comparisons for cryptocurrency.

Q2. Temporal evaluation. The original design kept programs fixed to test whether their applicability transferred. Your concern identifies a different question that it did not adequately address: how results change when discovery is repeated at a different date. We have now added annual discovery and validation refits across 2022–2025, using consistent rules and only labels available at each origin.

YearJoint RankICTwo-stage RankICJoint minus same-score always-on
2022+0.02460+0.03178+0.00956
2023+0.02802+0.02168+0.01399
2024+0.02745+0.02645+0.01046
2025-0.01542-0.00686-0.00135

Across three seeds per origin, the comparison shows temporal variation in both methods. Earlier applicability gains weaken in 2025, and refitting does not uniformly restore performance. Repeating discovery makes this sensitivity visible beyond the original split. These retrospective results are distinguished from the frozen extension (Table 5; Appendix H.2).

Q3. Why retain joint discovery? Thank you for your interest in the two-stage variant. We do not claim that joint discovery must outperform it. In F.2, two-stage discovery benefits more from applying its conditions, despite having weaker always-on performance; this explains its higher full return.

We retain joint discovery because we want the context in which a relationship works to guide the search for that relationship. Evaluating formulas within their conditions during generation gives that information a role in what is explored. The new matched comparison supports this choice: joint discovery retains broader expression diversity, while two-stage search concentrates more on repeated scoring expressions. How to preserve that breadth while improving returns is an important next step (Section 5.4; Appendix H.3).

Q4. Why did cryptocurrency performance decline? The main change is a weakening of the frozen formulas' ability to separate subsequent returns. Almost all of the examined spot-state programs deteriorate, while costs fall and support availability remains broadly similar. Reading the formulas and holdings reveals persistent preferences for low nominal prices and low absolute price variation; the payoff relationships associated with those preferences weaken or reverse in 2025.

This also explains why absolute performance and conditional-use value can move differently. At matched exposure, the conditions still improve on using the same formulas everywhere and on random supports, even though the underlying signals have deteriorated. We interpret this as a selective-use benefit during signal decay. Appendix H.8 gives the decomposition and separately records the uniform event-accounting correction, alongside the original reported returns.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of revisions and responses

We thank the reviewers for identifying gaps in the risk controls, temporal evaluation and explanation of what is learned. We have added the missing experiments and revised the paper around a clearer research question.

A formula says how to rank assets, but leaves open where that relationship should apply. Our expected-rank argument shows why this matters: changing the assets being compared can change the ranking target itself. We therefore study applicability as part of the factor hypothesis, with an explicit condition and a formula that can be inspected and evaluated together.

This separates three questions: which relationship is discovered, whether restricting its use helps, and how it performs in a portfolio. A relationship can weaken while selective use still reduces losses relative to using it everywhere. Establishing such an effect is different from establishing superior absolute returns. This distinction motivates the paper's formulation and evaluation framework; reliable transfer and stronger economic performance are the next objectives.

The revision adds annual discovery refits, capital and risk controls, an explanation of the cryptocurrency decline, and a controlled experiment that progressively relaxes portfolio constraints. These experiments address the reviewers' concerns and show why factor discovery, conditional use and portfolio adaptation need separate evaluation. We also updated the representative formulas and accompanying code materials.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.