acceptodds
Under review as a conference paper at ICLR 2027

AlphaRef: Reference‑Guided Autoregressive Formula Generation for Formulaic Alpha Mining

17 views

Abstract

Discovering predictive formulaic alpha factors from noisy stock-market data is a challenging problem in quantitative finance. Existing symbolic regression methods often search over a vast expression space with limited reuse of established factor structures. Recently, some LLM-based approaches, such as Alpha-GPT, incorporate existing factor knowledge into prompts to guide formula generation. However, effectively using reference formulas to guide generation while preserving the ability to explore beyond them remains a challenge. We propose AlphaRef, a reference-guided autoregressive framework for formulaic alpha mining. AlphaRef converts formulas from the Alpha101 and Alpha158 libraries into reverse Polish notation and progressively masks their suffixes to obtain variable-length prefixes. These prefixes provide generation conditions at different levels of specificity. A Transformer-based generator combines reference prefixes with market data and factor-pool information to autoregressively generate candidate formulas. During training, a reference-decay schedule gradually reduces the use of reference prefixes, enabling a transition from reference-conditioned completion to generation without explicit references. Experiments on the CSI 300, CSI 500, and CSI 1000 stock universes in the Chinese A-share market show that AlphaRef achieves higher mean information coefficients and rank information coefficients than the baselines in the main comparison. Under the evaluated backtesting settings, AlphaRef also achieves the highest Sharpe ratios and compound annual growth rates across all three stock universes. These results support the effectiveness of reference-guided autoregressive generation for formulaic alpha mining.

Primary area:
generative models
Keywords:
alpha factor mining, symbolic regression, generative model, quantitative finance, autoregressive, reference-guided generation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_NCbPOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

AlphaRef generates symbolic alpha factors by completing prefixes sampled from Alpha101 and Alpha158. A Transformer conditions on these prefixes, market data, and the current factor zoo, and is trained with a modified GRPO objective. Reference use decays during training. The mined factors are combined through dynamic OLS, with predictive and portfolio evaluation on three Chinese stock universes and an additional U.S. backtest.

Critical strengths
  1. Reusing established formulas is a sensible starting point for this problem. They contain structures that a generator trained from scratch would otherwise have to rediscover. Completing partial formulas gives a straightforward way to retain some of that knowledge while allowing changes, and the drop in performance when references are removed supports the usefulness of this idea.

  2. The reported improvements are reasonably consistent across the three Chinese stock universes and extend to portfolio returns. I also appreciate the comparison with the original factor libraries using the same combination procedure: it provides some evidence that refining these formulas is worthwhile in practice.

Critical weaknesses
  1. [Major] It remains unclear why prefix completion is the right way to use the reference library. Removing references shows that prior knowledge helps, but a simpler approach, such as initializing an evolutionary search with the same formulas, might obtain similar benefits. The paper also discusses TLRS, which learns from expert formulas through reward shaping, without comparing against it. These are particularly relevant controls because using existing formulas is the main contribution here. The subtree-masking experiment is useful, but the need to gradually remove references is still untested. I would like to see whether the proposed approach retains an advantage when another method receives the same references and search budget, and whether the decay schedule actually helps over keeping references throughout training.

  2. [Major] I cannot reconcile the stated KL penalty with the training procedure. In Eq. 10, the log-probability difference is averaged over formulas sampled from the old policy. With those samples fixed, the reference log-probability contributes no gradient, and minimizing the remaining term can simply lower the probability of the sampled formulas. This does not provide the claimed constraint on departure from the reference policy. The issue matters because the paper uses this regularizer to explain training stability and resistance to collapse. Is the implementation using a different estimator? If the equation is accurate, the authors need to correct the penalty and check whether the reported gains still hold.

  3. The main figure is very poor.

Questions for authors

See weaknesses

Reviewer_4vaoOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes AlphaRef, a reference-guided autoregressive model for mining formulaic alpha factors. Existing alpha-mining methods either search from scratch or use reference factors only as loose prompts. AlphaRef converts Alpha101 and Alpha158 formulas into reverse Polish notation, masks suffixes to obtain prefixes of different lengths, and conditions a Transformer generator on these prefixes together with stock data and the current factor pool. A reference-decay schedule gradually replaces reference prefixes with empty prefixes, so the model can move from formula completion toward unconstrained generation. The generated formulas are filtered and added to a factor zoo, which is then dynamically combined for portfolio construction.

The experiments use daily A-share data and chronological train/validation/test splits from 2010–2025 on the CSI 300, CSI 500, and CSI 1000 universes. The reported IC, Rank IC, ICIR, Sharpe ratio, and CAGR are generally higher than those of the listed baselines and the original Alpha101/Alpha158 libraries. Ablations suggest that the reference branch is particularly important, and an appendix experiment compares the proposed sequential masking with random subtree masking. The paper also reports a cross-market experiment on US equities.

Critical strengths
  1. A sensible way to use existing formula knowledge. Representing reference formulas as progressively shortened RPN prefixes is a concrete conditioning mechanism that matches the autoregressive decoding process. The decay schedule also addresses the obvious risk that the generator simply reproduces the reference library.

  2. Broad empirical coverage with interpretable outputs. The method produces explicit formulas rather than only a return predictor, and the experiments cover three stock universes, several families of baselines, ablations, and a separate US-market check. The reported gains are fairly consistent across the main predictive metrics.

Critical weaknesses
  1. The point-in-time protocol is underspecified. The method repeatedly updates a factor zoo, selects top candidates, and dynamically fits a combination. The paper needs a precise timeline showing which observations are available at each training, validation, generation, factor-admission, and backtest step. Without this, it is difficult to rule out look-ahead through factor admission, OLS fitting, hyperparameter choices, or the historical-performance ranking used for daily selection. This is consequential because the strongest claims are about out-of-sample trading performance.

  2. The evidence for robustness is limited. Results are presented as single numbers, with no standard deviations, confidence intervals, or tests across random seeds and time splits. The backtest uses one main test interval and a relatively idealized cost model, while the reported improvements over some baselines are not accompanied by turnover, exposure, capacity, or sensitivity analyses. It is therefore unclear how stable the gains are under reasonable changes to the sampling period, transaction costs, portfolio rules, and training randomness.

Questions for authors
  1. Please provide a fully point-in-time description of the factor-zoo and dynamic-combination procedure. At the start of each trading day, which data may be used to select factors, fit OLS weights, admit new formulas, and compute the trading signal? In particular, are all thresholds and model choices fixed before the test period?
  2. How many independent training runs and random seeds were used for each reported result? Please report mean and standard deviation (or confidence intervals) for the main metrics, and clarify whether the same seed and compute budget were used for every baseline.
  3. Please report turnover and results under a range of transaction costs and execution assumptions. The current backtest cost and the handling of liquidity, delisting, suspensions, and corporate actions are not sufficient to judge practical viability.
  4. The reward uses absolute IC and maximum correlation with the existing factor zoo, while the final portfolio uses dynamic selection and OLS. How sensitive are the results to the reward coefficients, the top-n rule, the zoo capacity, and the correlation/admission thresholds?
  5. How much of the final gain comes from reference conditioning itself versus the factor-zoo representation and dynamic combination? A stronger controlled ablation would keep the generator, reward, and portfolio procedure fixed while varying only the source and amount of reference information.
Reviewer_feB5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates how existing factor formulas can help a neural generator search for predictive formulaic alphas. The method converts formulas from Alpha101 and Alpha158 into reverse Polish notation and uses prefixes of different lengths as starting points for generation. A Transformer encodes market data, the current factor pool, and the partially completed expression, then generates the remaining tokens. Candidate formulas receive rewards based on their predictive correlation and their correlation with factors already in the pool. The generator is updated through a GRPO-style objective. During training, the proportion of samples conditioned on reference prefixes gradually decreases. The retained factors are subsequently selected and combined through a dynamic linear model. The experiment results show that the complete pipeline achieves the highest reported mean IC, Rank IC, Sharpe ratio, and compound annual growth rate among the evaluated methods. These findings support the pipeline’s predictive and simulated investment performance under the reported settings. The CSI 300 ablation also provides evidence that reference prefixes are useful. Whether gradually removing those prefixes improves exploration beyond the reference formulas is not established separately.

Critical strengths
  1. The proposed method is solid. Figure 1 turns existing formulas into generation contexts of different lengths. This gives the model a direct way to reuse expression structure and provides an explicit means of varying how much of a candidate is predetermined. This is the most distinctive and promising part of the method.

  2. Generation incorporates both market observations and previously discovered factors. Figure 2 describes how the generator uses market data and the current factor pool when constructing new candidates. Conditioning on these inputs could help the search adapt to the data and avoid repeatedly finding similar expressions.

  3. The main comparison covers several equity universes and relevant baselines. Sections 4.1 and 4.2 include conventional prediction models and formula discovery methods such as AlphaGen, AlphaForge, AlphaQCM, and AlphaSAGE. In Table 2, AlphaRef achieves consistently strong Rank IC results. The consistency across these universes is encouraging.

  4. Table 3 applies the same dynamic OLS combination procedure to AlphaRef’s factors and the reference libraries, reducing one source of confounding. Section 4.4 aligns the portfolio holding period with the 20-day prediction target, uses staggered sub-portfolios, and includes a transaction cost. These points make the evaluation informative.

Critical weaknesses
  1. The proposed method overlaps with AlphaGen on sequential formula generation, QuantEvolver on learning from seed factors, and AlphaPareto on factor-pool conditioning. The clearest contribution is prefix conditioning combined with its decay schedule. The authors should distinguish that mechanism from established components.

  2. Figure 3 shows that reference prefixes help, but cannot reveal useful financial structure from easier or shorter generation. It also does not test whether reference decay improves on a fixed conditioning rate. Matched-budget comparisons with fixed-rate conditioning and length-matched random prefixes would help identify the source of improvement.

  3. Tables 2 and 3 and Figure 3 omit seed counts and variation across runs. Candidate-evaluation budgets, pool capacities, and combination settings are also unclear. These reporting gaps weaken confidence in the comparisons.

  4. Sections 3.3 and 4.1 do not specify which matured 20-day return labels are available for daily factor selection and OLS updates, or how boundary-crossing labels are handled. This affects validity.

  5. For Equation (10), which formulas sample from the old policy and the reference policy detached, the stated loss has a gradient independent of the reference probabilities. It therefore does not establish the claimed constraint toward that policy. The authors need to provide the implemented objective.

  6. Table 2 reports the Sharpe ratios and annualized returns, but not the lowest drawdowns. CSI 300 drawdown is 34.38%, versus AlphaGen’s 30.30%. The cost application and execution timing are unclear in Section 4.4.

Questions for authors
  1. What regularization objective is actually implemented?

  2. Which labels are available at each dynamic update?

  3. How many runs underlie the reported results (Tables 2 and 3 and Figure 3)?

  4. Which search and combination conditions are shared across methods?

  5. What is the exact reference-decay schedule, and what existing evidence supports its contribution?

  6. What execution and cost assumptions are used in the backtest? Is the 0.05% cost charged per side or per round trip? When are signals formed, and at what subsequent time and price are trades executed? How are suspensions and price-limit restrictions handled?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.