acceptodds
Under review as a conference paper at ICLR 2027

AlphaPSC: A Pool-Subspace Conditioned Generative Framework for Complementary Formulaic Alpha Mining

Abstract

Formulaic alphas are symbolic predictive expressions widely used in quantitative trading, where newly discovered factors are typically integrated into an existing factor pool and linearly combined into a composite trading signal. Consequently, an effective factor must provide incremental predictive information beyond what the pool already captures. However, existing mining frameworks often generate candidates independently of the evolving factor library, yielding redundant and highly correlated representations. We propose AlphaPSC, a ool-ubspace onditioned generative framework for complementary formulaic alpha mining. AlphaPSC conditions generation on a continuous summary of the predictive directions already occupied by the current pool, and scores candidates by their incremental predictive contribution beyond that pool. On CSI300, CSI500, and CSI1000 under a shared expanding-window combiner with validation-selected and frozen hyperparameters, AlphaPSC improves combined-factor correlation and frictionless backtest metrics over representative formulaic miners. CSI300 ablations confirm that pool-subspace conditioning is important for these gains.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Quantitative Finance, Formulaic Alpha Mining, Generative Flow Networks, Conditional Generation, Preference-Conditioned Sampling

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_b1fNOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes AlphaPSC, a pool-conditioned generative framework for formulaic alpha mining. The central idea is to condition a GFlowNet-based formula generator on a representation of the existing alpha pool, including a sketched subspace projection, prediction residuals, and pool statistics. Candidate formulas are evaluated using a combination of standalone IC, marginal improvement in ensemble IC, and residual IC. The framework is evaluated on CSI300, CSI500, and CSI1000 against several existing formulaic alpha mining methods and conventional predictive models. The reported results show improvements in predictive correlations and frictionless portfolio metrics. The authors also provide ablation experiments to investigate the effects of pool conditioning and reward components.

Critical strengths
  1. Relevant problem formulation. The paper addresses an important limitation of formulaic alpha discovery: high standalone predictive performance does not necessarily translate into incremental predictive value when a factor is added to an existing pool. Explicitly considering the current pool during generation is a reasonable direction.

  2. Broad empirical comparisons. The paper compares against several established alpha mining approaches, including AlphaGen, AlphaQCM, and AlphaSAGE, across three equity universes. The inclusion of component ablations and repeated-seed results is also appreciated.

Critical weaknesses
  1. Portfolio evaluation. I have concerns about how portfolio returns are computed. Eq. (25) in Appendix B.1 sets the daily return to the average 20-day forward return of the selected stocks divided by 20. The authors treat this as the daily contribution of a staggered portfolio with a 20-day holding period. That isn’t the same as the actual daily mark-to-market PnL of the tranches being held. A real staggered portfolio earns, each day, the one-day price changes of positions opened on different dates. Spreading each tranche’s 20-day return evenly across the holding period smooths away that path. So the resulting series will probably understate volatility and drawdowns, and it will add serial correlation that isn’t really there. All of the annualized returns, Sharpe ratios, maximum drawdowns and cumulative wealth curves (Eqs. 26–29) come from this series, so I don’t think the portfolio results can be taken at face value yet. The backtest also ignores transaction costs and turnover. The authors acknowledge this, but together with the return definition it weakens the practical conclusions a lot. I’d ask the authors to rebuild the return series from daily tranche-level PnL and to report results net of reasonable costs.

  2. Attribution and cross-fitting. I’m not convinced the experiments isolate the effect of pool-subspace conditioning. AlphaPSC adds several things at once: pool-relative reward terms, pairwise-correlation and residual-energy gates, and a pool-dependent insertion rule. The insertion threshold of 0.3 in Appendix A.5 alone forces accepted factors to have low pairwise correlation. That means the lower correlations in Table 9 don’t show that conditioning helps find complementary signals. The reward-only ablation in Table 2 is a start. What’s still missing is a comparison with a strong pool-aware generation or selection baseline that uses the same candidate budget, insertion rules, rewards and compute. Given how far apart the ablation variants are, that control seems necessary. Separately, the cross-fitting setup may leak labels. The target is a 20-day forward return, but the embargo is only five days, so training and held-out labels near fold boundaries will overlap. Section 3.4 doesn’t explain how this is handled. Purged cross-fitting with an embargo of at least the prediction horizon would settle the question.

Questions for authors

Portfolio accounting: Can the authors clarify how Eq. (25) reproduces the realized daily returns of a portfolio containing 20 overlapping holding cohorts? Can the main portfolio results be reproduced using a conventional position-level, daily mark-to-market backtest?

Cross-fitting validity: Given the 20-session forward-return target, why is the embargo set to only five days? How are overlapping return-label intervals handled across the training and held-out folds? Would the results remain consistent under a properly purged 20-session separation?

Attribution of improvements: How much of the reported gain remains when AlphaPSC is compared against a pool-unconditioned generator using identical correlation gates, residual-energy gates, reward functions, candidate budgets, and pool insertion procedures?

Computational fairness: What are the actual formula evaluation counts, wall-clock costs, and total search budgets for each baseline? Are the reported performance differences maintained under a matched formula-evaluation budget?

3 comments

We thank the reviewer for the questions. The points raised under Critical Weaknesses are covered by the questions below, so we answer the questions directly.

Q1. Portfolio accounting: Can Eq. (25) reproduce the realized daily returns of a portfolio with 20 overlapping holding cohorts? Can the main portfolio results be reproduced with a conventional position-level daily mark-to-market backtest?

Eq. (25) does not reproduce, in a pathwise sense, the daily mark-to-market PnL of 20 overlapping cohorts. It attributes the selected top-20% basket’s hh-day forward return to the decision day and divides by hh. Main-table AR, SR, and MDD are computed on that attributed series under one frictionless protocol. They are not executable daily PnL. On the same CSI300 selection sets we ran a conventional position-level daily mark-to-market backtest. See Table 12. AlphaPSC still leads AlphaSAGE on AR (7.88% vs 6.57% under MTM; 8.65% vs 7.39% under paper RdR_d), but Sharpe falls sharply (PSC: 2.02 → 0.44). So paper SR and MDD are not tradable daily-PnL metrics. We will clarify Eq. (25) and report MTM as a supplement. We do not claim net tradable profit after costs.

Q2. Cross-fitting validity: Given the 20-session forward-return target, why is the embargo only five days? How are overlapping return-label intervals handled? Would results stay consistent under a purged 20-session separation?

We agree that a five-day embargo is short relative to a 20-session label. Cross-fitting is used only for mining rewards inside the training window. Test labels never enter generation, pool context, or admission. Without re-training, rescoring pool factors with embargo=20 leaves the acquisition scores essentially unchanged. This is not a full re-mine under embargo=20.

Q3. Attribution of improvements: How much of the reported gain remains against a pool-unconditioned generator with identical correlation gates, residual-energy gates, reward functions, candidate budgets, and insertion procedures?

The matched control is Reward-only: same correlation gates, residual-energy gates, pool-relative rewards, candidate budget, and insertion, with cPc_P zeroed for the policy. See Table 2. On CSI300, test IC falls from 0.048 (Full) to 0.013 (Reward-only). ICIR, AR, and SR fall as well (0.326 → 0.095, 8.65% → 3.21%, 2.02 → 0.69). Under the matched pipeline, most of the gap remains attributable to pool-subspace conditioning. The insertion correlation threshold of 0.3 encourages diversity, but it cannot alone explain Full ≫ Reward-only. The pairwise-correlation table is only a diagnostic, not the main evidence for conditioning.

Q4. Computational fairness: What are the formula evaluation counts, wall-clock costs, and total search budgets for each baseline? Do the gaps hold under a matched formula-evaluation budget?

See Table 8. Total search budget means the training schedule. Formula evaluations count candidates that receive a completed panel score at acquisition rounds. Pool capacity is the library size CC. AlphaPSC uses 10410^4 episodes with C=50C{=}50. Every 64 episodes it scores nacq=16n_{\mathrm{acq}}{=}16 candidates and admits at most one after the gates, so about 2,500 formula evaluations. AlphaSAGE is matched on episodes and capacity and counts one completed formula per episode, hence 10410^4 evaluations. From log timestamps, rough wall-clock on CSI300 is about 0.8–1 hour for AlphaPSC and 0.4–0.6 hour for AlphaSAGE. These are order-of-magnitude estimates, not clean single-GPU timers. The reported PSC–SAGE gaps (CSI300 IC 0.048 vs 0.040) are under that matched budget. Other formulaic miners follow their open-source default schedules and are then re-scored under the same second-stage combiner. We do not claim a matched formula-evaluation count against every baseline. Clean single-GPU wall-clock was not logged uniformly and is marked unavailable in the table.

Thank you for the detailed rebuttal and the additional experiments. I appreciate the authors' willingness to acknowledge the limitations of the original portfolio accounting and to provide more controlled evidence for the proposed conditioning mechanism. The matched Reward-only ablation substantially addresses my concern about attributing the improvements to pool conditioning. The additional mark-to-market evaluation and HAC-adjusted significance tests also strengthen the empirical evidence.

Nevertheless, two important concerns remain.

First, the original portfolio Sharpe ratio drops from 2.02 to 0.44 under conventional mark-to-market accounting. While AlphaPSC retains a performance advantage over AlphaSAGE on CSI300, this substantially weakens the economic interpretation of the original results. Moreover, the corrected portfolio evaluation remains limited to CSI300 and excludes transaction costs.

Second, the cross-fitting concern is not fully resolved. Rescoring an existing factor pool with a 20-session embargo does not establish that the generator training, factor selection, and final performance would remain unchanged under properly purged cross-fitting. A full re-mining experiment would be needed to verify this.

The revised search-budget disclosure is helpful, although matching training episodes does not establish computational equivalence across methods.

Thank you for the follow-up. We agree that the main-table Sharpe should not be read as tradable daily PnL, and that rescoring an existing pool with a 20-session embargo does not speak for a full re-mine. We add two experiments below.

  1. Portfolio accounting and costs

Under conventional mark-to-market on CSI300, AlphaPSC Sharpe falls from 2.02 to 0.44. We accept that this weakens the economic reading of the original frictionless (Rd)(R_d) numbers. We now run the same mark-to-market protocol on CSI500 and CSI1000, and subtract a one-way 10 bps cost. Turnover is half the L1 change in portfolio weights between consecutive days. On all three universes AlphaPSC still leads AlphaSAGE on annualized return, including after costs. Sharpe is much lower under mark-to-market, so we will stop presenting paper SR and MDD as tradable metrics and report the table below as a supplement.

UniverseMethod(Rd)(R_d) AR(Rd)(R_d) SRMTM ARMTM SRMTM−10bps ARMTM−10bps SR
CSI300PSC8.65%2.027.88%0.446.98%0.39
CSI300SAGE7.39%1.696.57%0.365.65%0.31
CSI500PSC11.27%2.3711.50%0.5510.59%0.50
CSI500SAGE10.16%2.1210.16%0.489.35%0.44
CSI1000PSC10.61%2.1210.66%0.499.80%0.45
CSI1000SAGE9.64%1.8110.00%0.439.12%0.39

We do not claim net tradable profit after costs. The point is only that the relative ranking under one shared protocol does not reverse.

  1. Cross-fitting and embargo=20 re-mining

Rescoring is not a purged re-train. We re-mined CSI300 pools under embargo=20, with Full AlphaPSC against the matched Reward-only control that keeps the same gates and admission rules and only zeros pool context. The second-stage combiner is unchanged. Results are as follows.

VariantICICIRARSR
AlphaPSC0.0510.3568.95%2.08
Reward-only0.0180.1253.47%0.91

Full remains clearly stronger. The numbers are even a bit above the original embargo=5 main result, but a single re-mine is noisy, so we do not claim that a stricter purge is uniformly better. The claim is narrower. After aligning the embargo with the label horizon, the gap from pool-subspace conditioning is still there. We will state the limitation of the default five-day embargo and report this re-mine.

  1. Search budget

Matching training episodes and pool capacity does not establish computational equivalence across methods. Table 8 reports training schedules and formula-evaluation counts. We did not log clean single-GPU wall-clock uniformly. We do not claim matched compute against every baseline. The AlphaPSC–AlphaSAGE comparison uses a matched budget. Other formulaic miners follow their open-source default schedules and are then re-scored under the same second-stage combiner.

Reviewer_A7wPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies formulaic alpha mining in the setting that actually matters for a library: a new expression is useful only when it adds predictive information to the pool already in hand. AlphaPSC encodes that pool as a fixed context vector, conditions a GFlowNet generator on the context and on a preference over three rewards, and admits factors with a separate insertion loop. Rewards combine standalone correlation, cross-fitted marginal gain of the pool, and cross-fitted residual correlation. Mined pools are frozen after 2020 and scored through 2026 by one shared expanding-window combiner on three Chinese index universes. The paper reports higher combined correlation and better frictionless portfolio statistics than genetic programming, reinforcement-learning miners, a GFlowNet miner, an LLM miner, and several end-to-end predictors.

Critical strengths
  1. The objective is the correct one. Standalone information coefficient keeps rediscovering correlated variants, and several strong prior miners consult the pool only through a terminal reward or a post-hoc filter. Putting a summary of the current pool into every generation step is a real change in the search problem, and it is the change this literature needs.

  2. The label discipline inside mining is better than the usual in-sample information coefficient. Marginal gain and residual correlation are cross-fitted in time with an embargo, the live pool is not updated by the gradient step, and the authors state that post-2020 labels do not enter generation, context, or admission. Scoring every formulaic method with the same frozen selection rule and the same expanding-window combiner is the right comparison.

  3. On the large-cap universe the ranking is not a one-seed accident. Across mining seeds AlphaPSC remains ahead of the closest GFlowNet baseline by a gap that is large relative to the reported seed variation, and the selected factors are less mutually correlated than the comparison pools. That is evidence about the pool, which is the object the paper claims to optimize.

Critical weaknesses
  1. The mechanism named in the title is not what the ablation isolates. Occupied directions are described as styles such as momentum or reversal, but the generator only receives a random sketch of the training score panels, with no check that the sketch preserves factor correlations. The ablation assigns more of the test correlation to the residual branch, which is a sketch of the training target, and generation without that context falls to a weak test correlation.

  2. The portfolio numbers are not evidence of economic value. The label is a twenty-session close-to-close return, so today's close is already inside the return being predicted, and the reported book is a frictionless long-only top quintile. Sharpe is annualized from the overlapping daily contributions of this staggered book, so a ratio near 2 is not a credible tradable result.

  3. Most of the reported prediction comes from a second-stage combiner that is not the one used during mining. Mining fits a ridge model, while the test numbers come from a different least-squares fit, rank-correlation gates, and a truncation to twenty factors. The paper never checks that the generator's samples match its pool-relative reward, and without code a reimplementation cannot tell whether it has reproduced the method.

  4. A twenty-day price-volume factor is a hard object, and the paper never shows one. The appendix lists the allowed fields and operators, but it does not give a single expression that actually entered the pool. At a holding horizon this long, a correlation can come from a slow risk exposure rather than from a mechanism that should persist for a month. Several concrete formulas, with an economic reading of why each one should forecast a twenty-day close-to-close return, are necessary before the pool can be treated as a set of valid factors.

  5. The only validation year is 2021, and the selection gates and combiner chosen on that year are then frozen for 2022 through mid-2026. The evaluated library is adapted to one market regime. That is a weak design for a twenty-day target, whose payoff depends on styles that can shift well inside a multi-year test. A time-series K-fold or rolling scheme would show whether the mining and selection still work when the held-out block is not 2021, and would be a more direct test of adaptation to regime change.

Questions for authors
  1. What episode count, insertion period, and baseline evaluation budget produced each reported pool?

  2. Does a low-dimensional summary of pool correlations recover the full model's test correlation?

  3. After a one-session execution lag and a stated cost, with overlapping holdings accounted for, does the ranking against the closest formulaic miner hold on more than one universe?

  4. Genetic programming builds expressions by randomly combining operators and fields, so why does its cross-correlation reach about 0.9 in the CSI300 heatmap, and how can that baseline's experimental setup be ruled out as the cause?

1 comment

We thank the reviewer for the comments. We agree that the object to optimize is pool complementarity, and that the original wording invited over-reading on mechanism names and on the economic meaning of the portfolio numbers. Details point to tables and figures in the revision. The correlation-summary check for Q2 is reported only here and is not added to the main ablation table.

W1. The mechanism named in the title is not what the ablation isolates

Momentum or reversal was only a loose analogy. It does not mean the model generates those style factors. The generator receives a random sketch of the training score panels, with no guarantee that factor correlations are preserved. The ablation isolates blocks of the continuous summary cPc_P , not a style taxonomy. See Table 2. Under matched rewards and admission rules, Reward-only zeros cPc_P and test IC falls from 0.048 to 0.013, so seeing the current pool summary is necessary for the reported gains. Whether a correlation summary can replace the full occupancy branch is answered in Q2.

W2. The portfolio numbers are not evidence of economic value

We agree. Main-table AR and SR are frictionless relative ranking, not tradable results. The label is a twenty-session close-to-close return, and the reported book is a frictionless long-only top quintile. See Table 12.

W3. Most reported prediction comes from a second-stage combiner that is not the one used during mining

Mining and reporting do use two different combiners. Mining uses a light ridge fit only to judge whether a new factor helps the current pool. Reporting scores every formulaic method, including baselines, with the same expanding-window second-stage rule so that pools can be compared fairly. See Appendix B.2. This is a design choice. Ridge is the mining-time score scale. The second stage is the shared report card. The main claim should be read as pool quality under that shared reporting rule. The original submission did not check that generated samples match the pool-relative reward. We will state this in the limitations. Code is not released with this submission per the venue practice. Method details are in the appendix.

W4. A twenty-day price-volume factor is never shown

We agree. The revision adds expressions that actually entered the pool, with a short reading of each. See Table 10. A twenty-day correlation can also come from a slow risk exposure, so the readings are descriptive rather than causal.

W5. The only validation year is 2021

We agree. Gates and combiner settings are chosen only on 2021 and then frozen through the end of the test window. This is a static out-of-sample test of a frozen pool, not a test of adaptation to regime change.

Q1.

See Table 8. AlphaPSC and AlphaSAGE are matched on episodes and pool capacity. The main-table lead still holds under that contrast. See Table 1. We do not claim a matched evaluation budget against every baseline.

Q2.

No. We ran a reply-only control that is not added to the main text. Residual and pool-stat blocks stay. Only the occupancy branch is replaced by a pairwise pool-correlation summary. Under the same window and second-stage protocol as Table 2:

VariantICICIRAR
Reward-only (Table 2)0.0130.0953.21%
Correlation summary (this reply)0.0250.1543.13%
Full AlphaPSC (Table 2)0.0480.3268.65%

The correlation summary beats Reward-only but stays well below Full. It helps. It does not recover full conditioning. The main mechanism evidence remains Full ≫ Reward-only in Table 2.

Q3.

Overlapping holdings under daily mark-to-market are in Table 12. AlphaPSC still leads AlphaSAGE on AR, but Sharpe falls from about 2 to 0.44, so the main-table Sharpe is not tradable. Next-close execution does not reverse the ranking under the frictionless protocol. A stated cost and additional universes are not ready, so we do not extrapolate.

Q4.

See Figure 3 and Table 13. GP has pool-average absolute correlation about 0.903. AlphaPSC is lowest. The heatmaps are built after the shared second-stage selection rule used for every formulaic method, not under a PSC-only metric. GP settings are in Appendix B.3 and Table 8 (population 1000, 40 generations, shared operator vocabulary). Mining mainly scores standalone fitness and favors IC. Redundancy is handled mostly after search. Under that objective, search tends to keep expressions that are mutually similar yet each has decent standalone IC, so pool correlation is high after they enter the shared combiner. This is more likely from GP favoring high-IC, high-correlation variants than from the second-stage protocol alone creating a measurement artifact.

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

The paper proposes AlphaPSC, a GFlowNet-based method for mining formulaic alphas. The generator uses a fixed-size summary of the current factor pool (cP). This summary includes a ridge-regularized projection matrix from random sketches of factor scores, a sketch of the pool's residual, and four statistics on pool size and weight concentration. The reward combines three components: standalone IC, cross-fitted marginal improvement in combined IC, and cross-fitted residual IC. Their weights are sampled from a Dirichlet distribution during training and set to uniform during factor insertion. Correlation and residual-energy thresholds filter candidates. The model is trained with Trajectory Balance on frozen pool snapshots. Each insertion cycle adds at most one factor, with a pool limit of 50.

The authors mine factors using 2010–2020 data and test them on CSI300, CSI500, and CSI1000 from January 2022 to June 2026. All methods use the same expanding-window OLS combiner, with settings fixed on 2021 data. AlphaPSC achieves the best results across all three universes in combined IC, ICIR, Rank IC, Rank ICIR, annualized return, Sharpe ratio, and maximum drawdown. The paper also reports CSI300 ablations for the context blocks and reward components.

Critical strengths
  1. The method is well motivated and includes safeguards against data leakage. Conditioning the generator on the existing factor pool helps avoid redundant candidates. The authors also take several steps to prevent leakage. Marginal gain and residual IC are cross-fitted with a five-day embargo. Reward normalization uses training data only. The pool is frozen after 2020, and the second-stage combiner uses a 21-session shift to avoid label overlap.

  2. The evaluation covers a broad range of settings. The experiments include three universes, a 4.5-year test period, a shared OLS combiner, five random seeds on CSI300, and ablations of the context components. This provides a useful basis for comparison, although the significance of the reported improvements remains unclear, as discussed below.

Critical weaknesses

The reported gains are small, and the comparisons may not be fully fair.

  • Statistical significance. The test set has 1,085 days, but the 20-day forward returns overlap. This leaves about 54 non-overlapping periods. On CSI300, IC = 0.048 and ICIR = 0.326 imply a daily IC standard deviation of about 0.15. The standard error of mean IC is roughly 0.02, larger than the 0.008 gain over AlphaSAGE. On CSI500, the IC and Rank IC gains are only 0.002. The five-seed results in Appendix C.1 measure variation across mining runs, not uncertainty over the test period. No paired tests or HAC-adjusted confidence intervals are provided.
  • Seed sensitivity. AlphaPSC's mean Sharpe ratio across five seeds on CSI300 is 1.92, below AlphaAgent's reported 1.96. The selected run in Table 1 has a Sharpe ratio of 2.02. Thus, the claim of leading on all metrics does not hold when comparing the reported seed mean with the baseline result.
  • Search budgets. The paper does not report comparable search budgets. AlphaPSC uses 10⁴ training episodes, each involving five-fold cross-fitted ridge evaluations. The reference to baseline budgets is also broken.
  • Combiner settings. The authors select combiner settings using validation IC but do not specify which method guides this selection. They also select the first 20 eligible factors in "pool order," which may affect methods differently.
  • ML baselines. The ML baselines use raw price-volume features and appear to be trained only once. Formulaic methods instead receive daily expanding-window refits of their combination weights. This makes the comparison less balanced.
  • Portfolio evaluation. The reported returns are absolute long-only returns, with no adjustment for an equal-weighted universe benchmark. The equity curves in Figures 2 and 4 move closely together, suggesting substantial common market exposure. The return label, Close[d+20]/Close[d] − 1, also assumes execution at the same closing price used to form the signal.
Questions for authors

Does the combined signal provide new alpha beyond known risk factors?

Please report the test-period IC and Rank IC for AlphaPSC and AlphaSAGE after neutralizing industry and common style exposures: size, beta, short-term reversal, momentum, volatility, and liquidity/turnover. Please also report IC after orthogonalizing against a standard factor library such as Alpha158. Low correlation within the mined pool does not necessarily imply new information beyond existing factors. Report the differences with HAC standard errors (lag ≥ 20) to account for overlapping return labels.

1 comment

We thank the reviewer for the comments. We agree that the reported gains need tighter uncertainty measures and that the comparison protocol should be stated more carefully. Our replies follow.

W1. Statistical significance and seed sensitivity

We agree that the 20-day forward returns overlap, so the 1,085 test days are not independent samples. Table 9 measures variation across mining runs, not uncertainty over the test period. For the gain over AlphaSAGE, the revision reports paired $\Delta\mathrm{IC}/\Delta\mathrm{RankIC}$ with Newey–West HAC standard errors (lag 20) and 95% CIs in Table 11. On CSI300 the CIs exclude 0. The five-seed mean Sharpe of AlphaPSC is 1.92, below AlphaAgent’s reported 1.96 in Table 1, so the claim of leading on all metrics does not hold under that mixed comparison. We soften the wording accordingly. The main-table claim is limited to the reported setting under the shared protocol. Multi-seed means are reported only for formulaic miners with a reproducible mining-seed protocol. AlphaAgent has no comparable seed-controlled pipeline in its released setup and is kept out of Table 9. Relative to AlphaSAGE, the five-seed means of IC, AR, and SR remain higher.

W2. Fairness of the comparisons

Search budgets. We fixed the broken reference and added comparable search budgets in Table 8. AlphaPSC and AlphaSAGE are matched on episodes and pool capacity. We do not claim a matched number of formula evaluations against every baseline.

Combiner settings. Appendix B.2 states that gates and capacity are chosen once by validation IC and then frozen. The first 20 eligible factors in pool order are taken under the same rule for every formulaic method. We do not retune the order for AlphaPSC alone.

ML baselines. The main comparison is among formulaic miners under the shared second-stage combiner. MLP, LightGBM, and XGBoost use raw price–volume features and are trained once as end-to-end predictors. They have no formula pool, so the daily expanding-window refit of combination weights does not apply. We keep them as black-box controls and do not claim a matched daily-refit protocol with the formulaic methods.

Portfolio evaluation. The reported returns are absolute long-only returns under a frictionless protocol. Table 12 gives the mark-to-market check. Sharpe falls sharply under MTM, so the main-table Sharpe is not read as tradable. Excess over an equal-weighted universe benchmark and same-close versus next-close execution are reported only here (CSI300, equal-weighted universe AR 3.05%):

MethodSame-close ARNext-close ARSame-close excess ARNext-close excess AR
AlphaPSC8.65%8.85%5.60%5.73%
AlphaSAGE7.39%7.53%4.34%4.40%

After subtracting the equal-weighted universe, AlphaPSC still leads AlphaSAGE. Next-close execution does not reverse the ranking. Both checks remain without trading costs.

Q1. After neutralizing industry and common style exposures, does the combined signal still provide new alpha?

See Table 11. After neutralization, IC and Rank IC stay positive for both methods, and the HAC 95% CIs for $\Delta$ exclude 0.

Q2. After orthogonalizing against Alpha158, does incremental predictability remain?

See Table 11 (Alpha158 residual). IC and Rank IC remain positive, and the CIs for $\Delta$ again exclude 0. We agree that low correlation within the mined pool does not necessarily imply new information beyond existing factors. The table shows that incremental predictability remains beyond the stated risk exposures and Alpha158.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.