acceptodds
Withdrawn from ICLR 2027

Scaling Prediction and Investment in Financial Machine Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Financial machine learning typically follows a predict-then-optimize pipeline: forecast returns, then construct portfolios. We study how model capacity should scale when returns are noisy and cross-sectionally dependent. We derive a finance-specific scaling law that separates diversifiable estimation error from the additional cost induced by common shocks. On a panel of 1,800 U.S. stocks with 153 characteristics, the law raises log-loss from to relative to an i.i.d. law and captures finite loss-minimizing model widths. A natural response is to select the loss-minimizing model, but our portfolio results show that this can discard useful information: smaller models achieve lower prediction loss, while larger models generate higher cumulative long–short returns. Together, these results show that financial scaling cannot be reduced to model selection alone; further scaling beyond the loss optimum requires better supervision. We therefore introduce hybrid weak-to-strong (W2S) learning, which combines smaller-model forecasts with realized returns to supervise the larger model. Under conditional orthogonality, prediction loss is quadratic in the mixing weight and admits an interior optimum. Empirically, hybrid W2S reduces prediction loss relative to direct strong learning and achieves higher terminal cumulative long–short returns than both the weak and directly trained strong models.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Finance, Financial Machine Learning, Scaling law

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

The paper proposes a financial scaling law separating stock count N from history T, and hybrid supervision mixing smaller-model predictions with realized returns. Across 1,824 configurations, the fitted law reduces scale-extrapolation log-loss RMSE by 35.3%; hybrid supervision lowers prediction loss relative to direct strong training, while paired investment-return gains are not statistically significant.

Critical strengths

Important problem with extrapolation evidence. Distinguishing stocks from dates matters for financial data allocation. The financial law improves fitted R² from 0.315 to 0.888 and reduces log-loss RMSE by 35.3% on 12 held-out scale configurations (B.2).

Supported prediction-loss gains. B.6 includes pure distillation, equal mixing, and direct-training controls. Paired log-MSE HAC t-statistics against direct strong training range from −2.25 to −3.29, supporting a predictive benefit.

Critical weaknesses

Unclear benefit beyond existing methods. Mixed teacher/label objectives are established (Cho & Hariharan, ICCV 2019); W2S scaling is studied by Ildiz et al., ICLR 2025 and Wu et al., ICML 2026. Their settings differ from financial panels, but the financial restriction's added value needs comparisons with weak/strong forecast ensembles and historically validated mixtures. Pure distillation often has lower MSE than the selected hybrid, while paired return t-statistics of 0.61–1.31 (B.6) leave investment gains inconclusive.

Uncertain identification of scaling mechanisms. Eq. (1) concerns excess risk, whereas Eq. (72) fits geometric-mean realized-return MSE without an explicit noise floor. Temporal dependence also requires justifying orthogonality between the trained predictor and future shocks. Moreover, all widths m ≥ 153 share the same function class (B.1), so single-seed, ten-epoch results do not separate capacity from optimization. These gaps limit the interpretation of fitted components.

Questions for authors

Risk mapping (clarification): What assumptions connect Eq. (1) to Eq. (72), accounting for irreducible noise, geometric aggregation, and dependence between historical estimation and future shocks?

Distinct contribution (comparison): How does the selected hybrid compare with weak/strong forecast ensembles and mixture weights chosen by ordinary historical validation, using the same information set?

Implementation (clarification): Which selector produced Figures 6, 7, 9 and Table 2: per-configuration 4/8/24-month selection or pooled 60-month selection? Please reconcile B.4.1, B.5, and B.6.

Reviewer_qE6kOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies financial prediction scaling along two dimensions of panel data, the number of assets and the number of dates. It also proposes a hybrid training objective for financial prediction. However, the two parts are only loosely connected and read more like separate studies than components of a unified contribution. Moreover, the analysis and experiments focus on a narrow model and market setting, and the theoretical results depend on several assumptions. These limitations make it difficult to assess how broadly the findings generalize and limit the paper’s overall contribution.

Critical strengths

The study of scaling laws for financial prediction is interesting and potentially novel. The paper treats the asset and date dimensions separately and develops a mathematical analysis of their effects on prediction loss. The discussion is generally well organized, and the empirical results are closely tied to the proposed theoretical framework.

Critical weaknesses
  1. The scope of the theory and experiments is limited. The paper studies a single, relatively basic model family on a fixed U.S. stock market dataset. It does not establish a general relationship between model size and predictive performance. It is therefore difficult to determine whether the findings reflect broader properties of financial prediction or are specific to the chosen model, data, and experimental setup.
  2. The empirical comparisons lack strong baselines. There are many existing approaches to stock prediction, with different model architectures and training objectives. However, the paper does not compare its methods with recent, relevant approaches, and instead focuses on variants of its own base model. This limits the evidence for the broader effectiveness of the proposed method.
  3. The two parts of the paper are insufficiently connected. The relationship between the scaling-law analysis and the proposed training objective is not developed in enough depth. As a result, the two contributions feel relatively independent rather than parts of a unified study.
  4. The investment evaluation is not sufficiently robust. The portfolio analysis does not account for realistic trading costs or other market frictions. It also omits comprehensive risk-adjusted metrics, such as the Sharpe ratio. Consequently, the reported results do not fully establish the method’s practical investment value.
Questions for authors

See my comments in weaknesses

Reviewer_qN6gOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This manuscript studies the relationship among prediction loss, model capacity and sample size. The authors design an extension to an IID benchmark, through an additional exposure term called the estimation cost, and claim that prediction loss does not necessarily improve by increasing the model size. From this view point, the authors departs to the usage of small models (W2S): controlling a hyperparameter rho that linearly interpolates between using true sample returns and predicted returns from small models. The authors show that this interpolation improves prediction loss compared to both weak and strong models. The current presentation makes the manuscript confusing and hard to follow.

Critical strengths

On the positive side, I find the intuition behind W2S interesting. Stock prices can be noisy inherently, which means that the ground-truth samples may contain non-systematic risks that are not predictable, therefore discarding them and using the model-predicted output could potentially eliminate this noise and improve later model performances (next predictions). The interpolation to strong models enhances the small model’s performance while ensuring that the generalization capability of small model is still effective.

Critical weaknesses

I have some issues with this work. The current presentation is poor. The notations are overloaded, yet their meanings are not described carefully. A notation table summarizing their meanings could substantially help readers to understand. In Proposition 1, which I believe is the core proposition of this paper, the authors should at least try to summarize assumption 8, 10, and 12, rather than simply referring readers to the appendix. These are key assumptions of the paper, they should appear in the main text. At line 195, the authors mention that G is bounded, but later at line 209, the authors contradict that G is unrestricted, which is true? I believe that a good interpretation for Proposition 1 (beyond the current one) could substantially improve the presentation. The current text is insufficient for a good interpretation on the financial scaling law. How does it different from the IID benchmark, and what role does the additional G term represent?

Questions for authors

I have a question that I would like the authors to address. Could the authors please provide a high-level interpretation on Proposition 1?

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

Related papers

Discussion (0)

Sign in to comment.