Diversification Across Predictors: Heterogeneous Ensemble Learning for Cross-Sectional Return Prediction
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Quantitative Finance. The day in quantitative finance: models of markets, risk and prices.
Ruth: We're looking at the paper: "Diversification Across Predictors".
Warren: Here's the gist of the paper. Cross-sectional stock return prediction remains challenging due to extremely low signal-to-noise ratios, high-dimensional factor spaces, and non-stationary predictor effectiveness The proposed framework extends the financial principle of diversification to the predictor level Through a bagging procedure combining linear predictors with nonlinear predictors, each model functions as a distinct specialist whose prediction errors exhibit low cross-covariance, enabling error cancellation upon aggregation On CSI300 daily-frequency Alpha158 factor datasets, the proposed ensemble achieves a Rank IC improvement exceeding 13% over standalone neural networks and surpasses tree-based state-of-the-art methods including XGBoost and Double Ensemble as measured by both predictive accuracy and portfolio-based evaluation metrics Problem Formulation The fundamental difficulty lies in the extremely low signal-to-noise ratio (SNR) inherent to daily-frequency return data, where the predictable component constitutes only a marginal fraction of total variation Linear models lack the representational capacity to capture nonlinear interactions among factors and thus suffer from persistent underfitting Neural networks are prone to fitting spurious patterns in noise, resulting in severe overfitting and poor out-of-sample generalization The prevailing paradigm is to engineer an ever more powerful individual model rather than organizing a collective of complementary predictors How the training signal is built The framework constructs the Diversified Expert Ensemble (DEE) by drawing predictors from two architecturally distinct families, linear and nonlinear models The triple diversification protocol involves three independent randomization axes to minimize pairwise error covariance These axes are: 1.
Ruth: Bootstrap subsampling, where each predictor trains on a pool Bk of ⌊βT⌋ observations drawn with replacement 2.
Warren: Feature subsampling, where each predictor observes a random subset Fk ⊂ ΣF of size ⌊γF⌋ 3.
Ruth: Type assignment, which populates the N slots by round-robin cycling over the six types to enforce balanced representation Where the gains come from The intellectual foundation for this alternative is already embedded in finance itself, where combining assets with imperfectly correlated returns reduces aggregate risk without proportionally sacrificing expected return The diversification benefit is governed by cbar, which determines if the ensemble improves over the mean individual expert only when cbar < vbar Heterogeneity drives rho → 0, which is the precondition for the 1/N variance decay in Eq.
Title and authors: Ruth: So, we're looking at this paper today, "Diversification Across Predictors: Heterogeneous Ensemble Learning for Cross-Sectional Return Prediction." It sounds like they’re trying to tackle one of those really tough problems in finance where the signal is buried under a lot of noise.
Warren: Yeah, and the authors are Danning Li from The Hong Kong University of Science and Technology. They’re proposing this idea that you should diversify not just your assets, but the models themselves when you're trying to predict stock returns across different stocks at the same time.
Iris: It seems like they're suggesting that instead of building one super powerful model, which can get stuck in noise or miss the underlying structure, you build a collection of specialists.
Ruth: Exactly. Think about it this way: if you have one giant AI trying to predict everything, it might just memorize the random noise in the data instead of finding anything meaningful. This paper is about extending that financial idea of diversification from holding different stocks to having different types of predictive tools working together.
Warren: It’s interesting because they specifically address how neural networks tend to overfit noise while linear models struggle with nonlinear relationships, so they combine both.
Iris: And the core idea is that when you combine these specialists—the linear ones and the nonlinear ones—their prediction errors have low cross-covariance, which means they cancel each other out when you average their predictions.
The paper's summary: Ruth: So, what’s this framework actually doing? Basically, they create a Diversified Expert Ensemble that mixes linear predictors with nonlinear predictors in a bagging procedure. Each one acts as its own specialist.
Warren: The mechanism they use to make sure these specialists are different is called the triple diversification protocol. They randomize three things: how many data points each model trains on through bootstrap subsampling, which features each model sees through feature subsampling, and which type of model it is through type assignment.
Iris: That randomization process is designed specifically to minimize the pairwise error covariance between these predictors, which is the mathematical way of saying they shouldn't make too many mistakes in the same way. They even have a protection filter that removes any predictor whose performance isn't good enough on some validation set.
Ruth: That’s a lot of detail to keep track of, but the big picture is that by making sure these models are structurally different—linear versus nonlinear—and training them on slightly different data subsets, they get a better overall prediction than any single model could achieve alone.
Warren: The summary points out that this approach establishes an exact mathematical link between portfolio diversification and predictor aggregation, which is pretty neat because it formalizes the financial idea in a learning context.
The paper's improvements: Iris: When we look at what they found, they show significant improvements over standalone neural networks and current state-of-the-art tree-based methods like XGBoost. They claim a Rank Information Coefficient improvement exceeding thirteen percent over standalone neural networks.
Ruth: That’s a big jump, especially since those neural networks can easily start fitting noise when the signal is weak, which is exactly where this framework shines.
Warren: They also say it surpasses tree-based methods like XGBoost and Double Ensemble when looking at both predictive accuracy and portfolio-based evaluation metrics. It's not just about one metric; they look at how well the resulting predictions actually translate into a portfolio performance improvement.
Iris: The paper points out that the heterogeneous ensemble achieves this by combining linear and nonlinear families, and they even show that adding different regularization variants to the linear family contributes meaningful decorrelation to the ensemble.
Ruth: So, what’s their conclusion on how to actually use this? They suggest a two-pronged design: first, maximize heterogeneity through that triple diversification protocol to minimize correlation between errors. Second, you need to increase the size of your ensemble, N, because they show that the benefit from adding more predictors decays at a rate of O of one over n.
Warren: So the strategy is to make sure you have as many different types of experts as possible while watching how much improvement you get for every new expert you add.
Conclusion: Iris: To wrap up the main findings, they conclude that their diversification law shows a diminishing marginal return that decays as O of one over n. This means adding more predictors eventually stops helping you out in terms of accuracy gains.
Ruth: They also note that they found the optimal size for this ensemble seems to be around N equals nineteen, where the performance curve starts to plateau according to their analysis.
Warren: And they report that this specific ensemble achieves a very high annualized return of thirty-eight point five six percent and a Sharpe ratio of one point eight four six, while also recording the second lowest maximum drawdown among all the methods they tested.
Iris: It’s interesting that the ensemble framework actually manages to find that good trade-off between high returns and low downside risk in these noisy environments.
Ruth: Exactly. So, to sum up this paper "Diversification Across Predictors: Heterogeneous Ensemble Learning for Cross-Sectional Return Prediction," it shows that by using a heterogeneous ensemble of linear and nonlinear predictors with specific randomization techniques, you can get substantial gains over single models in low signal-to-noise ratio settings.
Warren: It gives us a concrete way to structure our predictive systems so they are more robust, even when the data is noisy or when the underlying financial factors aren't behaving linearly.
Iris: We’ll keep an eye on how this heterogeneity idea plays out in other areas of econometrics and machine learning.
Ruth: We’ve got a lot to think about for our next paper, but for today, that's what we have with this one.
Danning Li
The Hong Kong University of Science and Technology (Guangzhou)
q-fin.CP
Submitted: 2026-11-09
Updated: 2026-11-09
Comments: Accepted to ICAIF 2026
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 94/100
The gist: Cross-sectional stock return prediction remains challenging due to extremely low signal-to-noise ratios, high-dimensional factor spaces…
Key concepts
- Signal-to-Noise Ratio (SNR)
- This measures how much of the predictable pattern in stock returns is visible amidst random noise. In daily frequency data, this ratio is extremely low, making prediction very difficult. The paper argues that diversity helps overcome this by creating a stronger effective signal.
- Heterogeneous Ensemble Learning
- Instead of using many models of the same type (like only neural networks), this approach mixes different model families—specifically linear and nonlinear predictors. This structural variety forces the individual errors to be less correlated, allowing them to cancel each other out when combined in the ensemble.
- Diversification Law (O(1/N) Decay)
- This law describes how the performance gain from adding more experts decreases as you increase the number of models (N). The paper finds that gains diminish following a 1/N decay, meaning there is an optimal point for ensemble size. This suggests a two-part strategy: maximize diversity first, then increase N.
- Randomization Axes
- These are three independent methods used to create diversity among the predictors. They include bootstrap subsampling (randomly selecting data), feature subsampling (randomly selecting input variables), and type assignment (cycling through different model types). These ensure each expert is trained on a unique, varied subset of the data.
Terminology
Summary
Cross-sectional stock return prediction remains challenging due to extremely low signal-to-noise ratios, high-dimensional factor spaces, and non-stationary predictor effectiveness The proposed framework extends the financial principle of diversification to the predictor level Through a bagging procedure combining linear predictors with nonlinear predictors, each model functions as a distinct specialist whose prediction errors exhibit low cross-covariance, enabling error cancellation upon aggregation On CSI300 daily-frequency Alpha158 factor datasets, the proposed ensemble achieves a Rank IC improvement exceeding 13% over standalone neural networks and surpasses tree-based state-of-the-art methods including XGBoost and Double Ensemble as measured by both predictive accuracy and portfolio-based evaluation metrics
Problem Formulation
The fundamental difficulty lies in the extremely low signal-to-noise ratio (SNR) inherent to daily-frequency return data, where the predictable component constitutes only a marginal fraction of total variation Linear models lack the representational capacity to capture nonlinear interactions among factors and thus suffer from persistent underfitting Neural networks are prone to fitting spurious patterns in noise, resulting in severe overfitting and poor out-of-sample generalization The prevailing paradigm is to engineer an ever more powerful individual model rather than organizing a collective of complementary predictors
How the training signal is built
The framework constructs the Diversified Expert Ensemble (DEE) by drawing predictors from two architecturally distinct families, linear and nonlinear models The triple diversification protocol involves three independent randomization axes to minimize pairwise error covariance These axes are:
-
Bootstrap subsampling, where each predictor trains on a pool Bk of ⌊βT⌋ observations drawn with replacement
-
Feature subsampling, where each predictor observes a random subset Fk ⊂ ΣF of size ⌊γF⌋
-
Type assignment, which populates the N slots by round-robin cycling over the six types to enforce balanced representation
Where the gains come from
The intellectual foundation for this alternative is already embedded in finance itself, where combining assets with imperfectly correlated returns reduces aggregate risk without proportionally sacrificing expected return The diversification benefit is governed by cbar, which determines if the ensemble improves over the mean individual expert only when cbar < vbar Heterogeneity drives rho → 0, which is the precondition for the 1/N variance decay in Eq. (7) This leads to an amplified signal-to-noise ratio SNReff = N · SNR, which is significantly higher than individual expert ratios
Performance and Validation
The heterogeneous ensemble achieves substantial improvements over both conventional neural networks and current state-of-the-art tree-based methods in rank information coefficient and downstream portfolio Sharpe ratio The full six-type DEE outperforms the three-type variant by 2.2% in RankIC, confirming that the additional regularization variants within the linear family contribute meaningful decorrelation The marginal gain decreases as O (1/N), matching Eq. p. 6 The full six-type DEE outperforms the three-type variant
(linear, MLP, tree) by 2. 2% in RankIC, confirming that the additional regularization variants within the linear family contribute
meaningful decorrelation. 4 Portfolio Backtest Performance
Table 3 translates predictive signals into portfolio-level outcomes
via the top-50 equal-weight strategy. DEE achieves the highest annualized return (38. The surviving predictions are combined through equal-weighted averaging, yielding
a collective forecast whose accuracy emerges from the complementarity of its constituents rather than the strength of any single
member. Evaluated on the CSI 300 universe with the Alpha158 daily
factor set [34], DEE yields substantial improvements over both conventional neural networks and current state-of-the-art tree-based
methods in rank information coefficient and downstream portfolio
Sharpe ratio.Where the paper says it
Where the paper says it
Where the paper says it
Conclusion
The resulting diversification law predicts a diminishing marginal return that decays as O (1/N), offering a principled account of when additional capacity ceases to help The analysis prescribes a two-pronged design: maximize heterogeneity to minimize rho, and increase N to exploit the resulting 1/N decay The IC-based quality filter ensures that increasing N does not introduce harmful members that would elevate rho Together, these avenues point toward ensemble systems that are not only more accurate but more reliable in the low-signal, high-noise environments that define financial markets
Related Work
Factor-based return prediction has evolved from linear specifications to increasingly complex nonlinear architectures Gradient-boosted decision trees currently achieve the strongest empirical results on standard benchmarks yet they constitute a single function class refined through sequential residual fitting, providing no explicit mechanism for incorporating architecturally diverse inductive biases Existing financial ensembles are predominantly homogeneous, composing multiple instances of the same architecture family, typically tree-based learners Our framework explicitly constructs heterogeneity by combining linear and neural families under a unified bagging protocol The mathematical foundation for combining imperfect estimates is well established in both finance and statistical learning theory The structural isomorphism between portfolio variance reduction and ensemble error reduction has not been explicitly leveraged as a design principle for financial prediction systems Beyond the mathematical decomposition, the theory of collective intelligence provides a complementary framework While these theoretical conditions map naturally onto predictive ensembles, their application remains predominantly confined to classification tasks The present work extends this reasoning to continuous financial return prediction by demonstrating that architectural heterogeneity and stochastic training structurally satisfy the necessary conditions The resulting collective exhibits emergent predictive capabilities exceeding those of any individual member
Experiments
DEE achieves the highest IC (0.048) and RankIC (0. p. 5 DEE
achieves the highest IC (0. 048) and RankIC (0. 048) and RankIC (0. 060), surpassing XGBoost by 4. 060), surpassing XGBoost by 4. 7% in RankIC, and MLP by 13. 1% in RankIC. Notably, the gap between DEE and the tree-based
SOTA is achieved without any tree component in the nonlinear
family contributing gradient-boosted structure; the soft decision
trees in DEE are bounded-depth weak learners serving purely as
decorrelation devices. 001 at 𝑁 = 50, confirming
that larger ensembles not only improve mean performance but also
stabilize it across random seeds, consistent with the diversification
interpretation of Section 4. The curve begins to plateau around
𝑁 = 20, aligning with the saturation estimate 𝑁
∗ ≈ 19 derived
from the residual correlation analysis in Section 4. The contrast
with MLP is more pronounced: DEE improves annualized return
by 6. 1 percentage points and Sharpe by 13.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Limitations
The improvements over strong baselines remain bounded in magnitude, reflecting the fact that variance reduction cannot compensate for fundamentally weak signal Our analysis also provides only a limited view of robustness: the behavior of ensembles under regime shifts, distributional stress, and realistic frictions has not been systematically characterized The theory rests on simplified correlation assumptions that abstract away market microstructure A deeper understanding of how heterogeneity gives rise to emergent collaboration within an ensemble could inform a principled design of model composition and interaction Equally important is robustness-oriented design, including regime-aware and uncertainty-driven strategies that preserve diversification benefits under non-stationary conditions The paper is licensed under a Creative Commons Attribution 4.0 International License The author sincerely thanks Chao Zhang for his valuable guidance and insightful discussions throughout this work
Acknowledgments
The author sincerely thanks Chao Zhang for his valuable guidance and insightful discussions throughout this work The paper is licensed under a Creative Commons Attribution 4.0 International License The author sincerely thanks Chao Zhang for his valuable guidance and insightful discussions throughout this work
References
[1] J Scott Armstrong. 1989. Combining forecasts: The end of the beginning or the beginning of the end? p. 7 References
[1] J Scott Armstrong. Combining forecasts: The end of the beginning or the
beginning of the end?Where the paper says it
[2] Kenneth J Arrow, Robert Forsythe, Michael Gorham, Robert Hahn, Robin Hanson, John O Ledyard, Saul Levmore, Robert Litan, Paul Milgrom, Forrest D Nelson, et al. 2008. The promise of prediction markets. p. 7 [2] Kenneth J Arrow, Robert Forsythe, Michael Gorham, Robert Hahn, Robin Hanson,
John O Ledyard, Saul Levmore, Robert Litan, Paul Milgrom, Forrest D Nelson,
et al. The promise of prediction markets.Where the paper says it
[3] David Austen-Smith and Jeffrey S Banks. 1996. Information aggregation, rationality, and the Condorcet jury theorem. p. 7 [3] David Austen-Smith and Jeffrey S Banks. Information aggregation, rationality, and the Condorcet jury theorem.Where the paper says it
[4] Leo Breiman. 1996. Bagging predictors. p. 0 We propose a heterogeneous ensemble framework that extends the financial principle of
diversification to the predictor level. Through a bagging procedure
combining linear predictors with nonlinear predictors, each model
functions as a distinct specialist whose prediction errors exhibit
low cross-covariance, enabling error cancellation upon aggregation.Where the paper says it
[5] Leo Breiman. 2001. Random forests. p. 4 2 Ensemble Strategies in Financial Prediction
Ensemble learning combines multiple predictors to reduce generalization error. Bagging [4] and random forests [5] introduce
diversity through bootstrap subsampling and feature randomization, while stacking [33] learns a meta-level combiner over base
predictions.Where the paper says it
[6] Tianqi Chen and Carlos Guestrin. 2016. XGBoost: A Scalable Tree Boosting System. p. 7 [6] Tianqi Chen and Carlos Guestrin. XGBoost: A Scalable Tree Boosting
System. XGBoost: A Scalable Tree Boosting
System. In Proceedings of the 22nd ACM SIGKDD International Conference on
Knowledge Discovery and Data Mining (San Francisco, California, USA) (KDD
’16).Where the paper says it
Where the paper says it
[7] Kyunghyun Cho, Bart van Merriënboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2014. Learning Phrase Representations using RNN Encoder–Decoder for Statistical Machine Translation. p. 7 2939785
[7] Kyunghyun Cho, Bart van Merriënboer, Caglar Gulcehre, Dzmitry Bahdanau,
Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning Phrase
Representations using RNN Encoder–Decoder for Statistical Machine Translation. Learning Phrase
Representations using RNN Encoder–Decoder for Statistical Machine Translation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language
Processing (EMNLP).Where the paper says it
Where the paper says it
[8] Peter Christoffersen. 2010. Backtesting. p. 1 Our contributions are summarized as follows:
• We formalize the transposition of financial diversification
theory from the asset-return space to the prediction-error
space, establishing a structural correspondence between portfolio variance decomposition and ensemble forecast error
decomposition. • We propose DEE, a heterogeneous bagging framework that
combines linear and nonlinear predictors under triple stochastic randomization, achieving state-of-the-art predictive
and backtesting performance on CSI 300. Backtesting. John Wiley and Sons, Ltd. John Wiley and Sons, Ltd. 1002/
9780470061602.Where the paper says it
Where the paper says it
Where the paper says it
[9] John H Cochrane. 2011. Presidential address: Discount rates. p. 7 eqf15018
[9] John H Cochrane. Presidential address: Discount rates.Where the paper says it
[34] Xiao Yang, Weiqing Liu, Dong Zhou, Jiang Bian, and Tie-Yan Liu. 2020. Qlib: An AI-oriented Quantitative Investment Platform. p. 7 [34] Xiao Yang, Weiqing Liu, Dong Zhou, Jiang Bian, and Tie-Yan Liu. Qlib:
An AI-oriented Quantitative Investment Platform. Qlib:
An AI-oriented Quantitative Investment Platform. arXiv:2009. arXiv:2009. 11189 [q-fin.Where the paper says it
Where the paper says it
Where the paper says it
[35] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. 2021. Understanding deep learning (still) requires rethinking generalization. p. 7 11189
[35] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning (still) requires rethinking generalization. Understanding deep learning (still) requires rethinking generalization. ACM 64, 3 (Feb.Where the paper says it
Where the paper says it
[36] Chuheng Zhang, Yuanqi Li, Xi Chen, Yifei Jin, Pingzhong Tang, and Jian Li. 2020. DoubleEnsemble: A New Ensemble Method Based on Sample Reweighting and Feature Selection for Financial Data Analysis. p. 7 1145/3446776
[36] Chuheng Zhang, Yuanqi Li, Xi Chen, Yifei Jin, Pingzhong Tang, and Jian Li. DoubleEnsemble: A New Ensemble Method Based on Sample Reweighting
and Feature Selection for Financial Data Analysis. DoubleEnsemble: A New Ensemble Method Based on Sample Reweighting
and Feature Selection for Financial Data Analysis. In 2020 IEEE International
Conference on Data Mining (ICDM).Where the paper says it
Where the paper says it
[15] Stuart Geman, Elie Bienenstock, and René Doursat. 1992. Neural networks and the bias/variance dilemma. p. 7 [15] Stuart Geman, Elie Bienenstock, and René Doursat. Neural networks and
the bias/variance dilemma.Where the paper says it
[20] Takashi Kimoto, Kazuo Asakawa, Morio Yoda, and Masakazu Takeoka. 1990. Stock market prediction system with modular neural networks. p. 7 pdf
[20] Takashi Kimoto, Kazuo Asakawa, Morio Yoda, and Masakazu Takeoka. Stock market prediction system with modular neural networks. Stock market prediction system with modular neural networks. In 1990 IJCNN
international joint conference on neural networks.Where the paper says it
Where the paper says it
[21] Anders Krogh and Jesper Vedelsby. 1994. Neural Network Ensembles, Cross Validation, and Active Learning. p. 7 [21] Anders Krogh and Jesper Vedelsby. Neural Network Ensembles, Cross
Validation, and Active Learning. LightGBM: A Highly Efficient Gradient Boosting Decision Tree. In Advances in Neural Information Processing
Systems, Vol.Where the paper says it
Where the paper says it
[24] Harry Markowitz. 1952. Portfolio Selection. p. 7 [24] Harry Markowitz. Portfolio Selection.Where the paper says it
[30] Naonori Ueda and Ryohei Nakano. 1996. Generalization error of ensemble estimators. p. 7 [30] Naonori Ueda and Ryohei Nakano. Generalization error of ensemble estimators. Generalization error of ensemble estimators. In Proceedings of International Conference on Neural Networks (ICNN’96),
Vol.Where the paper says it
Where the paper says it
[31] Ivo Welch and Amit Goyal. 2008. A comprehensive look at the empirical performance of equity premium prediction. p. 7 [31] Ivo Welch and Amit Goyal. A comprehensive look at the empirical performance of equity premium prediction.Where the paper says it
[32] Justin Wolfers and Eric Zitzewitz. 2004. Prediction markets. p. 5 Extensions of the
Condorcet jury theorem establish that aggregating signals from independent and competent agents converges to the optimal decision
as group size increases [3, 25]. This principle is empirically validated
across early statistical demonstrations [1, 14], modern prediction
markets [32], and decentralized complex systems [2].Where the paper says it
[33] David H Wolpert. 1992. Stacked generalization. p. 7 [33] David H Wolpert. Stacked generalization.Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Heterogeneous Ensemble Learning for Cross-Sectional Return Prediction can be implemented to improve cross-sectional stock return prediction by combining linear and nonlinear predictors. The proposed Diversified Expert Ensemble (DEE) achieves a Rank IC improvement exceeding 13% over standalone neural networks and surpasses tree-based state-of-the-art methods.
-
Error decorrelation is enforced through three independent randomization axes: bootstrap subsampling, feature subsampling, and type assignment. This protocol ensures that the
prediction errors exhibit low cross-covariance, enabling error cancellation upon aggregation.
-
The system can operate at an amplified signal-to-noise ratio by leveraging heterogeneity in the low-SNR regime. Proposition 4.1 demonstrates that under specific conditions, the ensemble effectively operates at an SNReff = N times SNR, substantially amplifying the signal when individual experts are weak but weakly correlated.
-
Model composition can be dynamically optimized to maximize gains from architectural diversity and ensemble size scaling. The analysis prescribes a
two-pronged design: (i) maximize heterogeneity to minimize rho; (ii) increase N to exploit the resulting 1/N decay.
Abstract
Cross-sectional stock return prediction remains challenging due to extremely low signal-to-noise ratios, high-dimensional factor spaces, and non-stationary predictor effectiveness. Individual model classes exhibit complementary failure modes in this regime: neural networks overfit noise due to high variance, while linear models underfit nonlinear structure due to high bias. We propose a heterogeneous ensemble framework that extends the financial principle of diversification to the predictor level. Through a bagging procedure combining linear predictors with nonlinear predictors, each model functions as a distinct specialist whose prediction errors exhibit low cross-covariance, enabling error cancellation upon aggregation. We formalize this mechanism via bias-variance-covariance decomposition, establishing an exact mathematical correspondence between portfolio diversification and predictor aggregation. On CSI300 daily-frequency Alpha158 factor datasets, the proposed ensemble achieves a Rank IC improvement exceeding 13% over standalone neural networks and surpasses tree-based state-of-the-art methods including XGBoost and Double Ensemble, as measured by both predictive accuracy and portfolio-based evaluation metrics.
Sources
Related papers
- AFNO: Universal Certifiably-Arbitrage-Free Operator Learning Framework for European Call-Price Surfaces
- Diffusion models for dynamic volatility surface generation and data-driven hedging
- Rethinking Synthetic Scenario Realism: Compatibility, Not Fidelity, Drives Hedging Performance
- CosmosAlpha: Multi-Relational Graph Search with Pareto-Guided Exploration for LLM-Driven Alpha Mining
- DSINC beyond Heston: the 3/2 and other non-affine stochastic-volatility models
- Generative AI, Managerial Expectations, and Economic Activity