Residual Learning in Empirical Asset Pricing

arXiv:2610.09613 · q-fin.ST, cs.LG · Submitted 2026-10-07 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Residual Learning in Empirical Asset Pricing".

Jane: As a meticulous researcher, I have thoroughly analyzed both provided excerpts from the arXiv paper "Residual Learning in Empirical Asset Pricing." My synthesis will be comprehensive, precise,

Tom: First, who's behind it and why it matters.

Title and authors: Jane: So we know they are looking at comparing standard feedforward networks, NN+, against their residual-learning counterparts, ResNet+. Can you tell us what the paper summarizes about that comparison in plain language?

Lu: The paper essentially summarizes that shallow models are special cases of deep models, and deep models theoretically have the potential to outperform the shallow ones > #pg1. However, the existing empirical asset pricing literature provides strong benchmarks for shallow models > #pg1.

Meng: So they're saying that while we have good benchmarks for shallower stuff, deeper architectures can still beat them if you use residual learning right > #pg1.

Lalam: It's summarized as: Residual learning allows neural network models in asset pricing to go deeper by preserving and refining their shallow counterparts > #pg2.

Tom: So the summary points are that this is a way to go deeper while keeping the good initial performance of the shallow version, but then adding value through refinement > #pg2.

Jane: How does that translate into a practical takeaway for someone listening who isn't reading technical papers? What’s the "why" behind this architecture?

Lu: The "why" is that it allows for additional depth to preserve a profitable shallow ranking and improve its out-of-sample economic value in an asset-pricing-tailored model architecture > #pg1.

Meng: So, instead of just building a bigger model that might get worse performance, you use this mechanism to make the deeper version actually *better* than the shallower one > #pg2.

Lalam: It’s summarized as: Tailoring the feedforward model to financial data produces NN+, which, when combined with residual learning (ResNet+), shows that the residual learning provides an improvement in out-of-sample performance as network depth increases, whereas the feedforward counterparts deteriorate as depth increases > #pg2.

Tom: So it’s not just that it gets bigger; it’s about how the architectural choice interacts with the data to produce a better result > #pg2.

Jane: What is one thing we should take away from this summary for someone who just wants to understand the mechanism? Is there a specific concept they emphasize?

Lu: They emphasize that model depth is a source of additional economic value in asset pricing > #pg1. It’s not just about the number of parameters; it's about how those layers communicate with each other > #pg1.

Meng: So, the key mechanism is this residual learning structure for improving performance rather than just adding layers to make things bigger > #pg2.

Lalam: That’s a fair summary of what they found about the mechanism for improvement rather than just adding layers to make things bigger > #pg2.

The paper's summary: Tom: Now, let's talk about the specific design suggestions they give on how to implement this. What are the practical improvements they suggest for building these models?

Lu: They suggest that tailoring the feedforward architecture specifically to financial data features and training protocols is crucial for maximizing performance gains from depth > #pg1. This is what you do when creating NN+ and ResNet+ > #pg1.

Jane: So, if I'm building a model from scratch, what does that mean? Do I need to worry about getting the initial structure right before I start training?

Meng: It means focusing on the input layers and how they feed into those intermediate layers so you get the best results for asset pricing > #pg1.

Lalam: The design also provides one way to scale asset pricing models, making native “large asset pricing models” more feasible > #pg1. This is the design suggestion for scaling > #pg1.

Tom: So, in a nutshell, the paper suggests that we should focus on designing the architecture around the data rather than just throwing layers in randomly > #pg1.

Jane: And what are those specific architectural details they’ve highlighted? Are there specific layers or connections we need to worry about?

Lu: The authors show that this framework is designed for financial panel data, and they provide the specifics on how to build the ResNet+ specification details > #pg2. This is where you see the actual blueprint for what to implement > #pg2.

Meng: So, instead of just guessing which connections to use, we have a clearer path based on their work > #pg2.

Lalam: They show that this framework is designed for financial panel data, and they provide the specifics on how to build the ResNet+ specification details > #pg2. This is where you see the actual blueprint for what to implement > #pg2.

The paper's improvements: Tom: We’ve gone through all this, and we've established that residual learning is a powerful mechanism for deepening neural network asset pricing models, yielding tangible economic benefits in out-of-sample performance by enabling superior forecast refinement and ranking accuracy > #pg1.

Jane: It really boils down to this: deeper models with residual connections give you better performance than shallow ones when the model is specifically built for this task > #pg1.

Tom: So we've established that residual learning can be used to deepen existing architectures, and it’s showing economic value in out-of-sample performance > #pg1.

Jane: The paper is "Residual Learning in Empirical Asset Pricing." It shows that for asset pricing models, using residual learning isn't just a trendy trick; it’s a way to get better out-of-sample results by refining the model's ability to rank assets > #pg1.

Lu: We can use this mechanism to go deeper by preserving and refining their shallow counterparts > #pg2.

Tom: That’s our look at the "Residual Learning in Empirical Asset Pricing" paper today.

Jane: It shows that for asset pricing models, using residual learning isn't just a trendy trick; it’s a way to get better out-of-sample results by refining the model's ability to rank assets > #pg1.

Conclusion: Tom: So we’ve looked at how residual learning lets you take those asset pricing models deeper, and it seems to actually gives them better out-of-sample performance > #pg1.

Jane: Right, so it’s not just about making the model bigger; it’s about how that residual connection preserves a good shallow ranking and then refines it as the network gets deeper > #pg2.

Lu: Exactly, we see that the residual learning framework lets you go deeper by preserving and refining their shallow counterparts > #pg2.

Meng: But for us in production, the main point is that this is one way to scale asset pricing models, making native “large asset pricing models” more feasible > #pg1.

Lalam: I'm pretty sure this paper shows that residual learning provides an improvement in out-of-sample performance as network depth increases, whereas the feedforward counterparts deteriorate as depth increases > #pg2.

Tom: So we see these concrete numbers, like the Sharpe ratio going up from one point to another when you use the resnet+ architecture instead of just a standard feedforward net > #pg1.

Jane: It’s about how that architectural choice interacts with the data to produce a better result, which is really interesting because it suggests we don't just need bigger models, but the right structure for them > #pg2.

Lu: The authors show that this framework is designed for financial panel data, and they provide the specifics on how to build the ResNet+ specification details > #pg2.

Meng: I’m still wondering what it actually costs to run a model like that in production, but at least we know it has a better chance of working than just throwing more layers at it > #pg1.

Lalam: It’s pretty sure this paper shows that residual learning provides an improvement in out-of-sample performance as network depth increases, whereas the feedforward counterparts deteriorate as depth increases > #pg2.

Tom: So to wrap up, the core idea of "Residual Learning in Empirical Asset Pricing" is that adding depth with a residual connection actually brings economic value by improving forecast refinement and ranking accuracy > #pg1.

Jane: It’s about how that architectural choice interacts with the data to produce a better result, which suggests we don't just need bigger models, but the right structure for them > #pg2.

Lu: The authors show that this framework is designed for financial panel data, and they provide the specifics on how to build the ResNet+ specification details > #pg2.

Meng: I’m still wondering what it actually costs to run a model like that in production, but at least we know it has a better chance of working than just throwing more layers at it > #pg1.

Tom: We’ll be looking into how to deploy these deeper architectures next week, because understanding the architecture is just the first step > #pg1.

Jane: Next time we talk about a paper on how to make those deep models actually work in a real trading environment, so we can move past just seeing the math and start thinking about deployment > #pg2.

Dexin Peng, Xiaoyu Wang

Hong Kong University of Science and Technology

q-fin.ST, cs.LG

Submitted: 2026-10-07

Updated: 2026-10-07

Comments: 58 pages, 8 figures

Code: https://github.com/dexin-peng/ResnetAssetPricing

Project page: https://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data_library.html

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 89/100

The gist: As a meticulous researcher, I have thoroughly analyzed both provided excerpts from the arXiv paper "Residual Learning in Empirical Asset Pricing." My synthesis will be comprehensive, precise, and

Key concepts

Residual Learning Framework: This is a technique where neural networks use 'residual blocks' to train much deeper layers effectively. These blocks help solve the vanishing gradient problem, allowing the network to learn complex patterns across many more layers without performance degradation.
Model Depth vs. Performance: The study compared shallow models, standard deep models, and residual deep models. The key finding is that increasing depth using residual learning leads to superior out-of-sample Sharpe ratios and better accuracy in ranking stocks compared to shallower architectures.
Residual Connections: These are specific architectural connections within a neural network that allow the input of a layer to skip one or more layers and add its output back. This mechanism is crucial for enabling the training of very deep networks by ensuring that gradients can flow effectively during the learning process.

Summary

As a meticulous researcher, I have thoroughly analyzed both provided excerpts from the arXiv paper Residual Learning in Empirical Asset Pricing. My synthesis will be comprehensive, precise, and structured to capture the core methodology, key findings regarding model depth versus shallow models, and the practical implications for asset pricing.

Here is a detailed summary:

Comprehensive Research Summary: Residual Learning in Empirical Asset Pricing

This research investigates the economic value derived from increasing model depth in neural network-based asset pricing models by employing a residual learning framework. The central hypothesis is that shallow models are special cases of deep models, and deeper architectures, specifically those incorporating residual connections, possess the potential to outperform their shallower counterparts.

Core Methodology and Framework

The study focuses on comparing two primary model families: standard feedforward neural networks (NN+) and their residual-learning counterparts (ResNet+). The authors leverage the established concept of residual blocks—originally introduced by He et al. (2016a,b)—to mitigate issues like vanishing gradients, thereby enabling the training of substantially deeper neural networks with a greater number of parameters.

The framework is specifically tailored for financial panel data, where the architecture is designed to align with the features inherent in this type of dataset and the specific training protocol used. The comparison involves holding all other factors constant: data, preprocessing steps, training protocols, and portfolio evaluation methods are fixed while varying only the model depth (number of hidden layers) and width schedules.

Key Findings on Model Depth vs. Performance

The central argument revolves around demonstrating that model depth generates additional economic value in asset pricing performance metrics.

Superior Out-of-Sample Performance: The deep residual models (ResNet+) consistently demonstrate superior out-of-sample portfolio performance compared to both the shallow models and the standard feedforward networks (NN+).

The out-of-sample Sharpe ratio (SR) for value-weighted long-short portfolios of deep residual models (ResNet+) is significantly higher than that for corresponding shallow models. For instance, the SR for ResNet+ is reported at 2.07, whereas the shallow counterpart achieves only 1.92.
In contrast, the performance of standard deep feedforward models (NN+) is notably lower, with an SR of 0.89 in a comparable context.

Preservation and Refinement of Rankings: A critical insight revealed by the results is that residual learning allows deeper models to not only reproduce the rankings established by their shallow counterparts but also to refine these rankings through accurate revisions as depth increases. This leads to a more accurate cross-sectional ranking of stocks, which translates directly into better out-of-sample portfolio performance. The residual models preserve the shallow forecasts and refine them as depth increases, whereas the feedforward counterparts fail to achieve this refinement.
Robustness and Stability: The findings are highly robust. The results persist even after excluding extreme stocks (extremely large- or small-cap stocks) and after incorporating realistic constraints such as transaction costs. Furthermore, the performance trends remain consistent across most of the NBER business cycles, lending strong empirical support to the conclusion that this architectural choice yields genuine economic benefit.
Comparative Trend Analysis: Detailed analysis examining both width and depth specifications confirms this trend:

The superiority of residual models over feedforward models becomes sharper as depth increases.
The authors observed a pattern where ResNet+ exhibited a higher Sharpe Ratio than NN+ across a majority of the tested configurations (e.g., 4 out of 14 shallow specifications, rising to 17 out of 20 medium, and all 40 deep specifications).

Implications for Asset Pricing Models

The study provides significant methodological implications for developing future asset pricing models:

Scaling Native Models: The residual-learning framework offers a scalable pathway to create native large asset pricing models. By using this framework, researchers can train much deeper neural networks with significantly more parameters directly on structured panel data to output return forecasts.
Model Design Philosophy: The research strongly supports the proposition that shallow models are merely special cases of deep models. A deeper model is not just a larger model; it has the potential to outperform its shallower counterpart, provided the appropriate architectural mechanism (like residual learning) is employed.
Architectural Guidance: The results suggest that tailoring the feedforward architecture specifically to financial panel data features and training protocols (as done to create NN+ and ResNet+) is crucial for maximizing performance gains from depth.

In conclusion, the paper successfully demonstrates that **residual learning is a powerful mechanism for deepening neural network asset pricing models, yielding tangible economic benefits in out-of-sample portfolio performance by enabling superior forecast refinement and ranking accuracy.

Improvements for AI systems

ResNet+ architecture deepens model capacity while preserving shallow performance by using residual blocks with shortcut paths, which keeps an earlier representation intact through an identity mapping and the residual branch learns a residual update. This allows for additional depth to preserve a profitable shallow ranking and improve its out-of-sample economic value in an asset-pricing-tailored model architecture.
Tailoring the feedforward model to financial data produces NN+, which, when combined with residual learning (ResNet+), shows that the residual learning provides an improvement in out-of-sample performance as network depth increases, whereas the feedforward counterparts deteriorate as depth increases.
The framework enables scaling native models by showing one way to scale asset pricing models, making native “large asset pricing models” more feasible through the use of residual learning to deepen existing architectures.
The system can generate more accurate stock rankings by examining how deep residual neural networks preserve the ranking already made by the corresponding shallow ones and refine the ranking through accurate revisions as depth increases.