Black-Box Model Risk in Finance

Samuel N. Cohen, Derek Snow, Łukasz SzpruchSSRN 2021

Abstract

Machine learning models are increasingly used in a wide variety of financial settings. The difficulty of understanding the inner workings of these systems, combined with their wide applicability, has the potential to lead to significant new risks for users; these risks need to be understood and quantified. In this sub-chapter, we will focus on a well studied application of machine learning techniques, to pricing and hedging of financial options. Our aim will be to highlight the various sources of risk that the introduction of machine learning emphasises or de-emphasises, and the possible risk mitigation and management strategies that are available.

The Black-Box Dilemma: Shifting Risks in Finance

Financial institutions are undergoing a paradigm shift, moving from traditional "handcrafted" models—like Black-Scholes or Heston—to powerful but opaque machine learning (ML) systems. While these new models effectively break the "computational curse of dimensionality," allowing for state-of-the-art pricing with minimal code, they introduce a "black-box" dilemma. The core issue is not that risk is eliminated, but rather that it is shifted. When a model is opaque, it becomes difficult to understand why it makes specific predictions, making the quantification of inherent risks—such as overfitting and data sensitivity—a critical challenge.

A Taxonomy of Machine Learning in Finance

The application of ML in derivative pricing and hedging can be categorized into several distinct modes, ranging from simple numerical speed-ups to fully data-driven systems.

Figure 1: Overview of the data, models, and outputs in financial machine learning workflows, highlighting the interplay between traditional and neural network approaches.

The authors identify five primary use-cases for ML in this domain:

  1. Numerical Approximations: Using neural networks to accelerate traditional models (generally low risk).
  2. Data-driven Models: Purely empirical models that rely on historical data and disregard traditional finance theory.
  3. Hybrid Models: Architectures that embed domain knowledge, such as no-arbitrage constraints, into the neural network.
  4. Online Optimization: Utilizing Reinforcement Learning (RL) to find optimal hedging strategies in real-time.
  5. Generator Models: Employing GANs or Variational Autoencoders (VAEs) to create synthetic market data.

Data and the Benchmark Gap

Unlike fields like computer vision, finance lacks standardized public benchmarks (e.g., ImageNet) due to proprietary data and competitive secrecy. This forces a reliance on diverse data sources, each with its own risks:

  • Historical Data: Often "backward-looking" and limited to a single realized trajectory, failing to capture "what might have happened."
  • Synthetic Data: Generated via Synthetic Data Generation (SDG) or Agent-Based Models (ABM) to augment datasets, though these generators carry their own modeling risks.

Key Insights: Risk Displacement and Complexity

The transition to ML shifts risk from structural bias (where a model is too simple) to variance (where a model is too sensitive to specific data points). Interestingly, complexity does not always yield better performance; research shows that simple, fixed hedging strategies sometimes outperform complex neural networks in out-of-sample tests.

Furthermore, financial ML models are uniquely susceptible to adversarial vulnerabilities, such as "Reward Hacking" (where an RL agent finds a loophole in the reward function) and "Model Stealing" (where competitors reverse-engineer a proprietary model by querying its outputs).

Results and Risk Mitigation

The paper highlights Neural SDEs (Stochastic Differential Equations with neural network coefficients) as a robust middle ground, offering ML flexibility within a rigorous mathematical framework. Key risk mitigation strategies include:

  • Adversarial Training: Training on "worst-case" scenarios to improve robustness.
  • Explainability: Using post-hoc analysis to visualize what a model has learned.
  • Expert Knowledge as a Regularizer: Forcing models to obey financial laws, such as ensuring option prices are convex relative to the strike price.

Limitations and Future Directions

The industry faces significant hurdles, including a lack of transparency and the potential erosion of human intuition as practitioners rely more on automated systems. Future research must focus on Provable Defenses—mathematically guaranteeing a model's robustness—and better Monitoring and Control systems to detect "concept drift" when market conditions change so drastically that training data becomes irrelevant.

Summary updated

Abstract

Machine learning models are increasingly used in a wide variety of financial settings. The difficulty of understanding the inner workings of these systems, combined with their wide applicability, has the potential to lead to significant new risks for users; these risks need to be understood and quantified. In this sub-chapter, we will focus on a well studied application of machine learning techniques, to pricing and hedging of financial options. Our aim will be to highlight the various sources of risk that the introduction of machine learning emphasises or de-emphasises, and the possible risk mitigation and management strategies that are available.
BibTeX
@misc{cohen2021black,
  author = {Samuel N. Cohen and Derek Snow and Łukasz Szpruch},
  title = {{Black-Box Model Risk in Finance}},
  year = {2021},
  eprint = {2102.04757},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2102.04757},
  howpublished = {SSRN},
}
Figure