Why do modern ML models in finance often appear to underperform?

Why do modern ML models in finance often appear to underperform?

Last week, our Head of Business Development EMEA, Ivan Dexeus, and our Head of Investment Strategies, Giovanni Beliossi, attended the London Quant Group session featuring Professor Semyon Malamud. 

The discussion of his paper, "Limits to (Machine) Learning," offered a fascinating look at the statistical barriers AI faces in asset management and validated several core principles of our approach to building systematic models:

• The "final mile" is critical. Data providers are often evaluated not just on signal strength, but on the resources required to integrate that signal. Ease of ingestion is relevant.
• Complexity can be a virtue. To approach true economic signals, complex models are often necessary. This supports Axyon AI’s process of model exploration and ensembling thousands of models rather than relying on over-simplified approaches.
• Evaluation needs more than headline R² *. We need diagnostics that estimate how much predictability is being left on the table. This aligns with our commitment to "transparent box" solutions rather than black-box alpha engines.

Prof. Malamud’s work suggests that many of the perceived limits of modern ML are statistical in nature. Capturing existing signals requires the right methodological and operational framework.

In modern ML for finance, success depends on the ability to capture, evaluate, and deploy signals at scale. This is where our focus lies: building transparent, production-ready AI that transforms theoretical predictability into actionable investment insight.

* R² is a common statistical metric used to show how well a model explains past data. In simple terms, it answers the question: “How much of what happened in the past can this model explain?”

Curious how our approach bridges AI research and real-world investment use cases?