Model risk is the possibility that a model's output is wrong, misleading, or unsuitable for the decision it informs. In trading it applies to quantitative models and to AI assistants, and it includes fabrication, overfitting, data errors, and misapplied outputs.
How it works
A model output carries risk whenever it feeds a decision: the number can be wrong, the sample can be unrepresentative, or the model can be used outside what it was built for.
Managing model risk means verifying outputs, understanding what data built the model, and keeping the final decision with a human who owns the risk.
Why it matters
An AI assistant that fabricates a number and a backtest that overfits are the same failure class: a model output that misleads the decision.
Keeping judgment on the human side is the main defense, because no model owns the consequences of its output.
A simple market example
A language model produces a confident revenue figure for a company. The figure is fabricated. Using it in an analysis inherits the error, which is model risk in its simplest form.
Common mistakes
Treating a model's confident output as verified truth.
Using a model for a task it was never built for, such as a language model predicting prices.
Frequently asked questions
Is model risk only about AI?
No. It applies to any model that feeds a decision, including statistical and backtested models.
How do I reduce model risk?
Verify outputs against data, understand the sample that built the model, and keep the final decision with a human.
Can model risk be eliminated?
Not entirely. The defense is verification and human ownership of the decision and its consequences.
Educational content only. Definitions describe common market usage and may vary by jurisdiction, instrument, or institution.