Model Uncertainty

Model uncertainty — also called Knightian uncertainty or deep uncertainty — is the condition in which an agent does not merely face unknown parameter values within a known model, but does not know which model is correct. It is distinct from risk, where probabilities are known, and from parameter uncertainty, where the model structure is given but its coefficients must be estimated. Model uncertainty is uncertainty about the model itself: the set of relevant variables, the functional form, the causal structure, and the boundary conditions.

In macroeconomics, model uncertainty was emphasized by John Maynard Keynes and later formalized by Gilboa and Schmeidler through maxmin expected utility: when the true model is unknown, the rational agent maximizes utility under the worst-case plausible model. In climate science, model uncertainty dominates policy debates: different climate models produce divergent projections not because of parameter differences but because of structural differences in how clouds, aerosols, and ocean circulation are represented. In artificial intelligence, model uncertainty is the problem of out-of-distribution generalization: a model trained on one data-generating process may fail catastrophically when deployed on another, not because of noise but because the underlying causal structure has changed.

The critical insight is that model uncertainty cannot be reduced by more data within the wrong model. Collecting more observations refines parameter estimates but does not test structural assumptions. The only response to model uncertainty is structural skepticism: maintaining multiple models, actively seeking disconfirming evidence, and designing policies that are robust across model specifications rather than optimal within a single one. This is the logic behind robust decision making and scenario planning — and it is the logic that rational expectations economics, with its assumption of a single known model, systematically suppresses.

The assumption that the model is known is the most dangerous assumption in applied science. It transforms statistical precision into intellectual complacency, and it produces policy recommendations that are optimal in imaginary worlds and disastrous in real ones.