Macrocosm Substack

How the gradient of a model tells us which questions the data can answer

Aug 6, 2026

Maarten Scholl and the team show how a simulator's gradients reveal which questions your data can actually answer — before any time goes into collecting it or compute into fitting the model. Testing the approach on a classic financial-market model, they find a transaction tax's effect is identifiable from price data while the share of trend-following traders is not, which is exactly the kind of thing worth knowing at the start rather than the end.