Evidence

Backtest

Would this rule have made money after costs, and how often does the forecaster's range contain what actually happened? Every number below is computed walk-forward, with no access to the future.

Set-up

Kronos foundation model · walk-forward evidence

Each row: 20-day forecasts from past origins, scored against what happened. Kronos gets weight only where it beat the random walk.

Skill = 1 − MSE(model) / MSE(random walk) on 20-day log returns; positive means it beat “no change”. A good 90% band covers about 90% of outcomes; much less means it is overconfident.

Pick an instrument and a rule, then run it. Indian listings use the .NS (NSE) or .BO (BSE) suffix.