How Firm Is a Forecast? Reading a Forecasting Model at Several of Its Own Settings

A new kind of AI model can forecast almost any series of numbers it is shown, such as power demand, sales or prices, without being trained on that series. Such a model gives one forecast. It does not tell the user how firm that forecast is. This note describes a way to find out. The series is kept exactly as it is. The model's own internal settings are moved a little, one at a time, and the forecast is read at each setting. If the forecasts are close together, the forecast is firm. If they are far apart, the answer changes whenever a setting is changed, so it cannot be trusted. The model and its settings must be looked at again before anyone uses that forecast. A computer-decided shunting action follows the verdict. The computer decides without waiting for a person, and it moves the work aside to a safer track instead of stopping it. The note lists the papers read on forecasting models, calibration, input attacks, uncertainty in AI models and Monte Carlo dropout, and explains why each differs from this method. The method is patent pending. Keywords time series foundation models, forecasting, TimesFM, forecast reliability, internal settings, spread of forecasts, shunting action, AI safety, uncertainty