A company develops an AI that is supposed to make forecasts for sales figures of various products.
The AI analyzes historical data and makes predictions for the future.
At first, the managers are satisfied because the forecasts seem realistic.
But over time it becomes apparent: the predictions tend to forecast rather average values.
Extreme highs or lows are rarely predicted, although such fluctuations occurred more frequently in the past.
The forecasts appear confident but often inaccurate when it comes to important peaks or drops.
Why does the AI always only hit the average, although it is actually supposed to deliver accurate forecasts?
Question: Why can an AI tend to avoid extreme values in forecasts and instead hit the average?
Solution follows tomorrow.
Solution
Many AI models, especially those with mean loss functions like mean squared error, optimize their predictions to hit the average of possible outcomes.
This means that with uncertain or highly fluctuating data points, they tend to avoid extreme values and instead give a forecast in the middle of the observed distribution.
This results in predictions that are good on average but rarely capture actual peaks or outliers.
The AI "protects" itself against large errors by choosing conservative values, which, however, reduces accuracy in important extreme events.
Result: An AI that bases forecasts on mean optimization can overlook important fluctuations and thus be unusable in critical situations. For more precise predictions, specialized models or methods are often needed that better represent uncertainties and extreme values.