AI prediktif

A prediction is conditional on the data and model assumptions. It is neither a guarantee nor evidence that the model has identified a causal relationship.

Key takeaways

• Specify the horizon and available inputs.
• Connect prediction quality with the action it supports.
• Evaluate uncertainty and performance over time.

Define the prediction time and horizon. A forecast for tomorrow, next month, and the next five minutes can require different inputs and evaluation. Check that every input would actually be available when the forecast is issued. Separate prediction from the action taken on it. An inventory forecast estimates demand; a replenishment decision also depends on lead time, storage capacity, shortage costs, and waste. A better numerical score is useful only when it improves the downstream decision. Evaluate against simple baselines and across time periods. Average error can conceal systematic underprediction during peak demand or poor performance on new products. Where appropriate, estimate uncertainty and check how often observations fall inside the reported intervals. Monitor both input changes and measured outcomes after deployment. Feedback may arrive late, and the model’s own decisions can change which outcomes become visible. Record overrides and corrections so a later review can distinguish model errors from missing measurements or policy changes.

1. For a hypothetical three-day period, actual demand is 10, 20, and 30 units. Forecast A predicts 12, 18, and 28.
2. Absolute errors are 2, 2, and 2, giving mean absolute error of 2 units. A constant forecast of 20 has errors 10, 0, and 10, averaging about 6.67 units.
3. Check additional periods and shortage costs before deciding that the first forecast is operationally better.

The invented figures illustrate an error calculation, not evidence about a deployed forecasting system.