预测人工智能测验

当您参加测验时，就会出现答案和解释。

1. A model estimates tomorrow’s demand accurately. What does that alone say about what causes demand?
   • Nothing definitive; prediction accuracy alone does not establish causation
   • It proves the strongest input causes demand
   • It identifies a causal effect automatically
   • It removes the need for study design
2. Why define the prediction time and horizon before choosing inputs?
   • All horizons use identical data
   • Different horizons can require different inputs and evaluation
   • It guarantees a precise forecast
   • It removes the need for a baseline
3. A model uses a feature that becomes available only after the forecast deadline. What is the issue?
   • It is a baseline by definition
   • It improves causal identification
   • It would not actually be available when the prediction is issued
   • It makes uncertainty intervals unnecessary
4. A model forecasts future product demand. What else affects the replenishment action?
   • Whether the forecast uses a chart
   • Only the forecast’s decimal precision
   • The model’s architecture name
   • Lead time, storage capacity, shortage costs, and waste
5. A model beats a simple baseline on one period. What should a team do next?
   • Assume the baseline is obsolete everywhere
   • Check additional periods and downstream costs before relying on it
   • Deploy without monitoring
   • Report the result as causal evidence
6. Average forecast error is low, but peak-demand periods are consistently underpredicted. What does the guide warn?
   • Peak cases should be excluded by default
   • Average error proves every subgroup is accurate
   • An average can conceal systematic errors in important periods
   • The model has established why peaks occur
7. A system reports uncertainty intervals. What should evaluation check?
   • How often observations fall inside the reported intervals
   • Whether all intervals have the same width
   • Whether the point forecast is always the maximum
   • Whether uncertainty can replace monitoring
8. After deployment, outcomes arrive late and model decisions affect which outcomes are observed. What should the team do?
   • Treat every correction as user error
   • Assume missing outcomes are successful predictions
   • Stop checking after launch
   • Monitor input and outcome changes and record overrides and corrections