Emergency management tracks overfitting when a wildfire-risk model memorizes last year’s incidents but fails on new weather. What training risk is illustrated?

Emergency management illustrates overfitting when a wildfire-risk model memorizes last year’s incidents but fails on new weather. Overfitting occurs when a model fits training incidents too closely—including noise and year-specific patterns—and then fails to generalize to new municipal conditions. Recognizing that risk matters because civic hazard models must support decisions under novel seasons, fuels, and weather regimes, not only replay prior years. Calling this underfitting—when a model is too simple to learn training incidents and performs poorly on past and new conditions—gets the diagnosis backward; the problem is excessive memorization, not insufficient learning. Attributing the failure to concept drift—when changing weather shifts incident patterns and reduces accuracy on new municipal conditions—mistakes overfitting for a shifting distribution; the model memorized last year’s incidents rather than the environment changing underneath it. Suggesting data leakage—when test incident details enter training and inflate measured accuracy on past wildfire data—does not fit; the issue is memorization of training incidents, not contamination from test data. Exam caveat: underfitting and data shift are related but distinct problems; a model can be simple yet still fail when the live weather distribution leaves the training support. Operational check: evaluate wildfire-risk models on held-out seasons and newer weather windows, track live performance against outcomes, and retrain or fall back to conservative rules when generalization metrics degrade.