Spare parts and slow-moving stock produce series that are mostly zeros. Standard forecasters quietly fail on them; Croston’s method and its successors are built for exactly this shape of data.
Missing data is not only a preprocessing nuisance. The reason data is missing can change model bias, fairness, monitoring, and deployment behavior.
Rare event models should be optimized for decisions, not only class balance. Cost-sensitive learning connects model thresholds to real operational consequences.
Decision curve analysis evaluates predictive models by asking whether acting on their predictions produces better decisions than simple alternatives.
Fourier analysis is more than a signal-processing trick. It is a way to ask which cycles, rhythms, and scales explain variation in data.
Counterfactual evaluation helps teams estimate how a new decision policy might perform before deploying it to users, patients, customers, or operations.
Sensor anomaly detection works best when statistical signals, domain constraints, and alert workflows are designed together.
Prevalence shift occurs when the base rate of the outcome changes, breaking thresholds, workloads, and probability interpretation even when the model ranking still looks good.
Uplift modeling estimates treatment effect heterogeneity so interventions can target the people, assets, or cases most likely to benefit.
Choosing the best model is the obvious strategy. Averaging several is usually better, and the reason is not that the average is smarter but that it is less wrong in a specific way.
