Forecasting public health during a pandemic means chasing a target that changes shape: case counts can jump due to new outbreaks, variants, or reporting delays and artifacts. A Canadian study used 190 weekly observations of Ontario COVID-19 case counts from January 2020 to October 2023 to compare three approaches: ARIMA, random forest, and XGBoost. Rolling-origin cross-validation preserved temporal order, preventing the models from peeking into the future during tuning and evaluation.

Across the three dimensions examined — responsiveness after turning points, forecast horizons from one to six weeks, and the amount of historical data used for training — the models behaved differently. ARIMA adapted quickly after changes in direction, but its normalized error grew over longer horizons. Random forest and XGBoost were less responsive initially but maintained more stable error over time. For two-week forecasts at the end of the study period, training on the most recent data outperformed using longer historical periods, especially for XGBoost.

These differences led to MLAMA, an ensemble that combines models with non-negative performance-based weights that vary by horizon and responsiveness setting. In retrospective comparisons, MLAMA achieved the lowest normalized mean absolute percentage error across most forecast horizons and ranked among the best methods across responsiveness scenarios. The result is not the celebration of a single model, but the demonstration that the forecast calendar and the type of event to anticipate radically change the best choice.

This is the structural point. In healthcare, the idea of a universal 'best model' is fragile: a surveillance system must live with changing case definitions, notification delays, and sudden spikes. An ensemble like MLAMA is not just a way to reduce error, but a decision framework that allows modulating the response according to context: more responsiveness when it is necessary to intercept a turning point, more stability when the goal is medium-term planning. This distinction challenges vendor narratives that present a single 'superior' model. The real value shifts from the individual algorithm to the pipeline that combines and validates multiple approaches.

For those working with health data, reproducibility and control remain central. The study notes that the Python package implementing MLAMA is maintained in a private repository while validation and reproducibility testing are completed. This is not a detail: in public institutions, the ability to inspect code, replicate results, and update models locally is a condition for trusting a forecasting tool. Even without explicit discussion of cloud, a modular forecasting framework lends itself to being run on local infrastructure, where sovereignty over surveillance data is easier to guarantee. The operational lesson is that effective forecasting does not live in a ranking of models, but in a selection process tied to field conditions.