New Hybrid Model Predicts COVID Waves Better Than Any Single AI

Forecasting disease outbreaks is like predicting the weather: you want to know what's coming so you can prepare — for hospital beds, vaccines, or lockdowns. This study compared three forecasting methods using 190 weeks of Ontario COVID-19 data from January 2020 to October 2023. The first, ARIMA, is a classic statistical method that finds patterns in past numbers. The other two, random forest and XGBoost, are machine-learning models — AI that learns from data automatically.

The researchers found that each method had its own personality. ARIMA adapted quickly when case counts suddenly spiked or dropped, but its accuracy faded when predicting more than a couple of weeks out. The AI methods were slower to react at first, but they stayed steadier over longer horizons. This means no single approach is always best — it depends on the situation.

So they built a new system called MLAMA (Machine Learning and ARIMA Model Averaging). It doesn't just pick one model; it blends all three, giving more weight to whichever has been most reliable recently. In tests, MLAMA had the lowest error across most forecast horizons and ranked among the best in every scenario. They also found that using only recent data worked better than using years of history, especially for short-term forecasts.

The big takeaway: combining different forecasting styles can beat any single one — a lesson that could improve how public health officials prepare for future outbreaks. But stay cautious: this paper hasn't been peer-reviewed, and the code isn't available yet, so it's promising, not proven.