Introduction The COVID-19 pandemic highlighted the importance for accurate and timely forecasts to support public health preparedness. Automated surveillance systems fit with pre-built model offers efficiency, but identifying optimal model difficult due to variations in populations. This study assessed forecasting accuracy in regions with substantial differences in population-at-risk sizes. Materials and methods COVID-19 daily incidence was forecasted for Wellington-Dufferin-Guelph Public Health (WDGPH) and Toronto Public Health (TPH), Ontario, Canada with population sizes differing by a factor of approximately 20. Datasets were split into training data (18 January 2020, to 5 November 2021) and validation data (6 November 2021, to December 3, 2021). The models applied were General Linear Autoregressive Moving Average, Seasonal Autoregressive Integrated Moving Average, and Regression with ARIMA errors, Neural Network Autoregression and Random Forest. Ensembles combining several models were then generated to investigate improvement in predictive performance. Results and discussion Random Forest provided the highest 28-day forecast accuracy with Mean Absolute Scaled Prediction Error (MASE) of 0.68, Root Mean Squared Prediction Error (RMSE) of 13.66 for TPH and MASE of 0.71 and RMSE of 4.40 for WDGPH. Statistical models, although simpler to implement, did not perform as well while ensemble modeling provided no improvement in forecast accuracy.