Table 2. Mean performance of the baseline dengue incidence prediction model across all cities, for the time period between January 2011 and July 2016.

The approach, a linear regression applied to the most recently available data points, is assessed across all four assumed delays in the reporting of epidemiological information, and compared with the random forest and LASSO-based approaches (with the AR+GT feature set). Numbers in bold represent the best performance for a given model and autoregressive lag across each of the metrics. This corresponds to the lowest value for the RMSE and relative RMSE metrics, and the highest value for the R^2 and Pearson correlation metrics.

Reporting Delay	Model and Feature Set	RMSE	Relative RMSE	R^2	Pearson Correlation
8 weeks	Baseline	-3.201	1.333	153.500	0.175
	Lasso, AR+GT	23.828	0.752	0.304	0.555
	Random Forest, AR+GT	23.355	0.737	0.331	0.596
6 weeks	Baseline	-1.054	1.135	127.146	0.372
	Lasso, AR+GT	22.615	0.717	0.368	0.608
	Random Forest, AR+GT	22.055	0.699	0.399	0.664
3 weeks	Baseline	-0.406	0.916	85.878	0.677
	Lasso, AR+GT	18.859	0.602	0.556	0.752
	Random Forest, AR+GT	17.613	0.562	0.612	0.793
1 week	Baseline	0.649	0.483	39.984	0.870
	Lasso, AR+GT	12.259	0.393	0.811	0.901
	Random Forest, AR+GT	11.027	0.354	0.847	0.924