A data-driven tool for tracking and predicting the course of COVID-19 epidemic as it evolves

Norden E Huang; Fangli Qiao; Ka-Kit Tung

Empirically, the derivative of log N(t) or log R(t) lies on a straight line, as shown in 231 Fig. 3 (although the scatter is larger as to be expected for any differentiation of 232 empirical data). The positive and negative outliers one day before and after 12 Feb 233 are caused by the spike up and then down, with little effect on the fitted linear trend 234 (but increases its variance and therefore uncertainty). Moreover, the straight line 235 extends without appreciable change in slope beyond the peak of N(t), suggesting 236 that the distribution of the newly infected number is approximately Gaussian. The 237 mean recovery time T can be predicted as t R −t N , where tR is the peak of R(t) and tN 238 is the peak of N(t). These two peak times can be obtained by extending the straight 239 line in Fig. 3 to intersect the zero line. This predicted result can be verified 240 statistically after the fact by the lagged correlation of R(t) and N(t). If the 241 distribution is indeed Gaussian or even approximately so, the slope in Fig. 3 would 242 be proportional to the reciprocal of the square of its standard deviation, σ, as (See 243 SI):