A retrospective assessment of short-horizon real-time COVID-19 projections in Tokyo

We examine the features of five real-time COVID-19 infection projections in Tokyo. We find that these projections tended to (i) overpredict subsequently reported infections, (ii) be less accurate when issued during the fifth infection wave, and (iii) underpredict subsequently reported infections when issued during the rising phase but overpredict them when issued during the peak and falling phases. If policymakers and the public were to utilize real-time projections in future pandemics, it would be useful for them to be aware of the possibility that real-time projections may have systematic patterns.