Estimation Review for 4/30 Estimates

The estimates made on 4/30 are now complete so it’s time to review how they did. Error is calculated using “Balanced relative Error (BRE)”.

Technique #1 did OK. Average error on the daily estimates was 33%, median was 19%, and 75% of the estimates were within 25% of actual (Pred(25)). On a cumulative basis, the average daily error was 10% and median was 9%. 100% of the estimates were within 25% of actual and 75% were within 10%.

Technique #2 did better. Average error on the daily estimates was 39%, median was 17%, and 58% of the estimates were within 25% of actual. This looks worse, but the cumulative numbers are the numbers that really matter. On a cumulative basis, the average daily error was 7%, as was the median. 100% of the estimates were within 25% of actual, 75% were within 10%, and 42% were within 5%.

Technique #1 was based on forcing the estimates onto a 12-day lag. Technique #2 started with the same cumulative target as Technique #1, but used positive test % (“positivity”) as a means of calculating the day-to-day variations.

New Estimates as of 5/13/20

This set of estimates uses the 12-day lag and the method of approximating CFR from the positive test %. The estimates are shown in the image below.