Forecasting the daily visits when I have the data for other stores

Question

I have the number of people visiting stores for each day, but sometimes one or several store do not send data for a particular day. How can I leverage the data I have for the stores that sent me data to predict the data that was not sent?

Here is what my data looks like:

day	store_a	store_b	store_c
2021-01-01	100	200	300
2021-01-02	110	220	290
2021-01-03	50	110	170
2021-01-04	110	220	290
2021-01-05	7	16	??
2021-01-06	90	??	??

These stores are affected by the same external factors: weather, strikes, lock-down measures, holidays, etc.

I could fit an estimator to predict the data for all stores, and fill the missing data with my prediction, but I think there has to be some powerful method to use the data that was sent on this particular day.

My first goal is to fill daily the missing data for the stores that did not send numbers: "As of 2021-02-19, the store B did not send data for yesterday, 2021-02-18. In light of the previous data and the data other stores sent for this day, I think store B welcomed [XX] customers."

Another goal would be to polish the data that I predicted in the past, doing something like "in the light of the data sent in January 2021, I think the store A welcomed [XX] customers on the 2021-01-15"

Answer 1

I have presented on this forum a novel way to account for (or forecast) the positivity rate over time for reported COVID-19 case data by states in America.

I noticed, in that exercise, that the reported COVID rates appeared to be connected by the 'significant' adjacent state. Significance was observed to be better defined by the population size of the adjacent neighbor together with the presence of major connecting transportation routes.

Interestingly, the underlying logic was more actually promulgated on a global countrywide level with speculated connections by air-routes. However, thinking at a local geographic level apparently may have value also.

So, my recommendation, do explore so-called nearest-neighbor paths examing both naive and perhaps more informed models. In particular, on the latter, for your missing store traffic data, look similarly at geographically route-connected stores, for example.

As missing data is eventually obtained, you may be able to further perfect the informed model.