Thank you @Aaron_Flaaen_Federal_Reserve_Board for presenting his research on “Explorations of High-Frequency Measures of Business Exit during COVID-19”. You are welcome to ask any additional questions here. The following questions were not answered due to limited time:
• Is balancing the type1 and type 2 errors a sort of rough fitting of the data to the model?id Dr. Flaaen mention using business closure data from BLS?
• To validate your 0.65 threshold, you mentioned predicting the “closed_on” variable. Are you referring to the closed_on variable we have in some releases of the Core places? Not sure I understood how you used this dataset that Safegraph used to train its model… Can you explain again please?
• Would all the exits in March perhaps be vulnerable restaurants that are struggling to begin with? Seems to be roughly two populations then? ‘Struggling/vulnerable’ and ‘robust’?
Regarding the “best” combination of parameters for defining temporary closures, I’m treating the predicted values from the operating/not operating SG dataset as truth. Then iterating over lots of parameters and (manually) finding the best combination. I’m not (yet, at least) minimizing some loss function.
The benchmark I’m using is the BLS Business Employment Dynamics dataset, which repots quarterly establishment deaths for some industries. (BDM Home : U.S. Bureau of Labor Statistics)
I validate the 0.65 threshold using the operating/not operating predictions from the SG special data product, as described above. I use the “closed_on” variables (yes, those that are attached to the CORE files) to help think through what visit patterns look like preceding and subsequent to an identified permanent closure.
Though note that using that variable would be another way to formally validate a threshold.
It is certainly possible that those retaurants that close (permanently) in March were struggling beforehand. In some exploratory work for fast food restaurants I was struck that the exit rates were closer to sit-down restaurants than I would have expected. (Given the ease with which those establishments can conduct carry-out, etc…) On a careful look I verified many of the largest chains involved in these closures with formal announcements, and it does appear that they were tied to longer-running challenges than just the pandemic. So that at least anecdotally supports that view…
cc @Mohsen_Bahrami_MIT who’s doing similar work and might have some feedback