Aggregating Raw Records Into a Company Panel

A satellite-imagery vendor does not sell you "Walmart's foot traffic." It sells you a raw table with one row per parking lot, per image, per day — thousands of individual observations that happen to belong to stores that happen to belong to one company. Turning that raw event table into a single, usable time series — one row per company per date — is the aggregation step, and the choices made here (which stores count, how missing images are handled, how outlier images are treated) shape the resulting signal at least as much as anything that happens afterward.

Choosing how to roll up

For a company $c$ on date $t$, built from raw location-level records $x_{c,i,t}$:

$$X_{c,t} = \sum_{i} w_{i} \, x_{c,i,t}$$

In words: the company-level value is a weighted sum of its individual locations' values on that date, where the weight $w_i$ decides how much each location counts — equal weight per location, weight by store square footage, or weight by that location's typical historical traffic. Changing the weighting scheme is a real research decision, not a cosmetic one, since a handful of very large flagship stores can otherwise dominate the whole company signal.

Worked example

A retailer has 4,200 stores; a satellite feed captures usable images for 3,100 of them on a given day, with coverage skewed toward larger urban locations. An unweighted average of the 3,100 available store-visit counts implicitly overweights the urban subset relative to the company's true store mix, which is 60% suburban and exurban. Reweighting by each store's share of the company's total historical square footage — rather than simply averaging whichever stores happened to have a clear satellite image that day — cuts the correlation between the resulting signal and "fraction of images that happened to be urban" from 0.41 to 0.06, a sign the reweighted series is tracking the company rather than a coverage artifact.

What this means in practice

Document the aggregation formula as carefully as any model specification, and test sensitivity to it — rerun the research with an alternative, reasonable weighting scheme and check whether the signal's performance survives. If a result only holds under one specific rollup choice out of several equally defensible ones, treat that as a warning sign rather than a footnote.