Cloud cover gaps and what they do to a retail chain's daily vehicle count series

A retail parking lot count built from satellite imagery is an optical read. The sensor needs a clear line of sight to the asphalt, same as any camera. When a storm front sits over a region for three days, every store under it goes dark in the feed for those three days. That's the trade-off of daily cadence from VHR optical imagery: you get a fresh pass most days, but "most" is doing real work in that sentence.
For a single-store check this barely matters. You wait a day, the sky clears, you get your count. For a chain-level read ahead of a print, it matters more than most people modeling same-store sales have thought through, because the gaps aren't random across your coverage universe. They cluster by geography and by season.
Why the gaps aren't evenly spread
A chain concentrated in the Pacific Northwest loses more November and December days to cloud cover than a chain concentrated in Arizona. A chain with exposure to the Gulf Coast can lose a full week in a single storm. If you're running a weighted index across, say, 400 stores and 60 of them sit in a region that just had a wet month, the chain-level aggregate for that stretch is built on a thinner store count than the rest of your series, even though the line on the chart looks continuous.
This is the part that trips up a quick read. A dip in the daily series that lines up with a known weather event in one region is not the same thing as a chain-wide traffic drop. It's a denominator problem: on a cloudy stretch, fewer stores clear the read, and the chain-level average is leaning on a smaller sample than the days around it. If you're trying to call a comp trend two or three weeks before the call, you want to know which days in your window had full coverage and which had partial coverage, because a partial-coverage day with no demand signal can still move a naive average.
How a gap-filled series should be read
Decent practice here looks like what most time series people already do with any sensor-based data that has missing observations: flag the gap, don't fake the number. A day with no clear look at a parking lot gets treated as missing, not silently dropped and not papered over with a flat interpolation that implies a reading nobody actually took. The seasonal adjustment step then has to account for the fact that some weeks have five clean observation days and others have two, which is a different problem than the usual day-of-week and holiday adjustment in a standard retail foot-traffic model.
Practically, this means a few things an analyst building a pre-print view should check before leaning on any single day's print in the series:
- Whether the dip or spike lines up with a regional weather event rather than a demand event
- Whether the store count behind a given day's chain-level figure is close to full or thin
- Whether the trend you're reading holds up on a trailing basis once a few cloudy days roll through, rather than on a single day's reading
None of this is a reason to discount vehicle counts as a nowcast input. It's a reason to treat the series the way you'd treat any field data with gaps: trust the trend, check the denominator before you trust a single data point, and don't confuse a weather week with a traffic week.
If you want to see how a daily, chain-level feed like this handles coverage and gaps across a real set of stores, the Sales Nowcast product page lays out what the delivered series looks like, index, adjustment, and all. Worth a look before you lock your next pre-print model.