Daily card-spend data covers 8% of transactions for 60 listed retailers and arrives with a 3-day lag; results come out 45 days after quarter end. How long is the information window, and how precise can the estimate be?

The clean window is about 42 days, from 3 days after quarter end to the results, and the data is precise enough to beat consensus only for the larger retailers. Sampling error falls with the square root of transactions captured, from about 3.8 points for a small chain to 0.24 for a large one, but a panel bias of about 1.8 points never goes away. Against consensus errors of 3 points, the edge is real for big names and absent for small ones.

Draw the calendar. Spend builds up through the 91 days of the quarter and arrives three days late, so the full quarter's data is in hand three days after quarter end, while the official number arrives 45 days after it: a 42-day window in which you know roughly what the results will say. Inside the quarter there is a partial window too; two thirds of the quarter is visible by about day 64. But the window is not all yours. Company trading updates, other data vendors and channel checks leak the same information, so the value is highest early in the window and decays towards the results date.

Think of an exit poll. Interviewing 2,000 voters gives a similar error whether the constituency has one lakh voters or ten lakh; what matters is the count of people asked, not the share. For card spend, the sampling error of a retailer's total is the variability of a single transaction divided by the square root of the number of transactions in the panel. With a transaction value spread of about 1.5 times the average, a large chain with 50 lakh card transactions a quarter puts 4 lakh in the panel and gets a sampling error of about 0.24 points; a small chain with 20,000 puts 1,600 in and gets about 3.8 points.

Then add the error sampling cannot fix. The panel's cardholders are not the retailer's whole customer base: cash buyers, other card networks and a different online share all sit outside it. That panel biasThe gap between what a data panel measures and the full population it stands for, caused by who is in the panel rather than by how many are in it. More data from the same panel does not shrink it. of about 1.8 points sets a floor, so even the largest chain's total error is about 1.8 points. For a small chain the two combine to about 4.2 points, worse than consensus. Setting the combined error equal to the 3-point consensus error gives the break-even size: about 49 thousand card transactions a quarter.

Three things, each testable. First, the history: how many past quarters exist, so the bias can be measured per retailer and not assumed; with 60 retailers and four quarters a year there are only 240 results a year to learn from. Second, stability: banks join and leave card panels, so check that coverage of each retailer has not jumped, which would look like a sales surge. Third, the legal basis: how the data was collected and anonymised, and whether compliance is satisfied that it is not inside information. Then estimate the economics: the trade only exists for the larger retailers, inside a window that shrinks as more funds buy the same feed.