Home / Insights / 02
DATA QUALITY

Silent biases in cross-source data

When two vendors report the same price differently, the difference is usually small. Over a year, those small differences compound.

The discrepancies you do not see

If you have ever compared a closing price from two vendors and found them the same, you have probably looked at a liquid, well-covered name on a normal day. On that day they will agree to within a fraction. The discrepancies that matter are the ones on less-liquid names, on dates around corporate actions, and at the boundaries of the trading day.

Time-zone boundaries

An Australian close recorded as 16:10 Sydney time and a New York vendor's "previous close" recorded in UTC are not the same instant. If you join the two series on a calendar date rather than on an instant, you will introduce a half-day lag that shows up as an unexplained correlation shift around daylight-saving changes.

Corporate-action timing

One vendor applies a dividend adjustment on ex-date, another on payment date. The difference is usually one to six weeks. If you are computing short-window returns around ex-dates, the choice materially changes the result.

Less-liquid names

For names with wide spreads or suspended periods, two vendors can legitimately report different last-trade prices because they are using different last-trade timestamps. Neither is "wrong", but if you treat them as interchangeable, your history will show phantom volatility.

Our view: The fix is not to average the two sources. The fix is to declare one source authoritative per field, document the convention, and quarantine the dates where the other source disagrees by more than a threshold. That rule, written down, removes most of the silent bias.