Carry-Forward Artifact
A stretch of a market dataset in which missing values were filled by repeating the last known price, manufacturing bars that look like market stability but actually record the absence of data.
When a data pipeline has no tick for an interval it must emit something, and the path of least resistance is to repeat the last known price. A bar builder obligated to produce one bar per minute emits a zero-range bar at the previous close. A feed merge papers over one source’s gap with its last value while the other side keeps updating. A vendor pads a missing session so downstream consumers do not encounter holes. In every case the output is a carry-forward artifact: data-shaped silence, indistinguishable at a glance from a genuinely quiet market.
The signatures are mechanical once looked for. Runs of identical open-high-low-close values far longer than a real market produces. Volume recorded as zero — or absent — beneath apparently stable prices. One side of the quote frozen while the other moves, because carry-forward was applied per side, producing spread structure no market maker ever quoted. Update cadence that collapses in exactly the windows where other venues show activity — which distinguishes an artifact born of pipeline failure from one born of censoring, though from the inside of the dataset the two can look identical. This is territory where quote staleness hardens from a latency property into fabricated bars.
The backtest consequences run in one direction: flattery. Carried-forward stretches understate volatility, since manufactured flatness enters every estimator as if it were market behaviour. Worse, a simulation will transact against the phantom prices — filling stops, limits, and entries at levels no order could have achieved, because no market existed there to fill them. A gap honestly recorded is information: the feed was down, the session was closed, the ticks were suppressed. A gap papered over is contamination that survives every downstream statistical check, which is why detection belongs at the data layer and in tick data provenance discipline. The Data Forensics instrument we are building is intended to surface exactly these artifacts in a user’s own dataset.