12 / 97
Censoring
Working definition
The thinning of a price feed — updates dropped, coalesced or rate-limited — during the intervals in which the underlying market is moving fastest, so that the record understates the spread widening and depth withdrawal that actually occurred.
A price feed is a sample of a market, not the market itself. Censoring is what happens when that sample is thinned at exactly the moments it matters — updates dropped, coalesced, or rate-limited during the seconds in which the underlying market is moving fastest.
The effect on the display is specific and counter-intuitive: the chart gets calmer as the market gets wilder. Fewer updates mean fewer visible spread excursions, fewer visible gaps, and a bid-ask that appears to hold its usual range through an event it did not actually hold through. What the trader sees is a stable, executable-looking quote. What arrives is a rejection, a requote, or a fill at a price the display never showed — the quote staleness that censoring manufactures.
Not every thinned feed is deliberate. Bandwidth limits, aggregation windows, conflation in the distribution layer and plain infrastructure saturation all drop updates, and they drop them hardest under the same conditions. That is the point: intent is not observable from the tape, and it does not need to be. What matters to a trader — and what is testable — is whether the feed being quoted preserves the event structure of the market being traded.
Which makes this a measurement problem rather than an accusation. Update rate per second through an event window against the same window on a reference feed, inter-tick interval distributions, the ratio of quote updates to trades, and whether gaps in one feed line up in time with spread excursions in another: all of it is recoverable from captured ticks, and all of it separates a feed that thinned from a market that quietened.
The consequence for research is the one most often missed. A backtest run on a censored history understates spread widening and understates the depth withdrawal underneath it, because the widening was real but the record of it was never kept. The simulation then fills orders at prices that the feed itself had already stopped being able to offer.
Commonly confused with
Neighbouring concepts that get used interchangeably, and the distinction that actually separates them.
- Quote staleness
Staleness is the symptom censoring manufactures. A thinned feed leaves each displayed quote representing an increasingly historical state, so the trader sees a stable executable-looking price and receives a rejection, a requote, or a fill the display never showed.
- Infrastructure limits
Bandwidth ceilings, aggregation windows, conflation in the distribution layer and plain saturation all drop updates, and they drop them hardest under exactly the same conditions. Intent is not observable from the tape — and does not need to be, because what is testable is whether the feed preserves the event structure of the market.
- Low volatility
The counter-intuitive part: a censored chart gets calmer as the market gets wilder. Fewer updates mean fewer visible excursions and a bid-ask that appears to hold its usual range through an event it did not hold through. Quiet on the display and quiet in the market look identical.
- Survivorship bias
Both are absences in a record, and they differ in what is missing. Survivorship removes whole entities from a universe; censoring removes moments from a series. One buries the losers, the other buries the worst seconds of everyone.
How to measure it in your own data
A definition you cannot test is a definition you have to take on trust. This is the shortest honest route from the concept to a number you computed yourself.
- Records you need
Captured ticks from the feed under examination and from a reference feed over the same windows, with timestamps. Nothing here requires the venue's cooperation or any claim about what it intended.
- What you compute
Update rate per second through an event window against the same window on the reference; inter-tick interval distributions; the ratio of quote updates to trades; and whether gaps in one feed line up in time with spread excursions in another.
- What the answer tells you
All of it separates a feed that thinned from a market that quietened, which is the whole question. The consequence most often missed is for research rather than for trading: a backtest run on a censored history understates both the widening and the depth withdrawal underneath it, because the widening was real and the record of it was never kept. The simulation then fills orders at prices the feed itself had already stopped being able to offer.
If this has already cost you
Whether a feed thinned during the minutes that mattered is measurable from captured ticks against a reference.
- Feed Fidelity Assay“Is the feed my terminal shows me behaving consistently?”Will not establish: A verdict on your broker. One exported log from one terminal measures your feed as you received it — venue-side behaviour, other account tiers and intent are all outside what this data can carry.
- Dataset Forensics Assay“Can I trust the data file my research stands on?”Will not establish: That the data is correct. Structure can prove a file defective; it cannot prove it faithful to the market — only a second independent source can, and the report names that limit.
Intake is not open yet, so none of these can be commissioned today. They are listed here so you know the measurement exists and what it would and would not settle — the launch list hears first.
Questions and answers
How can a chart look calmer during a violent market?
Because the chart is drawn from a sample, and thinning the sample removes the extremes first. Fewer updates mean fewer visible spread excursions and fewer visible gaps, so the bid-ask appears to hold its usual range through an event it did not hold through. The display becomes more comfortable as the market becomes less so.
Is censoring always deliberate?
No, and treating it as an accusation makes it harder to investigate. Bandwidth limits, aggregation windows, conflation in distribution layers and simple saturation all drop updates under precisely the conditions that matter. Intent is not observable from the tape. What is observable — and what actually affects you — is whether the feed preserves the event structure of the market you are trading.
How would I detect it?
By comparison, from captured ticks. Update rate through an event window against a reference feed, inter-tick interval distributions, the ratio of quote updates to trades, and whether gaps in one feed coincide with spread excursions in another. Together these separate a feed that thinned from a market that genuinely quietened.
Why does censoring matter for backtesting?
Censoring's backtest consequence is the one most often missed. A history recorded through a censored feed understates spread widening and understates the depth withdrawal beneath it — the widening happened, the record did not keep it. A simulation built on that history fills orders at prices the feed had already stopped being able to offer, and reports the result as achievable.
Related terms
Derived from the links this entry makes and the entries that link back to it.
Where the term is used
Instrument pages whose published copy uses this term. Each page states what it measures and what it does not establish.
In the research
Censoring comes up in three research notes on this site.