When Two Gold Feeds Disagree, Match the Sides and the Clocks

Write down what each feed measures

For each source, capture the precise product, currency per unit, venue or provider, quote side and available size. Store the provider timestamp and a separate local receipt timestamp. Also keep the raw message and its feed identifier. A last futures trade, an indicative spot midpoint and an account-specific bid are three distinct inputs even if a chart labels each one “gold.”

OANDA’s v20 pricing documentation is a useful concrete example. Its account price object separates bids, asks and available liquidity. Either side may be empty. The streaming endpoint delivers at most four price updates per second per instrument and subscribers’ sampling windows can differ. As a result, even two connections to that endpoint can observe different updates during rapid movement. A missing intermediate change is consistent with that documented sampling design.

Source notes: [1] OANDA — REST v20 Pricing endpoints · [2] OANDA — REST v20 Pricing definitions

Pair without borrowing a future quote

Suppose feed A reports a hypothetical bid of 2,999.80 and ask of 3,000.20 at 10:00:00.500 UTC. Its midpoint is 3,000.00. Feed B’s latest preceding quote arrived at 10:00:00.350 UTC, showing 2,999.90 bid and 3,000.30 ask, a midpoint of 3,000.10. A backward match gives a ten-cent midpoint difference with a 150-millisecond quote-age gap. These are invented inputs solely for the calculation.

If B’s next quote at 10:00:00.530 makes the difference disappear, substituting that later quote changes the question. It was unavailable at A’s observation time. For a historical display comparison either convention can be studied if named. For a reconstruction of information available at a decision instant, match only messages actually received by that instant, with synchronized and sufficiently accurate collection clocks.

Set the tolerances before inspecting the result

Choose the comparison grid, the quote side, the maximum acceptable age on either feed and the size being compared. Under a hypothetical 100-millisecond age limit the pair above is excluded. Under a 250-millisecond limit it is eligible if all other rules pass. Retain both thresholds as planned sensitivity checks. Widening the limit solely to rescue a favorite example makes the criterion difficult to defend.

For every scheduled observation record its eligibility. Summarize matched observations, stale exclusions and one-sided quotes separately. Then calculate the median signed difference to see direction and a high percentile of the absolute differences to see dispersion. Repeat by session using the same units. If a feed loses coverage during the busiest minutes, strong agreement in the remaining sample cannot establish agreement during those missing minutes.

Interpret a gap in the appropriate units

A midpoint gap describes a difference in two constructed centers. A claim about a transaction also needs the relevant buy-side ask, sell-side bid, comparable size, eligibility to trade and attainable execution. There are no collected feed observations or fills in this worked example. Its product is an audit procedure, including a way to mark “insufficient paired data” whenever the inputs cannot answer the question.