Why do my forex orders get rejected?

The short answer

Because your counterparty was allowed to say no. Across much of over-the-counter FX an order does not become a trade on arrival: it enters a short decision window — last look, covered in full in the glossary — and comes out the other side as a fill, a rejection, or a requote. A rejection is that window’s ordinary output, not a malfunction, and the window itself is a disclosed practice with a defensible history the glossary entry walks through.

What the glossary cannot do is read your log. The useful question about your own record is not why was this one declined — it is whether the declines fall evenly, because every one of them was a decision taken by an interested party, on every order, out of your sight, and even-handedness is precisely the property a decision like that cannot be assumed to have.

Why a rejection rate on its own tells you nothing

The number almost everyone reaches for is the share of orders that were declined. On its own it is close to uninformative.

Consider two counterparties, each declining two orders in a hundred. The first declines them without reference to what the market did during the hold. The second declines only where the interim move went against it, and fills every order where the move went its way. Both publish the same rejection rate. They are not the same instrument, and the difference is the entire subject.

What separates them is conditioning. Not how often were you declined, but what was true when you were. A rejection rate is a marginal statistic in a question that is conditional all the way down, which is why a fill ratio quoted without its conditioning is a brochure figure rather than a measurement.

The measurement, step by step

Every step below runs on records you already hold, or reveals that you do not hold them.

1. Establish whether your export retains rejections. Filled-only is the common case. If that is what you have, stop and note it: the question is unanswerable from your data, which is a real finding about your instrumentation rather than a failure of the method.
2. Recover the hold interval. Send timestamp to decision timestamp, per order. The distribution matters more than the mean — a tail of long holds is a different behaviour from a uniform short one.
3. Attach a reference price at both ends of each hold. This is the step that decides whether the exercise is worth anything, and it is the hard part: the comparison needs an independent record of the market across the window, not the same feed whose behaviour is in question.
4. Classify each order by interim drift. Did the reference market move in your favour, against you, or not materially, between send and decision?
5. Compute the decline rate within each class, and print n on every one. Three conditional rates with their sample sizes, not one headline number. Where your platform records requotes as their own outcome, count them beside the declines in each class — a requote is a third answer, not a subspecies of fill.
6. Do the same for hold duration. Compare the distribution for accepted orders against declined ones.

The tell is symmetry. If the three conditional rates from step five sit close together, the window is doing what its disclosure says it does. If the against-your-counterparty class carries most of the declines while the favourable class fills, you now hold a description of one-directional behaviour — three numbers with their sample sizes, stated without any claim about why.

The Survivor’s Fill

Here is the part that survives even if every rejection in your record turns out to be even-handed, and it is the reason this matters beyond one disputed order.

Your fills are a sample somebody else selected. Every trade in your history is there because a counterparty chose to let it be there, in a structure where it could have declined. That makes the filled set a survivor population, and every statistic you compute on it inherits whatever did the selecting.

This is survivorship bias pointed at your own order flow, and it is more awkward than the usual kind, because the selection is not an accident of record-keeping. Where declines fall on one side of the interim drift, the orders removed from your flow are precisely the ones that had already moved your way. What is left — what becomes your fills, your average slippage, your measured execution cost — is the remainder after the good ones were taken out.

None of this shows up in a slippage report. An order that never became a trade contributes no row; its replacement fills at whatever the market has since become; and the difference between the two is booked by nobody, because it is filed under market conditions. That is the defect I am naming the survivor’s fill: not that a fill is wrong, but that a record made only of fills cannot price its own rejections, and will report a cost that is too low by exactly the amount it cannot see.

Where the measured version publishes

Steps one through six are a spreadsheet and an afternoon, and step three is where the afternoon stops being enough. Setting each hold against the market that was actually running during it needs an independent, time-aligned record of that market — and that record, not the arithmetic, is the part no spreadsheet supplies.

No published instrument yet claims the rejection half of this measurement, and I would rather say that plainly than stretch a page to imply otherwise. What is published divides the problem in two. The fills half — signed slippage as a distribution, cost against the quote of record — is the Execution Cost Auditor, whose scope is bounded by what your statements contain: exactly the fills-only boundary this article is about. The reference-record half — an independent, time-aligned journal of what the market was doing during a hold — is the Broker Feed Auditor’s territory, and the subject of is my broker’s feed honest. And the single-event version — one disputed window reconstructed with requote and rejection context where your logs carry it — is the Single-Trade Investigation on the assay desk. Both instruments are pre-launch and carry outputs marked NOT YET PUBLISHED; pattern-level rejection analysis joins the published set when it can carry its receipts, and not before.

If the question you actually have is about one order rather than a pattern, the narrower version is what can a trading statement prove, which is about what a record establishes and what it cannot.