Why is my live spread wider than my backtest?

The short answer

In most cases the live spread is not wider than it should be. The backtest spread was narrower than anything that ever existed, and the gap you are looking at is the size of that fiction.

There are four causes, they are separable, and they have different remedies. Knowing which one is yours matters more than knowing the total, because three of the four are fixable in your own process and one is not fixable at all.

The four causes, in the order they are usually guilty

1. The backtest used a constant where a distribution belongs

The constant-spread error is the common case and the largest. A single spread value — whether the broker’s advertised typical, the average of the file, or a number chosen because it felt conservative — removes the property that matters most: spread and volatility are correlated, and so are volatility and your entries.

If your strategy trades breakouts, news reactions, or anything triggered by movement, its entries are concentrated in precisely the conditions where the spread is widest. Costing those entries at the all-hours average is not an approximation. It is a systematic underestimate whose size scales with how much your strategy depends on movement.

The test is direct: recompute the backtest using the spread that was actually quoted at each entry timestamp, and compare. The difference is what the constant was hiding.

2. The historical file’s spread column is reconstructed, not recorded

Many historical datasets carry a spread that was never quoted to anyone. Mid-price series get a synthetic spread applied afterwards; gaps get filled by interpolation that smooths away the widest moments; files exported from one account tier get traded by someone on another.

This is a tick data provenance problem, and it is checkable. Record your own quote stream for a period, then compare its spread distribution against the same period in your historical file — not the means, the upper tail. Reconstruction shows up in the tail, because that is where interpolation does its damage and where the average hides it.

3. Your fills are not at the quote you were shown

Distinct from spread, and often mistaken for it. The quote was real; you did not get it. This is slippage, and mixing it into the spread question means you will try to fix a routing problem with a cost assumption.

Separate them arithmetically: expected cost at the quoted spread on one side, realised implementation shortfall on the other. What remains after subtracting the first from the second is the part the spread never explained.

4. Your own size is moving the price

The smallest cause for retail, and the one nobody can remove. If your order consumes more than the top of book, the average price you pay is worse than the quote by construction, and no backtest that fills at a single price models it.

Check by binning realised cost against order size. A flat relationship means size is not your problem. A relationship that steepens past a threshold is your book depth, and it is the one cause on this list with no fix other than trading smaller or slower.

A measured floor, and the part of it that is not the spread

The artifact published with this article measures the quoted-spread distribution for twenty-eight pairs and, for the seven where the commission converts, adds commission to give a cost floor — the minimum move that has to happen before the trade breaks even.

Pips. Commission from a named raw/ECN schedule, supplied by me.

On every one of the seven, the commission is larger than the median spread — on EURUSD by six times. If you are arguing with your backtest about a tenth of a pip of spread while the commission line is six tenths, the spread was never the main term. That is worth knowing before you spend an afternoon tuning the assumption that matters least.

Read the account type before transferring this to your own: on a commission-free account the same cost reappears inside a wider spread, so the floor is comparable while its composition is not. The pairs, not the totals, are what transfer.

Three limits, all of them in the file rather than in this paragraph. Twenty-one of the twenty-eight pairs have no cost floor at all — they are crosses whose commission is quoted per USD notional, and converting it needs a third USD leg this lane does not hold, so each reads NOT YET COMPUTED rather than carrying a guess. Zero-spread and crossed ticks are counted and excluded, never averaged in, which makes every floor above conservative in a stated direction: folding them back would lower it, never raise it. And the floor is a threshold, not a prediction — it states what a move must beat, and nothing on that face implies a move that beats it exists.

What the decomposition is worth

A strategy that survives realistic costs and one that survives an average is not the same strategy, and the difference is not marginal. The category of system most damaged by this is the one that trades often for small edges, because cost error compounds with trade count while edge does not.

The uncomfortable version: a great many strategies are profitable under a constant spread and unprofitable under the distribution that actually occurred. Discovering which one yours is costs a few hours of recomputation. Discovering it live costs whatever you funded the account with.

Where this leads

Costing entries at the spread that was actually quoted, and attributing the remainder honestly, is what the Feed & Execution instruments do as a matter of routine rather than as an investigation you mount after a bad month. The spread regime and execution cost definitions are in the glossary if you want the terms used precisely, and why backtests fail in live trading covers the failures that survive after the cost model is fixed.