What is a break of structure?
A break of structure is price trading or closing beyond the most recent significant swing point in the direction of the prevailing sequence: a higher high in an uptrend, a lower low in a downtrend.
The related term, change of character, describes the first break against the prevailing sequence — the first lower low after a series of higher highs and higher lows. The distinction matters because the two are used for opposite purposes: one confirms continuation, the other flags a possible turn.
What makes structure appealing is that it is nearly mechanical already. Swing points are locally extreme prices; a sequence of them defines a direction; a break of the most recent one is an event with a timestamp. Very little judgement is required compared with, say, drawing a trend line.
The word 'nearly' is doing significant work, though, and it hides where every structure backtest goes wrong.
Why do most structure backtests use the future?
Most structure backtests use the future because a swing high is conventionally defined as a bar whose high exceeds the highs of the N bars on either side — and the N bars on the right have not happened yet when that bar closes.
This is the bug, and it is almost invisible because it feels like definition rather than leakage. If you compute all swing points across the dataset before walking through it, the swing high at bar 500 was only confirmable at bar 500 + N. A strategy that reacts to it at bar 501 is using information that did not exist. The result will look better than reality by a margin that grows with N.
The correct implementation confirms swings with a delay. A candidate high becomes a confirmed swing only once N bars have closed to its right, and only from that moment is it available to any rule. This makes every structure break arrive N bars later than the annotated chart suggests, which reduces the apparent quality of the setup — accurately.
The delay is not a flaw in the method; it is the honest cost of the definition. Anyone showing you structure annotations on a historical chart is showing you swings confirmed with hindsight, and the entries marked there could not have been taken at those prices.
How do you choose the swing lookback?
Choose the lookback based on the structure you actually intend to trade, then hold it fixed, because it controls both how much structure exists and how late every signal arrives.
A small lookback finds many swings, produces frequent breaks, and confirms quickly. It also finds structure in noise, so a large share of the breaks it reports are meaningless on any timeframe a human would care about.
A large lookback finds only major turning points and confirms slowly. The structure it finds is more likely to be visible to other participants — which is the mechanism that makes structure matter at all — but the confirmation delay means the break is old news by the time it is usable.
Because both effects are real and opposed, the value is genuinely an empirical choice per instrument and timeframe. Which makes it a parameter, which makes it something to fix in-sample and validate out-of-sample rather than to tune until the equity curve looks right.
Is a break of structure enough to trade on its own?
On its own a structure break is a description of what has already happened, and in most tests it needs to be paired with a location — a level, a zone, a gap — before it becomes a rule with any selectivity.
The reason is base rates. Breaks of structure are common; in any trending period there are many, and a rule that enters on every one is close to a trend-following rule with extra vocabulary. That is not automatically bad, but it should be compared against the simple trend-following rule rather than assumed to be better because the terminology is more specific.
The combination that most published sequences describe — a sweep of liquidity, then a break of structure, then a retrace into the zone left by the break — is more selective and more testable precisely because it is a sequence. Sequences can be encoded exactly: event A, then event B within N bars, then event C within M bars.
The tax on selectivity is sample size, and it is steep. Each additional required event can cut the occurrence count by a large factor, and three chained conditions can turn a thousand candidates into a few dozen. A few dozen is not enough to conclude anything, however clean the chart examples look.