Strategies · intermediate · 9 min

Liquidity Sweeps

Below every obvious low sits a cluster of stop-loss orders — protective sells left by people who bought and by traders positioned for a breakdown. A liquidity sweep (or stop hunt) is the idea that price sometimes dips just below that low, triggers all those stops, absorbs the liquidity they release, and then reverses back up — leaving a long wick and a failed breakdown. The popular strategy is to fade it: treat the reversal back above the swept low as a signal to go long, on the theory that the "real" move is the one that follows the trap.

The pattern is real and easy to point at after the fact. This lesson asks the harder question: does fading it survive a random control? The answer is the most useful "no" in this whole track.

The mechanical rule we tested

A liquidity sweep of a recent low, followed by a reclaim, is expressible cleanly:

• Entry: today's low undercuts the lowest low of the previous 10 bars and today closes back above that prior low (LOW < REF(LOWEST(LOW, 10), 1) AND CLOSE > REF(LOWEST(LOW, 10), 1)) — a false breakdown that reclaims.
• Exit: a fixed 5% take-profit or 3% stop, whichever comes first.

Run on SPY, five years of daily bars. Unlike the order block and fair value gap rules, this one does capture the essence of the strategy faithfully — the sweep-and-reclaim is a single-bar event, so no zone memory is needed. Which makes its failure all the more worth reading.

The honest verdict — read every gate

The rule traded 46 times and returned about +7% over five years — against a market that returned about +73% buy-and-hold. On the headline alone it was already a poor way to be long. But the gates say something sharper and more general than "it made less."

The out-of-sample gate fails with the classic curve-fitting signature. Split the timeline: on the first 70% the rule made about +17%; on the held-out final 30% it made about −9%. A rule that works on the data you can see and stops working on the data you held back is the fingerprint of a pattern that fit the past rather than found something durable. This is the single most important failure mode in all of strategy testing, and here it is on a named, popular idea.

The random control fails too — and fails badly. The real +7% beat only about 32% of its 500 randomly-timed twins, meaning roughly two out of three scrambled versions did better than the real signal. The null average return was about +18% — more than double the real result. Fading sweeps did worse than entering at random and holding the same target. The significance test is near zero (t = 0.38).

Why this one matters most

The order block and fair value gap failed quietly — they made money and simply couldn't out-time their own scramble. This one fails loudly and in two independent ways: it overfit (great in-sample, negative out-of-sample) and it was beaten by random timing. That combination is the thing every strategy seller's backtest is silently hiding. They show you the in-sample line going up and stop there. Visor shows you the held-out half and the scrambled twins — and on this famous, intuitive, heavily-taught idea, both say the same thing.

None of this proves that no one can read a stop hunt well in the moment. It proves that this mechanical version, on this data, was not an edge — and that the way you find that out is the held-out split and the random control, not the headline return.

Try it yourself

What to read next

• The Overfitting Trap — the in-sample/out-of-sample failure, in depth.
• Why Most Strategies Fail the Gate — the gallery of famous strategies that beat zero and fail the control.
• The Random Control — the machinery behind the scrambled-twin test.