Independent strategy research
Do mean-reversion strategies work? We tested 90+.
Short version: sometimes, and rarely. We put 90-plus mean-reversion strategies through the same real-cost model we run on everything, and about four in five failed. That leaves a minority — roughly one in six — that held an edge, but only under conditions the seller never mentions. Mean reversion is the counterexample to grid and DCA bots: it does not die at 100%. It dies at about 80%, which is a more interesting number.
What mean reversion is, and where it works
A mean-reversion system bets that price has stretched too far and will snap back. Fade the push, take the bounce, do it again. In a sideways market it is close to printing money. Every jab into the top of the range gets sold, every flush into the bottom gets bought, and the equity curve turns into a tidy staircase. That staircase is the sales screenshot. It is also the whole trap, because the market does not range forever.
What happened when we tested 90+
We ran 90-plus mean-reversion systems across CME futures, tick-level FX with real bid/ask, and crypto spot and perps. About 80% failed once real spreads and slippage were in the model. Put the survivors next to the whole audit and the ordering is honest: mean reversion rejects a little less than reversal and breakout systems, a little more than momentum, and it sits close to the middle of the pack. Nowhere near the best category. And nowhere near the clean 100% wipeout that grid and DCA post.
So it is not a dead category. It is a picky one. The failure is not that the idea is fake — price really does revert, in the right place. The failure is that most of the published versions revert in the wrong market, at the wrong speed, or on a signal that was never there.
Why four in five fail
Two failure modes do almost all the damage.
Trend runs them over. A mean-reversion book is short trend by construction. As long as price oscillates, fading pays. The moment a market commits to a direction and holds it, every fade is a new loss and the staircase unwinds in a straight line. Getting run over by a trend it kept fighting is one of the single largest causes in our whole reject pile. The curve above is not a metaphor. It is the shape of the drawdown.
Cost eats the fast ones. The higher-frequency versions fade constantly, and every fade pays the bid-ask spread and slippage. On the fastest systems the edge per trade is thinner than the spread it has to cross — cost-fatal before it books a dollar. Sub-15-minute FX is the cleanest example: the spread swallows the reversion edge before it can exist, which is why we stopped testing FX and CFDs below 30-minute at all. And a large share of the rejects had no real edge to begin with — the "reversion" was noise that happened to look mean-reverting on one chart, in one window, on one instrument.
The honest exception: the minority that survived
Here is where mean reversion parts ways with grid and DCA. A real minority held up. Not a rounding error, not a fluke of one lucky chart — roughly one in six carried an edge that stayed positive after full costs. We do not get to pretend those away, and we do not.
What separated them was boring and specific. They fixed on instruments that actually revert, and most instruments do not. They traded slowly enough that cost was a rounding error instead of the entire P&L. And they carried a hard exit for the day the range breaks, rather than leaning into the trend and hoping. That is a conditional edge, not a deployable-anywhere one. It lives on a specific market at a specific timeframe. Move it one instrument over and it is usually gone. Real, narrow, and nothing like the "works in any market" pitch.
How we test
Every strategy is ported to Python and run against real costs — spreads and commissions modeled from tick data, not a guessed flat number. Futures come from Databento, 13 years of CME. FX comes from Dukascopy with real bid/ask. Stocks run with liquidity-aware fills, crypto as spot and perps. A fast model does the bulk porting; the strongest model then tries to break every apparent survivor, hunting look-ahead bias and impossible fills, and an independent supervisor watches the pipeline around the clock. We hash the code, so a strategy re-published under three names gets tested once. It is the same process that rejects roughly 78% of everything we run — and about 80% of the mean-reversion strategies inside that. See the wider research hub for how other categories fared.
Research and education, not financial advice. No signals, no return promises. Independent, and not affiliated with TradingView.
Which ones survived, and who published them?
You now have the aggregate truth for free: most mean-reversion strategies fail, a narrow minority holds a conditional edge, and the split is about four-to-one. What this page does not give you is the names. Which specific published mean-reversion scripts we tested, who wrote them, the exact after-cost numbers, and which handful actually survived — with the verdict and the reason for each. That is The No List: every strategy we audited, named, graded, and explained.
Get The No List →FAQ
Do mean-reversion strategies actually work?
Some do, most do not. About 80% of the 90-plus mean-reversion strategies we tested failed a real cost model. Roughly one in six held a conditional edge — real, but tied to a specific instrument and timeframe, not the "works anywhere" version being sold.
Why do mean-reversion strategies fail?
Two reasons dominate. Trends run them over: a mean-reversion book is short trend, so a market that keeps going one way turns every fade into a loss. And cost eats the fast ones — on short timeframes the spread and slippage are larger than the edge per trade.
Is mean reversion better or worse than grid and DCA bots?
Better, and it is not close. Every grid, DCA and martingale bot we tested failed — 100% of the category. Mean reversion fails at about 80%, and a genuine minority survives. It is a picky category, not a dead one.
What makes the mean-reversion strategies that survive different?
They trade instruments that actually revert, they trade slowly enough that cost is a rounding error, and they cut out hard when the range breaks instead of averaging into a trend. The edge is real and narrow — change the market or the timeframe and it usually disappears.