Trading Strategies
Mean Reversion Trading: Fading Extremes Without Catching Knives
This method wins most of the time, which is exactly what makes it dangerous. The losses arrive in a few trades and the expectancy decides everything.
Seven winners in ten. That is what a well built mean reversion method will hand you. It is precisely why the method is dangerous. The win rate is the most flattering number in trading and the least informative one, because it says nothing about the size of the three losses, and in this method the three losses are the entire reason the equity curve goes where it goes.
Why prices overshoot in the first place
The premise has a mechanism behind it. That is more than most retail methods can claim.
A large seller has to be out by the close. Price is not his problem. Market makers absorbing that inventory need paying for the risk of holding it, so they quote lower, and once the seller is finished the inventory has to be worked back out and the quotes come up again. That sequence is set out in market makers and liquidity. It is why a reversion trade is more than a shape on a chart.
The horizon matters more than the signal. Over days to a couple of weeks, short term reversals are common. Over six to twelve months the documented tendency runs the other way, toward continuation, which is the subject of momentum trading. A rule that fades a one day drop and a rule that fades a nine month decline are two different trades wearing the same name.
Measuring an extreme
Define the extreme in units that adjust for how volatile the stock normally is. Otherwise a 5 percent fall in a utility and a 5 percent fall in a small cap look like the same event on your screen. They are nowhere near it.
| Measure | What it says | Where it misleads |
|---|---|---|
| RSI below 30 | Recent losses outweighed recent gains | Sits below 30 for weeks in a genuine downtrend |
| Close below the lower Bollinger Band | Price is beyond two standard deviations of its own average | Price walks down the band through an entire decline |
| Distance from a 20 day average in ATR units | The move in the stock’s own daily range | Says nothing about why the stock fell |
| Percent below a 50 day high | A plain drawdown measure | Treats a slow grind and a one day collapse alike |
The band and volatility measures are worked through in Bollinger Bands, ATR and Keltner Channels, the oscillators in RSI, MACD and momentum indicators. Both make the same point. An extreme reading describes the past. It owes the future nothing.
The filter is the strategy
The signal finds stretched prices. The filter decides which stretched prices are worth your money, and without one this method systematically buys companies in genuine trouble, because those are the names producing the most extreme readings. That is the failure mode. It is not occasional.
Three filters do most of the work. Require the stock to sit above a long term moving average, so the weakness you are buying sits inside an uptrend. Otherwise you are buying the next instalment of a decline. Exclude anything that fell on an event which repriced the business, an earnings collapse or a withdrawn outlook, because the average it would revert to no longer exists. And require ordinary liquidity, since the widest spreads in the market appear exactly when a stock is falling fastest and you will be paying them on the way in and the way out.
The stock screener applies the extreme and the trend filter in one pass. Across a large universe that is the only workable way to run it.
The expectancy trap
Put the win rate where it belongs, inside the formula. Watch what it is worth on its own.
expectancy = (win rate x average win) - (loss rate x average loss)
Case one, winners taken quickly and losers allowed to run. A 70 percent win rate, a $120 average win and a $300 average loss gives (0.70 x $120) - (0.30 x $300) = $84 - $90 = -$6 a trade. Seven trades in ten go your way and the account declines.
Case two, the same win rate with the loss capped. A $150 average win and a $250 average loss gives (0.70 x $150) - (0.30 x $250) = $105 - $75 = $30 a trade, or $6,000 across 200 trades.
Now subtract the toll. Those 200 round trips, each one 100 shares of a $50 stock with a two cent spread, cost 200 x $4 = $800 in spread, and $1 a side adds 200 x 2 x $1 = $400.
| Case one | Case two | |
|---|---|---|
| Win rate | 70% | 70% |
| Average win | $120 | $150 |
| Average loss | $300 | $250 |
| Expectancy per trade | minus $6 | $30 |
| 200 trades, gross | minus $1,200 | $6,000 |
| Costs at $6 a round trip | $1,200 | $1,200 |
| Net | minus $2,400 | $4,800 |
One behaviour separates those columns. Holding a loser past the stop, because the position is already oversold and therefore ought to bounce. Both columns advertise the same 70 percent. When somebody quotes you a win rate and not the average win, the average loss and the number of trades behind it, they have told you the least useful fact they possess.
Sizing when the tail is fat
The calculation is the one that runs every trade on this site, shares = dollar risk / stop distance. The discipline required is higher. The stop will feel wrong at the moment it is hit. That feeling is the method working.
On a $20,000 account risking 1 percent the budget is $200. A stop $3 below the entry, roughly two times the stock’s average daily range, gives $200 / $3 = 66 shares. A $2 stop gives 100 shares. The tighter stop is not the safer one here. It sits inside the noise you are deliberately buying, so it converts winners into losers at your expense.
These positions cluster. Extremes appear across a market at the same time. Four open reversion trades at 1 percent each is one 4 percent bet that the selloff ends this week. Formal sizing, including why full Kelly is far too aggressive once the inputs are estimates from your own small sample, is in risk management for traders and priced by the Kelly criterion calculator.
Where it fails
In trends, which is the mirror of momentum failing in ranges: a stock in a sustained decline produces oversold readings the entire way down, and every one of them is a valid signal by your rules and a loss on your statement.
At regime changes, when volatility rises across the market and daily ranges double. Stops calibrated to the old range get taken routinely. The average you are reverting to is itself moving.
On single stock events, where the fall is information and not flow. No degree of statistical stretch makes a repriced business revert.
And there is a fourth failure that only shows up in the record. Winners close quickly. Losers stay open. At any moment your screen is showing you the trades going against you, the book looks worse than the strategy is, and traders read that snapshot as proof the method has stopped working, then either abandon it at the wrong point or double the size to fix it. The journal is the only thing that sees through it. Measure expectancy across closed trades in blocks of fifty and compare blocks, never days.
The relative value version trades the spread between two related securities. It moves these problems around. They do not go away. Pairs trading and statistical arbitrage covers it.
Before running either with real size, work the sizing rules in how to start trading stocks. Then check yourself against the risk management and position sizing quiz.
Frequently asked questions
What is mean reversion trading?
Buying after a sharp fall or selling after a sharp rise, expecting price to come back toward a recent average within days. The trade takes the other side of whoever is being forced out of a position. It is the opposite stance to momentum trading, and it operates on a much shorter horizon.
Does mean reversion work on all timeframes?
The tendency for short term moves to reverse is documented most strongly over days to a couple of weeks. Over horizons of several months the evidence points the other way, toward continuation. Applying a reversion rule to a multi month decline means fighting the exact horizon on which momentum has been recorded.
Why do mean reversion strategies have high win rates and still lose money?
Because the wins are small and the losses are not. A method winning 70 percent of the time with a $120 average win and a $300 average loss has an expectancy of 0.70 times 120 minus 0.30 times 300, which is minus $6 a trade. The win rate feels excellent the whole way down.
What stops a mean reversion trade from becoming a falling knife?
A filter and a stop. The filter keeps you out of stocks that have broken down structurally, for example by requiring the name to sit above a long term moving average and excluding anything that fell on an event. The stop caps the loss in dollars, which matters more here than in any other method because the tail of losses is what kills the expectancy.
Is RSI below 30 a buy signal?
On its own, no. A reading under 30 says recent losses outweighed recent gains, which describes what already happened. In a real downtrend that condition can hold for weeks while the price keeps falling. It becomes usable only alongside a trend filter and a price at which you accept being wrong.
What is pairs trading?
The same reversion logic applied to the spread between two related securities rather than to one price. You buy the laggard and sell the leader, expecting the gap to close. The position is closer to market neutral, and it fails when the relationship itself has changed rather than merely stretched.