The three ways a mean-reversion trade dies

A mean-reversion system that wins 70% of the time is easy to build and hard to survive. The difficulty is not the win rate. It is that the 30% does not arrive evenly — it arrives in clusters, and each loss in the cluster is larger than any of the wins that preceded it.

So the useful question is not “how often is it right.” It is “what does being wrong look like, and how big is it.”

Three shapes. They are worth learning to tell apart on the way in, because the right response to each is different.

1. The anchor moved

The most common death, and the one that feels most like bad luck.

You buy something two standard deviations below its mean because you expect the gap to close. The gap does close — because the mean comes down to meet the price. Six weeks later the z-score is 0.0, the position is down 22%, and the screen no longer shows the name as oversold. Technically the trade worked. Financially it did not.

What it looks like going in: the recent selling is not one event but a persistent grind. Forward estimates are drifting down. The story you hear is about a structural change in the business — a lost contract, a product cycle ending, a competitor with a better cost structure — rather than a quarter that missed.

The response: this is what the anchor-stability filter on the screen is for, and it is why the exit rule has to be defined in terms of the mean rather than price. If the 50-day mean has fallen more than some threshold since entry, the thesis is dead regardless of what price did. Exit on that, not on a percentage drawdown.

2. The leash broke

Rarer, faster, and the only one that can end a career.

Fraud. A failed refinancing in a company whose leverage you thought was fine. A takeover struck at a price below your cost. A single regulatory decision that permanently removes the earnings you were valuing. There is no reversion because the thing you were anchored to no longer exists.

What it looks like going in: usually, honestly, it does not. That is the defining property. What you can see beforehand is whether the name has the capacity for total loss — leverage, a going-concern qualification, a single binary catalyst, an auditor change, a business model that requires continuous access to capital markets.

The response: this failure mode is not managed with stops, because it gaps through them. It is managed entirely with position size and with the quality filter that keeps the highest-capacity-for-zero names out of the universe in the first place. Anything that could be a 100% loss must be sized as if it will be.

3. Death by a thousand small right answers

The subtle one, and the one that quietly eats retail accounts.

The system works. Win rate holds. Every individual trade behaves roughly as expected. And the account goes sideways for a year, because the average win is 2.1% and the average loss is 6.8%, and 70% × 2.1% − 30% × 6.8% is negative before a single commission.

What it looks like going in: it does not look like anything, which is why it needs to be caught in arithmetic rather than in judgment. Expectancy is the whole of it:

E = (P_win × avg_win) − (P_loss × avg_loss)

A system is only worth running when E is positive after costs, slippage, and the tax treatment of short holding periods. Reversion trades are short-horizon by construction, which means they generate short-term gains taxed at ordinary rates — a real drag that backtests routinely ignore.

The response: measure expectancy on your own filled trades, not on a backtest, and re-measure it every quarter. If avg_win / avg_loss is below about 0.4, no achievable win rate saves the system.

The sizing arithmetic

All three failure modes have the same defense, which is why sizing is the strategy rather than an accessory to it.

Fix the loss, not the position. Decide what a single trade being wrong costs the account — call it R, and for most individual accounts R belongs somewhere between 0.5% and 1% of equity. Then:

shares = (R × account_equity) / (entry_price − stop_price)

The position size falls out of the stop distance. A wider stop means a smaller position, automatically. This is the mechanism that prevents the most dangerous habit in reversion trading, which is sizing up on the names that have fallen furthest — precisely the ones most likely to be anchor-moved or leash-broken.

Two constraints on top:

Correlation. Six oversold regional banks is one trade with six tickers on it. Reversion screens cluster by sector by construction, because sectors sell off together. Cap total exposure per sector, not just per name.

Adding to losers. Averaging down is structurally native to mean reversion — the further it falls, the better the signal says the entry is — and it is the mechanism by which failure mode 1 becomes catastrophic rather than annoying. If you scale in, the total planned risk across all tranches must be decided before the first one, and it must fit inside R.

The honest summary

Mean reversion is a strategy that pays you frequently and bills you occasionally, in size. Everything in the process — the filters on the screen, the exit tied to the mean rather than the price, the fixed-R sizing, the sector cap — exists to put a ceiling on the bill.

None of this is advice, and nothing here is a recommendation to buy or sell anything. It is what I do with my own book, published so that when it stops working I have a record of what I believed and why.