Statistical Arbitrage: The Count Does It
Statistical arbitrage trades a statistical relationship across many instruments at once, rather than a price difference between two identical ones. The relationship is expected to hold on average across the whole book, not on any single position, so the edge per trade is tiny and the count does the work.
A relationship traded across many instruments, not a mispricing between two. The arithmetic decides it, not the idea.
How it works
A statistical relationship traded in bulk. Instruments that have behaved a certain way relative to each other are expected to keep doing so on average — never on any one position.
True arbitrage exploits an identity; this exploits a tendency. One is enforced by the instruments being the same thing, the other only by history.
It needs many positions, not one good one. The edge per position is too small to survive alone; the result is the average across a large book. A handful is a different, riskier strategy.
Each edge is tiny and the count does the work. No single position is expected to be right; probability does the rest. Breadth is a requirement, not a preference.
What it usually is underneath
Most versions are mean reversion with better plumbing. A basket that has moved apart is expected to move back together, the relationship usually a correlation; pairs trading is the two-instrument case.
And a relationship can stop holding without warning. Nothing enforces it. A merger, a rule change or a shift in what a company does can end it in a week.
When everyone holds it, the unwind is the risk. The same measurements point many participants at the same positions, so if one must reduce quickly the selling moves the prices assumed independent. Relationships stable for years break in days.
It is an engineering problem more than a trading one. Three operational things decide viability: the cost of transacting, whether the short side can be borrowed, and whether small orders move prices.
In practice
At this edge size, costs are the whole question. A round trip on this site’s shared 576-bar history costs 0.0098 price units — 2% of a median bar’s range, 45% of the smallest.
And liquidity decides which instruments qualify. Thin volume widens the bid-ask spread you pay and makes your own orders visible, so an instrument can be well modelled and untradeable.
The results only make sense over thousands of trades. A short record cannot separate a working book from a broken one — backtesting carries the weight, and overfitting is the danger.
The rare shock is under-represented in every sample. An opening gap prices the position where the market opens, not where the model expected — and those events are rarest here.
A stop loss on one leg breaks the whole structure. Close one side and a hedged position becomes a directional one — and small edges are usually scaled with leverage, which makes it worse.
Every round trip costs 2% of a bar. It exceeds 10% of a bar’s range on 15 of the 576 bars here. A book turning over constantly makes that assumption the model.
Four questions for anyone selling one
Ask how many positions it holds at once. If the answer is a handful, it is not this strategy but a directional bet wearing the name, and should be judged as one.
Then ask the average holding period and the assumed cost per trade. Both have to be stated as numbers. A method that cannot say what it pays to transact has not been costed.
Then ask for the same result with the cost assumption doubled. At this edge size that change usually removes the whole reported return, which is why it is rarely shown.
If any of the four cannot be answered, there is nothing to evaluate. That is not a hostile position. It is the minimum a serious quantitative analysis produces for itself, and a missing answer settles it faster than checking the mathematics.
What statistical arbitrage is not
It is not arbitrage. That exploits an identity; this exploits a tendency that can end.
It is not high frequency trading. Speed helps it; speed is not what it is.
It is not a retail strategy. The constraints are borrow, cost and execution, none bought with software.
It is not a small strategy scaled up. Run it on a few positions and it becomes something else.
When it fails
The market went quiet
In a calm market it looks like a savings account. Inside a trading range the relationships hold, the book grinds upward, and a short record reads as proof.
Everyone was in the same trade
Crowding turns independent positions into one position. When the unwind starts, the correlations all go the same way at once and diversification disappears.
The short side could not be borrowed
A hedge you cannot put on is not a hedge. Borrow availability changes without notice, and the leg that vanishes holds the structure together.
The costs were assumed rather than measured
A round trip here costs 2% of a median bar’s range and 45% of the smallest. An edge of a couple of percentage points does not survive an optimistic cost figure.
The history was mined rather than tested
Test enough relationships and some will look permanent. Overfitting is not a beginner’s error here but the central methodological problem, and computing power makes it easier.
The market started travelling
The ten-bar efficiency ratio here has a median of 0.34, with 30% of bars above 0.5. Most movement is churn rather than travel, the condition the approach wants — and a directional run hurts it.
The original data
In research/broker-coverage.json, a scan of the 31,760 videos in
research/search-study-corpus.jsonl, statistical arbitrage returns 1 video, 13 views, 1
channel — the smallest audience recorded for any term measured on this site, against 896 videos
on scalping at a median of 19,949 views. Read that as a fact about what can be sold, not a verdict
on the strategy: it cannot be taught as something a person does at a screen, so nobody makes the
video and the vacuum is where the bot pitches live.
And research/series-measurements.json, built by site/measure_series.py, sets the size of the
edge in question. A higher close ten bars later occurred in 54% of 566 observations on this
576-bar history, and 52% over one bar across 571. Two percentage points over doing nothing is the
order of magnitude these strategies work with. So ask what happens to any reported result when the
cost assumption doubles — at this edge size that one question settles most of it.
Related
Arbitrage is the version enforced by an identity, and why this one needs a different name.
Mean reversion is what most implementations are underneath, once the infrastructure is stripped away.
And quantitative analysis is the measurement work it rests on, traps included.
I looked hard at this years ago, and the honest answer is that I cannot run it and neither can almost anyone reading this. The idea is not the hard part; the borrow, the cost per fill and the ability to work a large book of small orders without moving anything are the hard part, and none of that comes with a platform subscription. What I do respect about it is that you cannot fudge a strategy whose entire result lives in the arithmetic. I just work in a lane I can actually execute in.
— Michael Whitman
This page is educational, not financial advice. Test every idea on your own charts before risking money.