Andrea Unger on Statistical Arbitrage: Why “Free Money” Trades Aren’t Free

“Arbitrage” sounds like easy money. In finance textbooks, it essentially is: buy an asset at a lower price in one market, sell it at a higher price in another, and pocket the difference without taking on risk.

Statistical arbitrage, however, is not the same thing. It borrows the name, but not the guarantee. Understanding this difference is essential if you want a realistic view of what this strategy can actually offer.

Statistical arbitrage, commonly known as stat arb, refers to a family of trading strategies that seek to profit when a statistical relationship returns to its historical range after prices have diverged. Unlike pure arbitrage, it does not exploit a guaranteed price discrepancy.

There are no risk-free profits here. There is a probability, a model, and a real risk that the model could be wrong. This distinction matters more than any formula discussed in this article.

What Statistical Arbitrage Really Means

Statistical arbitrage is a short- to medium-term systematic trading approach that identifies pairs or groups of assets whose prices have historically moved together. The strategy is based on the expectation that when their prices diverge, they will eventually converge again.

The trade is not a bet on whether the market will rise or fall. It is a bet that a relationship, namely the spread between two prices, will return to normal.

This changes the type of risk involved.

A directional trader is exposed to the risk of the broader market moving against the position. A statistical arbitrage trader is, in theory, hedged against general market movements and primarily exposed to the possibility that the relationship between the two assets will break down.

In practice, that hedge is never perfect. The risk section later in this article explains why.

The word “statistical” is therefore crucial. Unlike pure arbitrage, which is based on a price discrepancy that can be exploited by buying and selling simultaneously, stat arb relies on a statistical estimate. This may involve a correlation, a historical spread, and the probability that convergence will occur before the position runs out of room.

That estimate can be wrong. This is the most common misunderstanding surrounding the strategy. It is also why the question “Is arbitrage trading legal?” often starts from the wrong assumption. Stat arb is not illegal, but it does not offer guaranteed results either.

It is a probability-based trade, not a certainty.

👉 Also watch Andrea Unger’s video: Trading: Tricks, Arbitrage & Market Inefficiencies – The Ugly Truth No One Talks About

How Statistical Arbitrage Works

At its core, statistical arbitrage relies on three elements:

• a relationship between two or more assets;
• a method for measuring when that relationship moves away from its normal range;
• a rule for deciding when to act.

Mean reversion is the central assumption behind all three elements. It follows the same basic logic as mean-reversion strategies on S&P 500 futures, but applies it to the spread between two assets rather than to the price of a single instrument.

Pairs Trading: The Simplest Form

Pairs trading is the most straightforward introduction to statistical arbitrage.

The process begins by selecting two assets that have historically moved together. These could be two oil companies, two banks, or two ETFs with overlapping sector exposure.

When the price relationship between them, known as the spread, moves farther away from its historical average than usual, the strategy takes opposite positions:

• a long position in the asset that has underperformed relative to the other;
• a short position in the asset that has outperformed relative to the other.

The trade is not based on the idea that “oil will rise” or “banks will fall.” The underlying assumption is that the two assets, which normally move together, will resume doing so after a temporary divergence.

If the spread converges, both legs of the position may contribute to the profit. The long position recovers while the short position declines, regardless of what the broader market does that day.

This is what “market-neutral” means in this context. The position is designed to seek profits from the relationship between the assets, not from the direction of the overall market.

To determine when a spread has moved far enough to justify a trade, traders commonly use a z-score. This measures how many standard deviations the current spread is from its historical mean.

A z-score of 2 indicates that the spread is unusually wide relative to its history. One common rule, although by no means a universal one, is to enter when the z-score moves beyond a threshold such as 2 and exit when it returns toward 0.

From Pairs to Baskets

Institutional stat arb rarely stops at two assets.

Funds may manage baskets containing dozens or hundreds of positions. They use statistical techniques such as cointegration to automatically identify and rank thousands of potential pairs or groups.

Cointegration is a more rigorous test than simple correlation. It helps determine whether a genuine long-term equilibrium relationship exists between the assets.

The underlying logic is the same as pairs trading, but it is applied on a scale that no individual trader could realistically manage by hand.

This is the version of statistical arbitrage commonly found in open-source projects and quantitative finance textbooks. It is also where many retail traders risk adding too much complexity.

Trying to replicate a hedge fund’s basket-level infrastructure with a spreadsheet can lead traders to underestimate both execution costs and data requirements.

It is generally better to start with pairs and increase the scale only after fully understanding the mechanics. The same progression applies when building any algorithmic trading system from scratch.

A Practical Step-by-Step Example

This example is provided for illustrative purposes only. It is not a trading recommendation and does not describe a real position. The mechanics matter more than the specific numbers.

Suppose two large-cap airline stocks, Company A and Company B, have historically maintained a stable price relationship.

For every one-dollar move in Company A, Company B has generally moved by a similar amount. This makes sense because the companies compete on similar routes and react to many of the same factors, including fuel costs and travel demand.

Over the past year, the average price ratio between A and B has remained close to 1, aside from normal daily noise.

Then a company-specific issue, such as a maintenance-related grounding or disappointing quarterly results, causes Company A to fall 8% in one week while Company B remains relatively flat. The price ratio and the spread’s z-score move well beyond their normal ranges.

At this point, the strategy buys Company A, which now appears statistically undervalued relative to B, and opens a short position in Company B. The position size is set to approximately balance the dollar exposure of the two legs.

This balancing process is what keeps the overall position close to market-neutral.

If the impact of the news does not fundamentally change the relationship between the two companies, the ratio may return toward its historical average over the following days or weeks. In that case, both positions are closed and a profit is realized on the spread.

However, if the divergence reflects a structural change, such as Company A permanently losing market share, the spread may never converge.

The position could then lose money, even if the performance of one leg partially offsets the other.

This point is often overlooked in overly simplistic explanations. The model suggests that convergence is probable. It never says that convergence is certain.

Statistical Arbitrage vs. Other Types of Arbitrage

People researching statistical arbitrage often start with a broader question about arbitrage in general. It is therefore useful to compare stat arb with the other main forms of arbitrage.

There are three broad categories.

Pure arbitrage, also known as risk-free arbitrage, matches the textbook definition. The same asset trades at two different prices in two different markets at the same time.

The trader buys at the lower price and simultaneously sells at the higher price, with no directional exposure and, at least in theory, no risk.

These discrepancies are rare and are usually eliminated within milliseconds by automated systems. As a result, pure arbitrage is not a realistic strategy for a discretionary trader.

Risk arbitrage, also known as merger arbitrage, is based on the outcome of a specific corporate event, typically a merger or acquisition.

In these situations, the target company’s stock trades below the announced deal price until the transaction either closes or falls through. The primary risk is that the deal will not be completed.

Statistical arbitrage, the focus of this article, involves neither a guaranteed price discrepancy nor a single binary event.

Instead, it relies on a statistical relationship that is expected to persist over time. This introduces model risk and the risk of correlation breakdown rather than pure execution risk.

The common thread is the attempt to exploit a mispricing. In that sense, the term “arbitrage” is appropriate. In stat arb, however, there is no guarantee.

Is Statistical Arbitrage Still Profitable?

The short answer is that the edge is real, but it has gradually weakened over the decades.

This has happened for a specific structural reason, not because the underlying logic has stopped working.

Statistical arbitrage was highly profitable for the funds that first adopted it in the 1980s and 1990s, partly because relatively few market participants were systematically searching for these relationships.

Any effective strategy, however, tends to attract capital. The more capital that targets the same inefficiencies, the faster those inefficiencies disappear.

This phenomenon is sometimes called alpha decay, or the deterioration of a strategy’s edge.

Stat arb is one of the clearest examples of alpha decay in modern markets. Simple pairs-trading setups that worked more reliably 20 years ago are now often crowded, low-margin trades.

The main advantage tends to go to market participants with faster execution and lower transaction costs.

This does not mean that this family of strategies is dead. It means that the version worth exploring today is different from the textbook version.

Traders need to:

• select a limited number of sound relationships instead of indiscriminately scanning every possible pair;
• account realistically for transaction costs and slippage before assuming that an edge found in a backtest will survive live execution;
• verify that the relationship still has an economic rationale, not merely a statistical one.

Two assets can be correlated by coincidence. A spread that converged during the backtest period is under no obligation to continue doing so in the future.

This is where a systematic approach demonstrates its value compared with a purely academic framework.

According to Andrea Unger, a four-time World Trading Champion who achieved his titles using systematic strategies, a trading system should be retired when its live results stop being consistent with its backtest. It should not be defended simply out of habit.

The specific relationship behind a stat arb pair can change, but the underlying discipline remains the same:

• test the idea rigorously;
• size the position so that a single mistake cannot cause permanent damage;
• stop trading the strategy when the data indicates that the edge has disappeared.

The fact that a strategy “worked for years” does not prove that it will continue to work.

That is precisely the alpha-decay trap that a systematic trader must monitor consistently.

The Real Risks of Statistical Arbitrage

Statistical arbitrage is often presented as safer because it is market-neutral. That is only partly true, and the distinction needs to be made clear.

The entire strategy is based on the assumption that a historical relationship will remain valid.

A regulatory change, a company-specific event, or a sector rotation can permanently break a relationship that appeared extremely robust in a backtest.

When this happens, the overall position can lose money if the loss on one leg exceeds the gain on the other. This creates the exact scenario that a market-neutral structure is intended to protect against.

Every stat arb strategy also involves model risk. It depends on several assumptions, including:

• how the spread is measured;
• which threshold triggers an entry;
• how long a position should be held before it is closed.

A model fitted too closely to historical data is subject to overfitting. It may look excellent in a backtest and then fail quickly under live market conditions.

This happens because the model learned the noise in the historical data rather than a genuine and repeatable relationship.

Execution costs also matter.

Stat arb strategies generally trade fairly frequently and hold positions for relatively short periods. Transaction costs, bid-ask spreads, and slippage can therefore erode returns faster than they would in a buy-and-hold approach.

A strategy that is profitable before costs can become unprofitable after costs. This is one of the most common gaps between backtest results and the performance of a live trading account.

Finally, there are crowding and liquidity risks.

When many funds run similar stat arb models, they may attempt to exit similar positions at the same time during a period of market stress. This can amplify losses across the board.

This dynamic played a visible role in the so-called Quant Quake of August 2007, when several large quantitative funds suffered simultaneous and correlated losses.

Being market-neutral does not protect you from the effects of other market-neutral traders selling at the same time under illiquid market conditions.

None of this means that the approach is necessarily reckless.

It means that “market-neutral” describes the strategy’s objective. It does not guarantee the outcome. The same discipline required for any systematic, rule-based approach must also be applied here.

Can an Individual Trader Use Statistical Arbitrage?

Yes, but only in a smaller and more realistic form, not through the large institutional baskets described earlier.

A retail trader who wants to explore statistical arbitrage without an advanced degree in mathematics or computer science should focus on the few elements that truly matter.

A beginner’s most common mistake is not necessarily choosing the wrong pair.

It is scanning hundreds of combinations until one displays a convincing correlation, then trusting it because the chart appears to confirm it. The trader never asks whether that correlation reflects a genuine economic relationship or merely a few months of coincidence.

A single well-understood pair, tested across a broad historical window and treated with caution if the correlation only emerged in recent months, is often preferable to an extremely broad scan.

The same principle applies to costs.

A backtest result that ignores commissions, bid-ask spreads, and slippage is not realistic. These costs also carry proportionally more weight in a smaller trading account than they do in a fund’s portfolio.

Position sizing is what actually provides protection when the relationship breaks down, not the elegance of the model.

This is why we at Unger Academy teach a systematic, rule-based approach in which position sizing and risk management play a central role, regardless of the strategy family being used.

Ultimately, statistical arbitrage is a useful framework for thinking about mean reversion and relative-value trades.

It is not a substitute for a diversified and thoroughly tested portfolio of systematic strategies. Treating it as the only tool in your trading toolbox means underestimating the risk of correlation breakdown.

The real barrier to entry is not a doctorate. It is the discipline required to test an idea correctly, without assuming that the word “arbitrage” will manage the risk for you.

This is the same principle behind building algorithmic trading strategies for specific assets: an idea must be rigorously tested before real capital is put at risk.

The same process of searching for reliable statistical relationships also applies to newer and more volatile markets, where correlations tend to be even less stable.

Frequently Asked Questions

Is Arbitrage Trading Legal?

Generally, yes, provided that the trader complies with applicable market rules and regulations.

Arbitrage, including statistical arbitrage, is a legal and widely used trading approach. It is employed by everyone from individual traders to some of the world’s largest hedge funds.

Certain illegal practices are sometimes confused with arbitrage, such as trading on material nonpublic information or manipulating prices to create an artificial discrepancy.

The strategy itself is not the issue.

How Do Traders Make Money With Statistical Arbitrage?

The most common example is pairs trading.

A trader identifies two historically correlated assets, takes a long position in the asset that has underperformed relative to the other, and takes a short position in the asset that has outperformed.

Both positions are then closed when the price relationship returns toward its historical mean. The practical example above illustrates the full process.

What Are the Three Types of Arbitrage?

Pure arbitrage, also known as risk-free arbitrage, exploits a simultaneous price discrepancy for the same asset in two different markets.

Risk arbitrage, also known as merger arbitrage, is based on the outcome of a specific corporate event, such as an acquisition.

Statistical arbitrage seeks to profit from the return toward the mean of a historical price relationship between different assets. There is no guaranteed price discrepancy or single binary event determining the outcome.

Is Statistical Arbitrage Still Profitable Today?

The edge has weakened since the strategy’s early years as increasing amounts of capital have eliminated many of the simplest opportunities. This phenomenon is known as alpha decay.

The strategy may remain viable for traders and funds willing to test relationships rigorously, account accurately for transaction costs, and retire strategies when the data shows that the edge has disappeared.

Traders should not expect exceptional, risk-free returns that are easily available to everyone.