Cointegration

Cointegration requires that each individual time series is integrated of order one, denoted as I(1). This means the series itself is non-stationary (possessing a unit root) and its variance and mean change over time, but its first difference is stationary. Common examples include asset prices, GDP, and interest rates. A stationary I(0) series cannot be cointegrated with another I(0) series in the traditional sense; cointegration specifically addresses the shared stochastic trends in non-stationary data.

The defining feature is the existence of a cointegrating vector that, when applied to the I(1) variables, produces a new series that is I(0) or stationary. This stationary residual represents the long-run equilibrium error. For example, if two stock prices are cointegrated, their price spread will be mean-reverting, fluctuating around a constant mean. This stationary combination implies that the variables cannot drift arbitrarily far apart over the long term.

Cointegrated variables share one or more common stochastic trends. If a system of n variables has r cointegrating relationships, it has n - r independent stochastic trends driving its non-stationary behavior. This is a key insight from the Granger Representation Theorem, which states that a cointegrated system can be represented as a Vector Error Correction Model (VECM). The shared trends explain why the variables move together over time.

Cointegration implies an error correction mechanism (ECM). This means that any short-term deviation from the long-run equilibrium will be corrected over time. The dynamics are modeled by a VECM, where the change in a variable depends on the lagged equilibrium error (the deviation) and past changes of all variables. This captures both the long-run equilibrium relationship and the short-run dynamic adjustments back toward it.

A classic approach for testing cointegration between two variables. The first step is to estimate the long-run equilibrium relationship using ordinary least squares (OLS). The second step tests the residuals from this regression for a unit root using an Augmented Dickey-Fuller (ADF) test. If the residuals are stationary, the series are cointegrated. This method is straightforward but assumes a single cointegrating vector and is sensitive to the choice of dependent variable.

A multivariate framework for testing cointegration among more than two variables. The Johansen test uses a Vector Autoregression (VAR) model and applies maximum likelihood estimation to determine the number of cointegrating relationships (r). It provides two test statistics: the trace test and the maximum eigenvalue test. Unlike the Engle-Granger method, it can identify multiple cointegrating vectors and treats all variables as endogenous.

The canonical application of cointegration. A trader identifies two cointegrated assets, such as Coca-Cola (KO) and PepsiCo (PEP) , and monitors the spread between their prices.

• When the spread widens beyond a historical threshold (e.g., 2 standard deviations), the strategy shorts the outperformer and goes long the underperformer.
• The position is closed when the spread reverts to its mean, capturing the convergence profit.
• This is a market-neutral strategy, as the long and short legs hedge broad market risk.

A generalized form of pairs trading applied to a large basket of securities. Instead of two assets, a synthetic asset is constructed from a portfolio of stocks that is cointegrated with a target asset or index.

• Principal Component Analysis (PCA) or the Johansen test is used to identify cointegrating vectors across hundreds of assets.
• The residual of the cointegrating regression is the trading signal. A mean-reverting residual triggers a trade.
• High-frequency execution is critical, as these arbitrage opportunities often exist for milliseconds.

The relationship between an Exchange-Traded Fund (ETF) and its underlying basket of stocks is inherently cointegrated. Authorized Participants (APs) exploit this.

• If the ETF price deviates from its Net Asset Value (NAV), an AP will simultaneously buy the cheaper asset and sell the richer one.
• For a premium, the AP buys the underlying basket and redeems it for new ETF shares to sell. For a discount, the process is reversed.
• This mechanism, driven by cointegration, keeps ETF prices tightly aligned with their NAV.

Many fundamental and quantitative hedge funds use cointegration to build dollar-neutral portfolios that are insulated from broad market moves.

• A fund might identify a cointegrated cluster of airline stocks. They go long the stocks with the most positive residual (undervalued) and short those with the most negative residual (overvalued).
• The portfolio's beta to the overall market is engineered to be near zero.
• Profit is generated purely from the idiosyncratic convergence of the stocks within the cluster, not from directional market exposure.

Cointegration is used to trade economic relationships between different commodities or their processing chain, known as crack spreads (oil to gasoline) or crush spreads (soybeans to meal and oil).

• For example, a trader models the cointegration between West Texas Intermediate (WTI) and Brent crude oil benchmarks.
• A temporary dislocation in the transatlantic spread, driven by a geopolitical event or pipeline disruption, triggers a mean-reversion trade.
• The trader goes long the temporarily cheap benchmark and short the expensive one, betting on the restoration of the historical equilibrium.

In foreign exchange, three currency pairs that share a common base and quote currency form a cointegrated system. Any deviation from the synthetic cross-rate is an arbitrage opportunity.

• Example: EUR/USD, USD/JPY, and EUR/JPY. The price of EUR/JPY must equal the product of the other two.
• Algorithmic traders monitor this relationship tick-by-tick. If a lag in one pair causes a mispricing, simultaneous trades lock in a risk-free profit.
• This enforces the no-arbitrage condition and keeps global FX markets consistent.