The Vector Error Correction Model

Two stock prices can each wander like a random walk on their own — no predictable direction, no mean to revert to — and yet the gap between them behaves completely differently: it wanders in the short run but never strays far before snapping back. That combination (individually unpredictable, jointly anchored) is cointegration, and it's the mathematical basis of pairs trading. But cointegration by itself only tells you the relationship exists; it doesn't tell you how fast the gap closes, or how today's gap should adjust tomorrow's prices. The vector error correction model (VECM) is the regression that answers exactly that: it says each series' next move is partly its own short-run wiggle, and partly a correction proportional to how far the pair currently sits from its long-run equilibrium.

An analogy: two dogs on a leash

Picture two dogs on a leash of fixed length, both wandering around a park. Each dog's individual path, watched alone, looks aimless — no destination, no pattern. But the leash means neither dog can wander arbitrarily far from the other: the moment the leash goes taut, there's a pull back. Watch closely and you'll see it's not perfectly symmetric — maybe the smaller dog gets yanked back harder and faster than the bigger one drags forward. That asymmetric pull, proportional to how taut the leash currently is, is exactly the "error correction" term. The VECM measures how taut the leash is right now (the equilibrium error) and how hard each dog gets pulled back per unit of tautness (the adjustment speed).

The model, one symbol at a time

Let $y_t$ and $x_t$ be two cointegrated price series (or a vector of several). First define the equilibrium error, the current gap from the long-run relationship:

$$z_{t-1} = y_{t-1} - \beta x_{t-1} .$$

In words: $z_{t-1}$ ("z sub t-minus-1") is how far apart the pair sat last period, after adjusting for the hedge ratio $\beta$ ("beta") that defines their long-run one-to-one relationship — this is the spread you'd trade. The VECM then models each series' change as:

$$\Delta y_t = \alpha_y\, z_{t-1} + \sum_j \phi_j \,\Delta y_{t-j} + \sum_j \psi_j\, \Delta x_{t-j} + \varepsilon_t .$$

In words: the change in $y$ this period equals an adjustment speed $\alpha_y$ ("alpha sub y") times last period's gap, plus the usual short-run dependence on recent changes in $y$ and $x$, plus noise. A parallel equation holds for $\Delta x_t$ with its own adjustment speed $\alpha_x$. The critical quantity is $\alpha_y$: if $\alpha_y$ is negative and $z_{t-1}$ was positive (meaning $y$ was too high relative to $x$), then $\alpha_y z_{t-1}$ pushes $\Delta y_t$ downward — $y$ falls back toward equilibrium. That's the "error correction": a term whose entire job is to shrink last period's disequilibrium.

Worked example 1: computing the correction term

Stock $y =$ $102, stock $x =$ $100, hedge ratio $\beta = 1.0$ from a prior cointegrating regression. The equilibrium error is $z_{t-1} = 102 - 1.0(100) = 2.0$ — $y$ is $2 rich relative to $x$. Suppose the fitted VECM gives $\alpha_y = -0.30$ and $\alpha_x = 0.10$ (with short-run terms set to zero for simplicity). Then:

$$\Delta y_t = -0.30 \times 2.0 = -0.60, \qquad \Delta x_t = 0.10 \times 2.0 = 0.20 .$$

So the model predicts $y$ falls by $0.60 and $x$ rises by $0.20 next period — the gap shrinks from $2.00 toward $2.00-0.60-0.20 = $1.20, about 60% closed in one step ($0.80/2.00$, since the two moves jointly close $0.30+0.10=0.40$ of the $1 hedge-adjusted unit... more directly: predicted new gap $= 1.20$, a reduction from 2.00, i.e. $0.80/2.00 = 40\%$ closed this period). This is a directly tradable prediction: go short $y$, long $\beta \times x$, and the VECM says the position should profit as the gap mean-reverts, at a rate set by $\alpha_y - \alpha_x = -0.40$ per unit of gap per period.

Worked example 2: half-life from the adjustment speed

If short-run terms are negligible, the gap itself roughly follows $z_t \approx (1 + \alpha_y - \alpha_x)\, z_{t-1}$, an AR(1)-like decay with persistence $\lambda = 1 + \alpha_y - \alpha_x = 1 - 0.30 - 0.10 = 0.60$. The half-life of the mean reversion is

$$\text{half-life} = \frac{\ln(0.5)}{\ln(\lambda)} = \frac{\ln(0.5)}{\ln(0.60)} = \frac{-0.693}{-0.511} = 1.36 \text{ periods} .$$

In words: starting from a $2.00 gap, you'd expect it to shrink to $1.00 in about 1.36 periods, matching the arithmetic above (gap fell from 2.00 to 1.20, about 60% of the way to zero, faster than one full half-life — consistent, since 1.36 periods should get you to exactly half). This half-life is the single most useful number for sizing a pairs trade's holding period: if daily data gives a half-life of 1.4 days, you're running a fast mean-reversion strategy; a half-life of 40 days means much slower, longer-held convergence bets with correspondingly different risk.

Watch how the running quantity in the explorer settles toward a stable value as more steps accumulate — the equilibrium error in a VECM behaves the same way, oscillating but pulled back, rather than drifting off freely like the individual series do.

What this means in practice

• This is pairs trading's engine. The hedge ratio $\beta$ sets the position sizes; the equilibrium error $z_t$ is the entry/exit signal; the adjustment speed $\alpha$ tells you the expected holding period and, combined with the spread's volatility, the expected profit per trade.
• You need cointegration first. A VECM only makes sense if $y$ and $x$ are actually cointegrated — running one on two unrelated random walks manufactures the appearance of mean reversion out of noise, a specific case of spurious regression.
• Adjustment speed can be asymmetric or one-sided. Sometimes only one series does the correcting ($\alpha_x \approx 0$) — the "weakly exogenous" series leads, the other follows, which matters for deciding which leg to trade more aggressively.
• Compare to a simple OLS spread model. A VECM is strictly richer than just trading the residual from a static cointegrating regression, because it also models the short-run dynamics ($\phi_j, \psi_j$) that a plain spread z-score ignores.