The Vasicek Model

Short-term interest rates don't wander off to infinity or drop through the floor forever — they hover in a range that reflects a central bank's target and the broader economy, occasionally spiking or dipping before drifting back. A model of rates that ignores this tendency (like plain Brownian motion) will eventually predict absurd things: rates of −40% or +900%. The Vasicek model was the first widely used fix: it lets the short rate move randomly, tick by tick, but always with a gentle pull back toward a long-run average.

An analogy: a rate on a bungee cord

Imagine the short rate as a ball on a bungee cord anchored at some long-run level, say 3%. Random gusts of wind (economic shocks) push the ball around, but the further it drifts from the anchor, the harder the cord pulls it back. A big gust can still send the ball far from 3% temporarily, but over time it keeps snapping back toward the anchor rather than drifting away for good. The strength of the cord (how fast it snaps back) and the strength of the wind (how volatile rates are) together determine how the ball dances around its resting point.

The mathematics, one symbol at a time

The Vasicek model describes the short rate $r_t$ with the stochastic differential equation

$$dr_t = \kappa(\theta - r_t)\, dt + \sigma\, dW_t,$$

where $\theta$ is the long-run mean rate (the anchor), $\kappa > 0$ is the speed of mean reversion (the cord's stiffness), $\sigma$ is the volatility (the wind strength), and $dW_t$ is the random Brownian shock. In words: the rate's instantaneous drift is proportional to how far it currently sits from $\theta$ — the further away, the stronger the pull back — plus a constant amount of random noise on top. This is exactly the Ornstein-Uhlenbeck process (see The Ornstein-Uhlenbeck Process) applied to interest rates. Because the model is linear and Gaussian, the rate at any future time is normally distributed, with a known mean and variance that both converge to fixed limits as time grows — which is what makes Vasicek tractable enough to price bonds in closed form, at the cost of allowing (rare) negative rates.

Worked example 1: where the rate is heading

Suppose $\theta = 3\%$, $\kappa = 0.5$ (per year), and the rate currently sits at $r_0 = 6\%$. The expected rate at time $t$ is

$$\mathbb{E}[r_t] = \theta + (r_0 - \theta)e^{-\kappa t}.$$

At $t=1$ year: $\mathbb{E}[r_1] = 3 + (6-3)e^{-0.5} = 3 + 3(0.6065) \approx 4.82\%$. At $t=3$ years: $3 + 3e^{-1.5} \approx 3 + 3(0.223) \approx 3.67\%$. The expected path decays exponentially from 6% back toward 3%, halving the gap roughly every 1.4 years — it never expects to overshoot, just to relax back.

Worked example 2: how uncertain is that forecast?

The variance of $r_t$ around that expected path is $\mathrm{Var}(r_t) = \frac{\sigma^2}{2\kappa}\left(1 - e^{-2\kappa t}\right)$. With $\sigma = 1.5\%$ and $\kappa = 0.5$: as $t \to \infty$, $\mathrm{Var}(r_\infty) = \frac{1.5^2}{2(0.5)} = 2.25$, so the long-run standard deviation of the rate is $\sqrt{2.25} = 1.5$ percentage points. That means even far in the future, the rate is expected to sit within roughly $3\% \pm 1.5\%$ about two-thirds of the time — mean reversion caps how spread out the distribution can ever get, unlike ordinary Brownian motion where variance grows without bound.

Drag the mean-reversion speed above: turn it up and paths snap back to the anchor almost immediately; turn it down toward zero and the process starts to look like plain, unanchored Brownian motion.

What this means in practice

Vasicek is the textbook entry point to affine term-structure models (see Affine Term Structure Models): because rates are Gaussian, bond prices and yields come out as simple exponential-affine functions of the current rate, letting a whole yield curve be generated from three parameters. Trading desks use mean-reversion models like this to size how much a rates position should be discounted for "it'll likely revert" versus a pure random-walk assumption, and to build intuition for swap curve dynamics (see Interest Rate Swaps).