Jorge Quitério Search
Mathematics

Mean reversion in interest rates: the Vasicek model

Interest rates wander but tend to return towards a long-run level. The Vasicek model describes this with a mean-reverting stochastic process, which is an AR(1) model in continuous time, gives bond prices and the yield curve in closed form, and allows negative rates, as the euro area experienced between 2014 and 2022.

A household choosing a mortgage faces a forecasting problem, whether or not it describes it that way. A variable rate follows the short-term interest rate set by the market and the central bank; a fixed rate locks in today's long-term rate for ten or twenty years. Which is cheaper depends on where short-term rates will go. A pension fund faces the same question on a larger scale, because the value of its future obligations depends on the rates at which they are discounted.

Short-term rates have a characteristic behaviour. They move unpredictably from month to month, but they do not drift without limit: periods of very high or very low rates have tended to end, and rates have moved back towards a more usual range. The euro area provides a recent example at the low end. The European Central Bank's deposit facility rate became negative in June 2014, reached −0.50% in September 2019 and returned to zero only in July 2022, after which it rose quickly.

The simplest stochastic model of this behaviour is due to Oldřich Vasicek. This article presents it, shows that it is the continuous-time version of the autoregressive model, and derives the shape of the yield curve it implies.

The relevant literature is as follows.

  • Uhlenbeck and Ornstein (1930) introduced the mean-reverting process now named after them to describe the velocity of a particle subject to friction and random collisions.
  • Vasicek (1977) used it as a model of the short-term interest rate and derived the prices of bonds of all maturities in closed form.
  • Cox, Ingersoll and Ross (1985) proposed a variant in which the volatility grows with the square root of the rate, which keeps rates non-negative under a condition on the parameters.
  • Estrella and Mishkin (1998) found the spread between long and short yields to be one of the most useful single predictors of US recessions.
  • Brigo and Mercurio (2006) is a standard reference for the many interest-rate models built on these foundations.

The model

Vasicek describes the short rate rtr_t by

drt=κ (θ−rt) dt+σ dWt. dr_t = \kappa\,(\theta - r_t)\, dt + \sigma\, dW_t .

The first term pulls the rate towards the long-run level θ\theta at speed κ\kappa; the second adds random shocks with volatility σ\sigma. Given today's rate r0r_0, the rate at time tt is normal with

E[rt]=θ+(r0−θ) e−κt,Var⁡(rt)=σ22κ(1−e−2κt). E[r_t] = \theta + (r_0 - \theta)\, e^{-\kappa t}, \qquad \operatorname{Var}(r_t) = \frac{\sigma^2}{2\kappa}\big(1 - e^{-2\kappa t}\big).

A deviation from the long-run level halves every ln⁡2/κ\ln 2/\kappa years. With κ=0.15\kappa = 0.15 that is about 4.6 years. As tt grows, the distribution settles at a normal with mean θ\theta and standard deviation σ/2κ\sigma/\sqrt{2\kappa}: with σ=1%\sigma = 1\% a year, about 1.8 percentage points.

The connection with the AR(1) model

Observed once a month, the Vasicek rate follows exactly an AR(1) process:

rt+1/12=θ+e−κ/12 (rt−θ)+εt, r_{t+1/12} = \theta + e^{-\kappa/12}\,(r_t - \theta) + \varepsilon_t,

with independent normal shocks. Everything in the autoregressive model of a time series applies: the coefficient e−κ/12e^{-\kappa/12} is the memory, and the half-life follows from it. It also explains a familiar difficulty. With κ=0.15\kappa = 0.15, the monthly coefficient is about 0.988, so close to 1 that standard unit-root tests, as discussed in least-squares estimation of the AR(1) coefficient in three regimes, often cannot distinguish an interest rate from a random walk with a few decades of data. Whether interest rates are mean-reverting at all has therefore been debated for as long as the data have been studied.

Negative rates

Because the Vasicek rate is normal, it can become negative. For many years this was regarded as a defect of the model, and the Cox-Ingersoll-Ross model was often preferred because it rules negative rates out. The euro area, Switzerland, Denmark, Sweden and Japan then operated with negative policy rates for several years in the 2010s, and the defect became a feature. With today's rate at zero, a long-run level of 3% and the parameters above, the model gives about a 9.5% probability that the rate is negative ten years from now.

The yield curve

Vasicek showed that, with no risk premium, a zero-coupon bond maturing in τ\tau years costs

P(τ)=exp⁡(A(τ)−B(τ) r0),B(τ)=1−e−κτκ, P(\tau) = \exp\big(A(\tau) - B(\tau)\, r_0\big), \qquad B(\tau) = \frac{1 - e^{-\kappa\tau}}{\kappa},

with A(τ)=(θ−σ22κ2)(B(τ)−τ)−σ2B(τ)24κA(\tau) = \big(\theta - \tfrac{\sigma^2}{2\kappa^2}\big)\big(B(\tau) - \tau\big) - \tfrac{\sigma^2 B(\tau)^2}{4\kappa}, and its yield is y(τ)=−ln⁡P(τ)/τy(\tau) = -\ln P(\tau)/\tau. The yield curve has a direct interpretation: long yields average the expected future short rates, minus a small convexity term that grows with volatility.

When today's rate is below its long-run level, short rates are expected to rise and the yield curve slopes upwards. With r0=0r_0 = 0, the parameters above give yields of about 0.2% at one year, 1.4% at ten years and 2.2% at thirty years. When today's rate is above its long-run level, rates are expected to fall and the curve is inverted: with r0=6%r_0 = 6\%, about 5.8%, 4.5% and 3.5%. For the household choosing a mortgage, the model therefore says that a fixed rate above today's variable rate is not in itself a sign of a bad deal; it reflects the expected rise of short rates over the life of the loan.

An inverted curve signals that short rates are expected to fall, which usually happens when a central bank cuts rates in a downturn. In the United States, an inversion of the yield curve has preceded each recession since the late 1960s, which is why the spread between long and short yields is closely watched. How recessions themselves can be modelled is the subject of Hamilton's regime-switching model.

The figure below simulates the model and draws the yield curve it implies. Rates fluctuate around θ with a half-life of ln 2/κ; the long-run spread is σ/√(2κ); the curve slopes upwards when today's rate is below θ and is inverted when it is above.

Suggested settings:

  1. Low rates today. With the defaults (κ = 0.15, θ = 3%, σ = 1%, r₀ = 0), the expected path rises towards 3% and the band widens to about ±3.6 points. The yield curve slopes upwards, and the probability of a negative rate in ten years is about 9.5%.
  2. High rates today. Press "High rates today (r₀ = 6%)". The expected path falls, and the yield curve is inverted: one-year yields exceed thirty-year yields.
  3. Speed of reversion. Lower κ to 0.03. The half-life grows to about 23 years, the band keeps widening for decades, and the monthly AR coefficient is 0.9975: practically a random walk on the horizon of a mortgage. Raise κ to 1 and deviations disappear within a year.
  4. Volatility and negative rates. With r₀ at −0.5%, raise σ to 2%. The probability of a negative rate in ten years rises sharply, and the convexity term pulls the long end of the curve down.

Derivation of the distribution. Multiply the equation by eκte^{\kappa t}: d(eκtrt)=κθeκt dt+σeκt dWtd\big(e^{\kappa t} r_t\big) = \kappa\theta e^{\kappa t}\, dt + \sigma e^{\kappa t}\, dW_t. Integrating from 0 to tt gives rt=θ+(r0−θ)e−κt+σ∫0te−κ(t−s) dWsr_t = \theta + (r_0 - \theta)e^{-\kappa t} + \sigma\int_0^t e^{-\kappa(t-s)}\, dW_s. The stochastic integral is normal with mean zero and, by the Itô isometry, variance σ2∫0te−2κ(t−s) ds=σ2(1−e−2κt)/(2κ)\sigma^2\int_0^t e^{-2\kappa(t-s)}\, ds = \sigma^2(1 - e^{-2\kappa t})/(2\kappa). Taking t=1/12t = 1/12 gives the AR(1) form above, with shock variance σ2(1−e−κ/6)/(2κ)\sigma^2(1 - e^{-\kappa/6})/(2\kappa).

The Vasicek model is too simple for a trading desk, but it states clearly three properties that more elaborate models retain:

  • Interest rates revert towards a long-run level, slowly enough that decades of data are needed to measure the speed.
  • The yield curve is largely an expected path of future short rates, which explains why it slopes upwards when rates are low and inverts when they are high.
  • Negative rates are a possibility to be modelled, not excluded, as the euro area showed between 2014 and 2022.

References

  1. G. E. Uhlenbeck and L. S. Ornstein, "On the theory of the Brownian motion", Physical Review, 36(5), 823-841, 1930.
  2. O. Vasicek, "An equilibrium characterization of the term structure", Journal of Financial Economics, 5(2), 177-188, 1977.
  3. J. C. Cox, J. E. Ingersoll and S. A. Ross, "A theory of the term structure of interest rates", Econometrica, 53(2), 385-407, 1985.
  4. A. Estrella and F. S. Mishkin, "Predicting U.S. recessions: financial variables as leading indicators", Review of Economics and Statistics, 80(1), 45-61, 1998.
  5. D. Brigo and F. Mercurio, Interest Rate Models: Theory and Practice, 2nd edition, Springer, 2006.