Finance:Hull–White model

From HandWiki
Short description: Model of future interest rates

In financial mathematics, the Hull–White model is a model of future interest rates. In its most generic formulation, it belongs to the class of no-arbitrage models that are able to fit today's term structure of interest rates. It is relatively straightforward to translate the mathematical description of the evolution of future interest rates onto a tree or lattice and so interest rate derivatives such as bermudan swaptions can be valued in the model.

The first Hull–White model was described by John C. Hull and Alan White in 1990. The model is still popular in the market today.

The model

One-factor model

The model is a short-rate model. In general, it has the following dynamics:

[math]\displaystyle{ dr(t) = \left[\theta(t) - \alpha(t) r(t)\right]\,dt + \sigma(t)\, dW(t). }[/math]

There is a degree of ambiguity among practitioners about exactly which parameters in the model are time-dependent or what name to apply to the model in each case. The most commonly accepted naming convention is the following:

  • [math]\displaystyle{ \theta }[/math] has t (time) dependence — the Hull–White model.
  • [math]\displaystyle{ \theta }[/math] and [math]\displaystyle{ \alpha }[/math] are both time-dependent — the extended Vasicek model.

Two-factor model

The two-factor Hull–White model (Hull 2006:657–658) contains an additional disturbance term whose mean reverts to zero, and is of the form:

[math]\displaystyle{ d\,f(r(t)) = \left [\theta(t) + u - \alpha(t)\,f(r(t))\right ]dt + \sigma_1(t)\, dW_1(t), }[/math]

where [math]\displaystyle{ \displaystyle f }[/math] is a deterministic function, typically the identity function (extension of the one-factor version, analytically tractable, and with potentially negative rates), the natural logarithm (extension of Black–Karasinksi, not analytically tractable, and with positive interest rates), or combinations (proportional to the natural logarithm on small rates and proportional to the identity function on large rates); and [math]\displaystyle{ \displaystyle u }[/math] has an initial value of 0 and follows the process:

[math]\displaystyle{ du = -bu\,dt + \sigma_2\,dW_2(t) }[/math]

Analysis of the one-factor model

For the rest of this article we assume only [math]\displaystyle{ \theta }[/math] has t-dependence. Neglecting the stochastic term for a moment, notice that for [math]\displaystyle{ \alpha \gt 0 }[/math] the change in r is negative if r is currently "large" (greater than [math]\displaystyle{ \theta(t)/\alpha) }[/math] and positive if the current value is small. That is, the stochastic process is a mean-reverting Ornstein–Uhlenbeck process.

θ is calculated from the initial yield curve describing the current term structure of interest rates. Typically α is left as a user input (for example it may be estimated from historical data). σ is determined via calibration to a set of caplets and swaptions readily tradeable in the market.

When [math]\displaystyle{ \alpha }[/math], [math]\displaystyle{ \theta }[/math], and [math]\displaystyle{ \sigma }[/math] are constant, Itô's lemma can be used to prove that

[math]\displaystyle{ r(t) = e^{-\alpha t}r(0) + \frac{\theta}{\alpha} \left(1- e^{-\alpha t}\right) + \sigma e^{-\alpha t}\int_0^t e^{\alpha u}\,dW(u), }[/math]

which has distribution

[math]\displaystyle{ r(t) \sim \mathcal{N}\left(e^{-\alpha t} r(0) + \frac{\theta}{\alpha} \left(1- e^{-\alpha t}\right), \frac{\sigma^2}{2\alpha} \left(1-e^{-2\alpha t}\right)\right), }[/math]

where [math]\displaystyle{ \mathcal{N}( \mu ,\sigma^2 ) }[/math] is the normal distribution with mean [math]\displaystyle{ \mu }[/math] and variance [math]\displaystyle{ \sigma^2 }[/math].

When [math]\displaystyle{ \theta(t) }[/math] is time-dependent,

[math]\displaystyle{ r(t) = e^{-\alpha t}r(0) + \int_{0}^{t}e^{\alpha(s-t)}\theta(s)ds + \sigma e^{-\alpha t}\int_0^t e^{\alpha u}\,dW(u), }[/math]

which has distribution

[math]\displaystyle{ r(t) \sim \mathcal{N}\left(e^{-\alpha t} r(0) + \int_{0}^{t}e^{\alpha(s-t)}\theta(s)ds, \frac{\sigma^2}{2\alpha} \left(1-e^{-2\alpha t}\right)\right). }[/math]

Bond pricing using the Hull–White model

It turns out that the time-S value of the T-maturity discount bond has distribution (note the affine term structure here!)

[math]\displaystyle{ P(S,T) = A(S,T)\exp(-B(S,T)r(S)), }[/math]

where

[math]\displaystyle{ B(S,T) = \frac{1-\exp(-\alpha(T-S))}{\alpha} , }[/math]
[math]\displaystyle{ A(S,T) = \frac{P(0,T)}{P(0,S)}\exp\left( \, -B(S,T) \frac{\partial\log(P(0,S))}{\partial S} - \frac{\sigma^2(\exp(-\alpha T)-\exp(-\alpha S))^2(\exp(2\alpha S)-1)}{4\alpha^3}\right) . }[/math]

Note that their terminal distribution for [math]\displaystyle{ P(S,T) }[/math] is distributed log-normally.

Derivative pricing

By selecting as numeraire the time-S bond (which corresponds to switching to the S-forward measure), we have from the fundamental theorem of arbitrage-free pricing, the value at time t of a derivative which has payoff at time S.

[math]\displaystyle{ V(t) = P(t,S)\mathbb{E}_S[V(S) \mid \mathcal{F}(t)]. }[/math]

Here, [math]\displaystyle{ \mathbb{E}_S }[/math] is the expectation taken with respect to the forward measure. Moreover, standard arbitrage arguments show that the time T forward price [math]\displaystyle{ F_V(t,T) }[/math] for a payoff at time T given by V(T) must satisfy [math]\displaystyle{ F_V(t,T) = V(t)/P(t,T) }[/math], thus

[math]\displaystyle{ F_V(t,T) = \mathbb{E}_T[V(T)\mid\mathcal{F}(t)]. }[/math]

Thus it is possible to value many derivatives V dependent solely on a single bond [math]\displaystyle{ P(S,T) }[/math] analytically when working in the Hull–White model. For example, in the case of a bond put

[math]\displaystyle{ V(S) = (K-P(S,T))^+. }[/math]

Because [math]\displaystyle{ P(S,T) }[/math] is lognormally distributed, the general calculation used for the Black–Scholes model shows that

[math]\displaystyle{ {E}_S[(K-P(S,T))^{+}] = KN(-d_2) - F(t,S,T)N(-d_1), }[/math]

where

[math]\displaystyle{ d_1 = \frac{\log(F/K) + \sigma_P^2S/2}{\sigma_P \sqrt{S}} }[/math]

and

[math]\displaystyle{ d_2 = d_1 - \sigma_P \sqrt{S}. }[/math]

Thus today's value (with the P(0,S) multiplied back in and t set to 0) is:

[math]\displaystyle{ P(0,S)KN(-d_2) - P(0,T)N(-d_1). }[/math]

Here [math]\displaystyle{ \sigma_P }[/math] is the standard deviation (relative volatility) of the log-normal distribution for [math]\displaystyle{ P(S,T) }[/math]. A fairly substantial amount of algebra shows that it is related to the original parameters via

[math]\displaystyle{ \sqrt{S}\sigma_P =\frac{\sigma}{\alpha}(1-\exp(-\alpha(T-S)))\sqrt{\frac{1-\exp(-2\alpha S)}{2\alpha}}. }[/math]

Note that this expectation was done in the S-bond measure, whereas we did not specify a measure at all for the original Hull–White process. This does not matter — the volatility is all that matters and is measure-independent.

Because interest rate caps/floors are equivalent to bond puts and calls respectively, the above analysis shows that caps and floors can be priced analytically in the Hull–White model. Jamshidian's trick applies to Hull–White (as today's value of a swaption in the Hull–White model is a monotonic function of today's short rate). Thus knowing how to price caps is also sufficient for pricing swaptions. In the even that the underlying is a compounded backward-looking rate rather than a (forward-looking) LIBOR term rate, Turfus (2020) shows how this formula can be straightforwardly modified to take into account the additional convexity.

Swaptions can also be priced directly as described in Henrard (2003). Direct implementations are usually more efficient.

Monte-Carlo simulation, trees and lattices

However, valuing vanilla instruments such as caps and swaptions is useful primarily for calibration. The real use of the model is to value somewhat more exotic derivatives such as bermudan swaptions on a lattice, or other derivatives in a multi-currency context such as Quanto Constant Maturity Swaps, as explained for example in Brigo and Mercurio (2001). The efficient and exact Monte-Carlo simulation of the Hull–White model with time dependent parameters can be easily performed, see Ostrovski (2013) and (2016). An open-source implementation of the exact Monte-Carlo simulation following Fries (2016)[1] can be found in finmath lib.[2]


Forecasting

Even though single factor models such as Vasicek, CIR and Hull–White model has been devised for pricing, recent research has shown their potential with regard to forecasting. In Orlando et al. (2018,[3] 2019,[4][5]) was provided a new methodology to forecast future interest rates called CIR#. The ideas, apart from turning a short-rate model used for pricing into a forecasting tool, lies in an appropriate partitioning of the dataset into subgroups according to a given distribution [6]). In there it was shown how the said partitioning enables capturing statistically significant time changes in volatility of interest rates. following the said approach, Orlando et al. (2021) [7]) compares the Hull–White model with the CIR model in terms of forecasting and prediction of interest rate directionality.

See also

References

  1. Fries, Christian (2016). "A Short Note on the Exact Stochastic Simulation Scheme of the Hull-White Model and Its Implementation". SSRN. doi:10.2139/ssrn.2737091. https://ssrn.com/abstract=2737091. Retrieved October 15, 2023. 
  2. "HullWhiteModel.java". finmath.net. https://github.com/finmath/finmath-lib/blob/master/src/main/java/net/finmath/montecarlo/interestrate/models/HullWhiteModel.java. 
  3. Orlando, Giuseppe; Mininni, Rosa Maria; Bufalo, Michele (2018). "A New Approach to CIR Short-Term Rates Modelling". New Methods in Fixed Income Modeling. Contributions to Management Science (Springer International Publishing): 35–43. doi:10.1007/978-3-319-95285-7_2. ISBN 978-3-319-95284-0. 
  4. Orlando, Giuseppe; Mininni, Rosa Maria; Bufalo, Michele (1 January 2019). "A new approach to forecast market interest rates through the CIR model". Studies in Economics and Finance 37 (2): 267–292. doi:10.1108/SEF-03-2019-0116. ISSN 1086-7376. 
  5. Orlando, Giuseppe; Mininni, Rosa Maria; Bufalo, Michele (19 August 2019). "Interest rates calibration with a CIR model" (in en). The Journal of Risk Finance 20 (4): 370–387. doi:10.1108/JRF-05-2019-0080. ISSN 1526-5943. 
  6. Orlando, Giuseppe; Mininni, Rosa Maria; Bufalo, Michele (July 2020). "Forecasting interest rates through Vasicek and CIR models: A partitioning approach" (in en). Journal of Forecasting 39 (4): 569–579. doi:10.1002/for.2642. ISSN 0277-6693. https://onlinelibrary.wiley.com/doi/10.1002/for.2642. 
  7. Orlando, Giuseppe; Bufalo, Michele (2021-05-26). "Interest rates forecasting: Between Hull and White and the CIR#—How to make a single‐factor model work" (in en). Journal of Forecasting 40 (8): 1566–1580. doi:10.1002/for.2783. ISSN 0277-6693. 
Primary references
  • John Hull and Alan White, "Using Hull–White interest rate trees," Journal of Derivatives, Vol. 3, No. 3 (Spring 1996), pp. 26–36
  • John Hull and Alan White, "Numerical procedures for implementing term structure models I," Journal of Derivatives, Fall 1994, pp. 7–16.
  • John Hull and Alan White, "Numerical procedures for implementing term structure models II," Journal of Derivatives, Winter 1994, pp. 37–48.
  • John Hull and Alan White, "The pricing of options on interest rate caps and floors using the Hull–White model" in Advanced Strategies in Financial Risk Management, Chapter 4, pp. 59–67.
  • John Hull and Alan White, "One factor interest rate models and the valuation of interest rate derivative securities," Journal of Financial and Quantitative Analysis, Vol 28, No 2, (June 1993) pp. 235–254.
  • John Hull and Alan White, "Pricing interest-rate derivative securities", The Review of Financial Studies, Vol 3, No. 4 (1990) pp. 573–592.
Other references