struments. So for those instruments at least, all the models give the same price. But the risk management can be widely di\ufb00erent. We propose comparison on simple instruments (swaptions) on a simple risk measure (\ufb01rst order risk to the underlying yield curve). The hedging can vary widely (up to 10% of the underlying), specially with the dynamic of the model. The shape of the smile has also an impact but to a lesser extend.

The standard practice of option pricing is to calibrate the models to market prices of liquid options. Consequently all the models (with enough degree of freedom) will give the same price for standard options. Nevertheless, due to the di\ufb00erent dynamics implicit in the di\ufb00erent models, the risk inferred will be di\ufb00erent. In this note we study the risk from the view point of the change of price for small changes of the market rates, also called sensitivity to rates. We describe di\ufb00erent ways to compute it, not only using di\ufb00erent models but also di\ufb00erent ways to pass the market movements to the models.

There are certainly several places where studies of this kind are done (in particular Rebonato [7]). Here we study models widely used in practice and try to give actual \ufb01gures on the size of the impact.

We analyse several models. The \ufb01rst models are the classical Black model for swaption (geomet- ric Brownian motion of the forward swap rate) and its normal version(arithmetic Brownian motion of the forward swap rate). The next is a Heath-Jarrow-Morton model with Hull and White volatil- ity structure [4] (arithmetic Brownian motion of the continuously compounded rates) for which an explicit formula exists for European swaptions [2]. The last three are stochastic volatility models based on a constant elasticity volatility (CEV) extension of Black model, known SABR (stochastic, alpha, beta, rho)[1]. We use three versions of the model, one with the elasticity parameter\u03b2 equal to 0 (normal), one with 1 (log-normal) and one with no correlation between rates and volatility (\u03c1 = 0). For the Black, normal and Hull-White models we compute the theoretical in-the-model delta equivalent (amount of the underlying equivalent to the option to the \ufb01rst order) and use the sensitivity of the underlying multiply by this theoretical number. Moreover for the Black model, we compute the total delta and the one coming from the change of forward rate (no adjustment on the annuity or present value of a basis point). More details on the numerical procedures used are given in the Appendix.

of smile di\ufb00erence. For this we take two versions of a model with di\ufb00erent underlying dynamics, calibrate them (very closely) to the same smile and still obtain a quite di\ufb00erent delta. The model we use are the SABR models with elasticity parameter (\u03b2) equal to 0 and equal to 0.50.

As an example, we take a 1yx5y receiver at-the-money (ATM) swaption in USD on 25 June 2004. We use an ATM Black volatility of 23.5%, for the SABR model with 0 elasticity (\u03b2 = 0, normal-like model), we use (arbitrarily) a correlation (\u03c1) of 10% and a volatility of volatility (\u03bd) of 30%. The notional amount is 10 millions.

We calibrate the ATM price and the smile using the same SABR model but this time with an elasticity of 0.50 (between normal and log-normal model). For those models we compute the sensitivity to the market rates of the swaption. These sensitivities are reported in Table1 for di\ufb00erent tenors. The relative di\ufb00erence is computed with respect to the sensitivity of the underlying ATM forward swap. As can be seen from the numbers, the di\ufb00erence is close to 5%. The version of the model with elasticity 0 has the largest sensitivity (delta).

The impact of the dynamic depend on the swaption moneyness. We compute the sensitivities of receiver out-of-the-money (OTM) swaptions with di\ufb00erent strikes (1, 2, and 3% below and above forward rate). The total sensitivities are reported in Table3.

We have the same pattern with the 0 elasticity model giving a larger delta. The di\ufb00erence being between 0 and 5% of the risk of the underlying. Note that for ATM+3%, the sensitivity of the swaption is larger than the sensitivity of the forward ATM swap. The reason is that the swaption as a larger coupon and is so much in the money that its delta is almost 1.

As expected the sensitivity depends more on the underlying dynamics for ATM options than for OTM ones. For OTM options, the time-value, which is closely linked to the dynamic, is less important.

and volatility of volatility\u03bd). For the parameters we use ranges that, to our experience, are in line with market \ufb01gures. We use correlation between -10% and 70% and vol of vol between 20% and 50%. The results are given in Table4. We use as base case for comparisons the model used in the previous section with a correlation at 10% and a vol of vol at 30%.

In this section we study the dependency of the sensitivity to the choice of the model. For that we calibrate the models described in the introduction to an ATM swaption. For models with more freedom (SABR) we calibrate the smile at best to the smile as observe on 25 June 2004.

For the Black, normal, and Hull-White models we look at the sensitivity of thedelta equivalent. By this we mean that we compute the theoretical in-th-model delta of the swaption and multiply the sensitivity of the underlying swap by that amount. The results are detailed in Table5. Our

For the Balck model, we add the sensitivity when the Black volatility is changed1. Here we increase the volatility by 5%. With the change of volatility, the prices changes also. This delta is not included in the title of the note. It is a eleventh delta but with a second price. It is interesting to note that the change of sensitivity coming from the (very) wrong volatility is a lot smaller than the change of model. At the level of the table with the di\ufb00erent model sensitivities, it is (almost) impossible to distinguish the Balck model with the correct volatility from the one with the (very) wrong one.

It is interesting to note the following fact [6]. Suppose thecorrect model is the SABR 0 one. The Black model gives a completely di\ufb00erent dynamic. Correcting thesymptom of theincorrect model by introducing a smile through the SABR 1 model worsen the hedging results. The SABR 1 sensitivity (1,836) is further appart from the SABR 0 (2,260) than the simple Black (2,029).