RISK NEUTRAL VERSUS REAL WORLD VALUATION Pieter de Boer July 2009 Zanders T +31 35 692 89 89 Postbus 221 Brinklaan 134 1400 AE Bussum 1404 GV Bussum F +31 35 692 89 99 E info@zanders.eu Risk neutral versus real world valuation Page 2 of 38 Contents 1 2 Introduction ............................................................................................................3 1.1 Introduction ....................................................................................................... 3 1.2 Objective ........................................................................................................... 3 1.3 Approach ........................................................................................................... 4 1.4 Structure of this report ........................................................................................ 4 Risk neutral valuation ..............................................................................................5 2.1 What is risk neutral valuation? .............................................................................. 5 2.2 An intuitive justification of risk neutral valuation ...................................................... 5 2.3 Problems using real world valuation ....................................................................... 7 2.4 Risk neutral valuation in a continuous framework ..................................................... 7 2.5 From a real world to a risk neutral measure............................................................. 9 3 Guaranteed product ............................................................................................... 11 4 Black-Scholes-Hull-White model............................................................................. 12 5 4.1 Hull-White model .............................................................................................. 12 4.2 Black-Scholes model.......................................................................................... 14 4.3 Correlation between the interest rate and the stock price ......................................... 14 4.4 BSHW model .................................................................................................... 16 Results of the risk neutral valuation of the product ................................................. 19 5.1 6 7 Results ............................................................................................................ 19 The stochastic discount factor ................................................................................ 22 6.1 Theory on the stochastic discount factor ............................................................... 22 6.2 The stochastic discount factor in the BSHW model .................................................. 23 Results of the real world valuation of the product ................................................... 27 7.1 Results ............................................................................................................ 27 7.2 Backtest .......................................................................................................... 30 8 Summary and conclusions ...................................................................................... 34 9 References ............................................................................................................ 36 Appendix explanation used data .................................................................................. 37 Risk neutral versus real world valuation Page 3 of 38 1 1.1 Introduction Introduction The importance of market consistent valuation (MCV) has risen in recent years throughout the global financial industry (Gibbs & McNamara, 2007). This is due to the new regulatory landscape1 and because banks and insurers see the necessity of better understanding of the uncertainty in the market value of their balance sheet. Part of these balance sheets on the liability side often contain guaranteed contracts, which are difficult to value with standard valuation techniques due to the optionality and nonlinearity embedded in the product. Valuation of such guaranteed contracts at the present time uses the concept of risk neutral valuation, for example in Nielsen & Sandmann (1995). However, determining the uncertainty in the future value, for example needed for regulatory or economic capital calculations, of such products is more difficult, because future outcomes are not simulated correctly. For example, when using risk neutral simulations, stock prices are assumed to grow with the risk free interest rate, which is not realistic. Using real world simulations, future outcomes are correctly simulated, since the stock prices grow at the actual expected return (the risk free rate combined with a risk premium). However, the valuation of a product using a standard (risk neutral) discount factor is inconsistent, since the returns are not risk neutral but the discount factor is. Therefore, a combination of these two methods is needed to be able to correctly simulate future outcomes, value the (guaranteed) products and determine the uncertainty in it‟s future value. Real world simulations are needed to correctly simulate future values of the variables and a proper (stochastic) discount factor is needed to value the products in the future correctly. In this thesis the focus will be on the MCV of an equity-linked guaranteed contract. This product is a combination of a savings product and investment product. It contains a guaranteed return, a part of the 1-month Euribor interest rate, and the possibility of an extra investment return if the underlying index performs above a contractually determined level. 1.2 Objective The objective of this thesis is: Combine risk neutral valuation and real world simulations to correctly measure the uncertainty in future risk neutral value. This objective is achieved by answering the following questions: 1 For example, Basel II, IFRS and Solvency II, which forces banks and insurers to report the values and risks of products based on a market consistent valuation method. Risk neutral versus real world valuation Page 4 of 38 1. What is risk neutral valuation and why does valuation based on real world simulations not result in a correct future value? 2. What is the risk neutral value of the equity linked guaranteed contract at the present time? 3. What is the uncertainty of this risk neutral value in the future, i.e. one year? 1.3 Approach This thesis determines the risk neutral value of the product at the present time and measures the uncertainty in the future risk neutral value. For the risk neutral valuation at the present time we use a stochastic simulation model. This stochastic simulation model assumes a geometric Brownian motion for the stock prices, just as in a Black-Scholes model (Black & Scholes, 1973), where the volatility of the stock price is obtained by extracting the implied at the money (ATM) volatility of an option on the stock. Furthermore, a one factor Hull-White model (Hull & White, 1990) is used to simulate the interest rate. This arbitrage free interest rate model is entirely determined by the current state of the market. The parameters will be determined by the concept of calibration. This stochastic simulation model is estimated in a risk neutral setting. The next step is to combine risk neutral valuation and real world simulations to correctly measure the uncertainty in future risk neutral value. In order to do so, real world simulations, by assuming realistic returns, are used instead of risk neutral simulations, in the same setting as described above. Since an appropriate discount factor is not directly available when real world simulations are used, it‟s necessary to find a stochastic discount factor or “deflator” consistent with the real world simulations. Using this stochastic discount factor, a valuation method equivalent2 to risk neutral valuation can be used to determine the risk neutral value of the product at the present time and measure the uncertainty in the future value. 1.4 Structure of this report In the next chapter, the concept of risk neutral valuation is explained. Chapter 3 describes the chosen guaranteed contract in detail. Chapter 4 describes the Black-Scholes-Hull-White (BSHW) model used for the risk neutral valuation at the present time. The results of the BSHW-model are presented in chapter 5. Chapter 6 describes the properties of deflators and their usefulness in correctly measuring the uncertainty in future risk neutral value. This chapter describes the incorporation of deflators in the BSHW model. Chapter 7 contains the outcomes of the model described in the preceding chapter. Finally, chapter 9 gives a conclusion. The appendix discusses the way the parameters are obtained. 2 I.e., not exactly the same, since the calculations are different, but the outcomes are identical for the numbers of simulations approaching infinity. Risk neutral versus real world valuation Page 5 of 38 2 Risk neutral valuation This chapter explains the methodology of risk neutral valuation. First, section 2.1 explains what risk neutral valuation is, then a justification is given in section 2.2, illustrated by a simple example. This example is also used to clarify why real world valuation cannot be used in a straightforward way in section 2.3. Furthermore, a simple derivation of risk neutral valuation in a continuous framework is given in section 2.4, as well as a short elaboration on changing from a real world to a risk neutral probability measure in section 2.5. 2.1 What is risk neutral valuation? Risk neutral valuation is a method frequently used in derivatives pricing. “It is the single most important tool for the analysis of derivatives” (Hull J. , 2006). Common risk neutral models are the Black-Scholes (BS) and the Hull and White (HW) models, which are used for valuing derivatives. Both models are arbitrage-free models, based on the assumption that no profit can be made without being exposed to risk. The HW-model is used for pricing interest rate options, where the BS-model is used for pricing stock options. The underlying assumption in the BS-model is that the option price and stock price depend on the same underlying source of uncertainty. This enables the construction of a particular portfolio that consists of part of the stock and the option which eliminates this source of uncertainty. This portfolio is instantaneously riskless and has to earn the risk-free rate due to the exclusion of arbitrage opportunities (Hull J. , 2006). Using the above portfolio, the BS differential equation can be derived. This differential equation is independent from all parameters that reflect a risk preference. The solution to the differential equation is derived in a risk-free world and can be used in the real world, which leads to the principle of risk neutral valuation. Risk neutral valuation, as described above, can be used for valuation of a portfolio at the present time. 2.2 An intuitive justification of risk neutral valuation An example can be used for an intuitive justification of risk neutral valuation. Consider for simplicity a two-state binomial model (Rendleman & Bartter, 1979). In this case, the current stock price is €100 and can either go up to €200, the ‟up state of the world‟, or down to €50, „the down state of the world‟, within one year. Stock price today Stock price one year 200 100 50 Risk neutral versus real world valuation Page 6 of 38 Assume a call option with strike price €100 is available in the market and the risk free rate is 10%. It is now possible to create a riskless portfolio consisting of (buying) one share and shorting a fraction (∆) of call options, with the price of the call option equal to C. If the stock price will go up, the payout of the portfolio is €200 - ∆ call options (C). If the stock price goes down, the payout will be €50. Creating a riskless portfolio means having the same payout in both states of the world. Clearly, in the down state of the world, the payout is €50. To have the same payout in the up state, 1.5 call options should be sold, since the payout of the option in the up state of the world is €100. If arbitrage opportunities3 are excluded, the return of the portfolio must equal the risk free interest rate, i.e. 10%. Since the portfolio costs 100-1.5*C at the present time and returns 50 in one year, the value of C can be calculated: 50 − 100 − 1.5 ∙ 𝐶 = 10% 100 − 1.5 ∙ 𝐶 (1) Solving this equation for C yields a value of € 36.36, the price of the call option. Interesting to note is that in the calculations so far, the investor‟s risk preference and the probabilities of the stock going up or down are not used. Thus, the fact that the option prices are calculated without any knowledge or assumptions about the risk preference of the investor, enables one to assume risk neutrality in computing option prices. This assumption is only made for convenience, what can be clarified by an expansion of the previous example. Assuming risk neutrality of the investors means that investors do not require a risk premium for any sort of investment. Therefore, the expected return on the stock must be equal to the risk free interest rate, 10%. Therefore, it‟s possible to calculate the probability for the stock price to go up (p) under a risk neutral probability measure: p ∙ 200 + 1 − p ∙ 50 = 100 ∙ 1 + 0.1 (2) This results in a risk neutral probability of 2/5 to reach the up state of the world. The risk neutral probability can be used to determine the price of the call option, C. The value of the call option is €100 if the stock price goes up and 0 otherwise: 2 3 100 + 0 = 40 5 5 (3) The expected value of the call option at the end of the year is €40 and the discount factor equals the risk free rate. This results in a value of the call option of €36.36, the same value as calculated using no arbitrage arguments. This means that by assuming risk neutral investors, the same values 3 An arbitrage opportunity is an opportunity that guarantees a certain riskless positive excess return. Risk neutral versus real world valuation Page 7 of 38 are found. One major advantage of assuming risk neutrality of the investors is that the discount factor is equal to the risk free rate. On the contrary, it‟s not easy to find the future discount factor in the real world, which will be illustrated in the next section. Due to this advantage, among others, risk neutrality of the investors is an assumption often used in option valuation. 2.3 Problems using real world valuation Assume that the expected return on the stock in the real world is 15%, i.e. the market price of risk is 5%. Now, it is possible to calculate the probability for the stock price to go up under a real world probability measure: p ∙ 200 + 1 − p ∙ 50 = 100 ∙ (1 + 0.15) (4) This results in a value of 0.433 for the real world probability of the stock price moving up. The expected payoff of the option at the end of the year in the real world is then: 0.433 ∙ 100 + 0.567 ∙ 0 = 43.33 (5) Discounting this value at 15% (the expected return of the stock) results in a value of the call option at the present time of 37.68, which is known not to be the correct value in the (real world) market. In the preceding section, this value was found to be 36.36, by assuming no risk free arbitrage opportunities. This example illustrates that the proper discount rate is higher than 15% due to uncertainty, but it‟s not straightforward what this discount factor should be. It is now clear to see that risk neutral valuation is more convenient than real world valuation, since both the expected return of all the assets as well as the discount rate are all known in a risk neutral world. This in contrast with real world valuation, where it‟s difficult to determine the proper discount factor. Above it is illustrated that the real world probabilities can not be used for proper valuation of the option. However, these real world probabilities are needed to be able to assess the risk in holding the option. The above is merely an illustration of the difficulties that arise when using real world probabilities for valuation of options. 2.4 Risk neutral valuation in a continuous framework In the preceding sections, the model was set in a discrete environment. Its purpose was to clarify the method of risk neutral valuation, but it does not claim to be set in a realistic world. However, risk neutral valuation can also be used in a continuous, more realistic, framework. This section gives a justification for the use of risk neutral valuation in a continuous setting. At the same time, the Black-Scholes-Merton differential equation is (partly) derived. Risk neutral versus real world valuation Page 8 of 38 Consider a derivative, for example a call option C, whose underlying stock price, S, follows a geometric Brownian motion process (which in fact is the same assumption as in the Black-Scholes model): dSt = μSt dt + ςSt dWtS (6) where Ws is a Wiener process and µ and σ, respectively the expected return on the stock and the volatility of the stock price, are constants. Now Itô‟s lemma (Ito, 1951) can be employed to equation 6 in order to express dC, i.e. the change in the value of the call option, as (Broadie & Detemple, 2004): dC = ∂C ∂C 1 ∂2 C 2 2 ∂C μS + + ς S dt + Sς dW 2 ∂S ∂t 2 ∂S ∂S (7) The next step is to create a riskless portfolio consisting of the stock and the derivative. Consider a portfolio consisting of shorting one call option and buying ∆ shares. To create a riskless portfolio, the appropriate amount of shares is equal to ∂C ∂S (Hull J. , 2006). The change in the created riskless portfolio, dV, can be expressed as a combination of equations 6 and 7: dV = −dC + ∂C ∂C ∂C 1 ∂2 C 2 2 ∂C ∂C ∂C dS = − μS + + ς S dt − Sς dWs + μSdt + ςS dW ∂S ∂S ∂t 2 ∂S2 ∂S ∂S ∂S (8) where the last two terms of the right-hand side equation cancel out with its first and fourth term. The above equation is a riskless portfolio and due to the assumption of no arbitrage, it has to earn the risk free interest rate, r. So, this leads to the following equality: dV = rV dt or − ∂C 1 ∂2 C 2 2 + ς S dt = r ∂t 2 ∂S 2 −C + ∂C S ∂S (9) This equation can be rewritten as: ∂C 1 ∂2 C 2 2 ∂C + ς S − rC + rS = 0 ∂t 2 ∂S 2 ∂S (10) The above equation is the Black-Scholes-Merton (BSM) partial differential equation. The only difference is that in this equation the derivative is specified as a call option, whereas in the BSM partial differential equation no such specification is given. Solving this equation yields the value of C. Risk neutral versus real world valuation Page 9 of 38 From equation 10 is noted that no element of investor risk preference enters the BSM equation. In particular, the expected return on the stock, µ, is absent from the equation and since any form of risk preference is absent, it can‟t affect the equation. This leads to the concept of risk neutral valuation. Since risk preference is apparently not of any influence, risk neutrality can be assumed. This means that in this model, the expected return as well as the discount factor can be set equal to the risk free rate. Assuming risk neutrality of the investors is a matter of convenience. Other assumptions about risk preferences can be made as long as a proper discount factor is accompanying the assumption. The solution to the Black-Scholes-Merton equation is valid in both cases, as well in the real world as in a risk neutral world, but under a risk neutral measure it‟s easier to calculate. Risk neutral valuation does not claim to be a realistic reflection of the real world, but it does appropriately value a derivative. 2.5 From a real world to a risk neutral measure When using risk neutral valuation, the assumption is made that the stock price increases with the risk free interest rate. However, in equation 6, µ is the expected rate of return on the stock, which does not necessarily equal the risk free rate, r, in the real world. Equation 6 is a stochastic differential equation under the real world probability measure, but this equation can be rewritten to an equation under the risk neutral probability measure, a requirement for using risk neutral valuation. For this purpose, assume a probability space (Ω, F, ℱ, ℙ), where Ω is the sample space, F is the sigma field, ℙ is the real world probability measure and ℱ is the natural filtration Ft 0≤t≤T (Bingham & Kiesel, 2000). Suppose the stock price is a ℱ-adapted random process, the risk neutral probability measure is ℚ and the real world probability measure is ℙ. Recall WtSP is a Wiener process (or Brownian motion) under the real world probability measure ℙ, but now define SQ Wt = WtSP + μ−r t ς SQ WtSP = Wt or − μ−r t ς (11) At this moment, it is possible to make use of Girsanov‟s theorem (Etheridge, 2002). It states that WSQ is also a Brownian motion process, but now under a different probability measure, ℚ. Furthermore, it states that that this process is a martingale. Girsanov‟s theorem states that this is achieved by merely changing the drift of the stock price process. In this case, the change of drift was the market price of risk, i.e. μ−r ς , which makes sure that the drift of the new process equals the risk free rate (since the expected return of the stock can be seen as the interest rate plus the market price of risk). It can be shown that by changing the drift in the above fashion, the process is changed to a risk neutral process. By substituting the result of equation 11 into equation 6 this result is made obvious and yields Risk neutral versus real world valuation Page 10 of 38 SQ dSt = rt St dt + ςs St Wt (12) Thus, by rewriting the original stock price process and using Girsanov‟s theorem, it can be seen that the drift of the stock price process under the risk neutral measure equals the risk free rate. The above assumptions imply for the expectation of the stock price under a risk neutral probability measure that St = E ℚ e−r (T−t) ST | Ft (13) Thus, the stock price is discounted by the risk free interest rate. Similarly, for any derivative with payoff CT , such as a call option, it implies that Ct = E ℚ e−r (T−t) CT | Ft (14) equals the arbitrage-free value. This risk neutral value solves the BSM partial differential equation, as shown in equation 10. So far, it is shown how risk neutral valuation works and how it can be derived, as well as how to rewrite a process under a real world probability measure to a risk neutral probability measure. With explaining risk neutral valuation and what the difficulties are with real world simulation, the first of the subquestions in the introduction is answered. Using this information, it is possible to determine the risk neutral value of a product. Risk neutral versus real world valuation Page 11 of 38 3 Guaranteed product In this thesis, the focus is on a fictitious product. This product can be described as an equity linked guarantee product. Equity linked guarantees are not uncommon among insurance companies and banks, since such a product is appealing to consumers, because it offers a combination of saving and investing. The specific product offers a minimum guaranteed return, which is equal to 50% of the 1-month Euribor interest rate. Furthermore, the product contains the possibility to gain an additional investment return. The extra investment return will be δ, a constant, of the return on the Amsterdam Stock Exchange (AEX) index under the condition that the return is positive. This last part is seen as a (part of a) call option on the AEX index. In this thesis, the value of the product will be determined if it (δ) incorporates a quarter, half, or three-quarter of the return of the call option on the AEX index. The attractive property of this product is that it guarantees a certain interest rate and on top of that the return of a part of the possible rise of the stock market is also incorporated in the product. A possible set of payouts of the product is given in table 3.1. δ = 1/2 Year 1 Year 2 Year 3 Year 4 Year 5 Interest rate 5% 4,5% 4% 4,5% 4% Stock price change -2% 7% 3% -5% 1% Total return 2,5% 5,75% 3,5% 2,25% 2,5% Table 1 Example of possible payouts of the product To be able to asses the current value and the future development of the value of the product, it is necessary to investigate the development of the variables that influence the value of this product the most. In this case, it is clear to that the product is sensitive to changes in the stock price and the interest rate. These risks are examined throughout the thesis. Risk neutral versus real world valuation Page 12 of 38 4 Black-Scholes-Hull-White model In this chapter, a stochastic simulation model is constructed to value the product described in the previous chapter. As mentioned in chapter 3, the interest rate and the stock price were the main determinants of the value of the product. Therefore, the purpose is to build a proper stochastic model containing these variables. For the (stochastic) simulation of the interest rate, a one factor HW model is used. The HW model is fitted to the current term structure for the risk neutral parameters. In the BS model, a lognormal distribution is assumed for the stock prices. The one factor HW model can be incorporated in the Black Scholes world, which results in the BSHW model. The BSHW model is then used for the (risk neutral) valuation of the product. Throughout the chapter the equations are assumed to be under a risk neutral probability. 4.1 Hull-White model In using interest rate models, two kinds of models can be distinguished: a general equilibrium model (for example the Vasicek or CIR model) and the no-arbitrage model (for example the Ho-Lee model). Equilibrium models use assumptions about economic variables to estimate the interest rate (Hull J. , 2006). No-arbitrage models exactly fit today‟s term structure, i.e. the current interest rate term structure is fitted into the model in such a way that arbitrage opportunities for interest rate derivatives are excluded. In this thesis, the one-factor (no-arbitrage) HW model is used. Reason for this choice is that the HW model incorporates mean-reverting features and, with proper calibration, fits the current interest rate term structure without arbitrage opportunities (Rebonato, 2000). Furthermore, an appealing future of the HW model is its analytical tractability (Hull & White, 1990). Now, assume a probability space (Ω, F, ℱ, ℚ), where Ω is the sample space, ℚ is the risk neutral probability measure, F is the sigma field and ℱ is the natural filtration Ft 0≤t≤T as described in chapter 2. Suppose the interest rate is also a ℱ-adapted random process. The specification of the HW model for the process of the short rate under a risk neutral probability measure can be expressed as: drt = θ t − art dt + ςr dWtrQ (15) where a and σ are constants, WrQ is a Wiener process and θ(t) is a deterministic function, chosen in such a way that it exactly fits the current term structure of the interest rates. For θ(t) to exactly fit the current term structure, Hull and White (1990) provides the following solution θ t = ∂ f M (0, t) ςr 2 + a f M 0, t + 1 − e−2at ∂t 2a (16) Risk neutral versus real world valuation Page 13 of 38 Where f M 0, t is the market instantaneous forward rate at time 0 with maturity t. Now integrating equation 15 yields r t = r s e−a t−s t + e−a t−u s = r s e−a t−s t θ u du + ςr + α t − α s e−a t−s e−a t−u dWrQ u s t + ςr (17) e−a t−u dWrQ u s with α t = f M o, t + ςr 2 ( 1 − e− at )2 ) 2a2 (18) Thus, it can be concluded that 𝑟 𝑡 conditional on the natural filtration Fs is normally distributed with (Brigo & Mercurio, 2001): E r t Var r t Fs = r s e−a (t−s) + α t − α s e−a (t−s) Fs = ςr 2 2a 1 − e− 2a (19) (20) t−s Now for purposes clarified later on in this chapter, it is possible to split up r(t) in a stochastic and a deterministic part as: (21) r t = x t + α t Hence, it is possible to rewrite equation (19) to E r t Fs = x s e−a (t−s) + α t (22) To be able to use the above stochastic differential equation for Monte Carlo simulations, equations (19) and (20) have to be rewritten to a discrete process. This solution can be written as (Rebonato, 2000): rt+∆t = rt e−a ∆t + α t + ∆t − α t e−a ∆t + εr ςr 1− e −2a 2a ∆t εr ~ N 0,1 (23) Using equation (23), the simulation process can be started, but it can be noted that two constants are incorporated in equation(15). Of course, these two constants, the mean reversion speed a and Risk neutral versus real world valuation Page 14 of 38 the volatility ςr , could be chosen arbitrarily or based on historical data, but in order to achieve a no arbitrage model, values have to be chosen in such a manner that the difference between the market price of an (chosen) interest rate derivative and the HW price of the specific derivative is minimized. This process is called calibration and is elaborated on in chapter 8. 4.2 Black-Scholes model In 1973 Black and Scholes published a paper that turned out be a breakthrough in option valuation. In this paper, the Black-Scholes model was presented and from this model a closed form solution for the price of a European option could be derived. The BS model, possibly with proper adjustments, is still used for many option valuation problems in the current financial industry. One of the main assumptions made in the model is that the stock price follows a geometric Brownian motion and stock price changes are log-normally distributed. In this thesis, a stochastic simulation model is presented, which assumes the Black-Scholes world (i.e., it is assumed that the stock price follows a geometric Brownian motion and the changes in the stock price are log-normally distributed), but with stochastic interest rates. The stock price process follows equation (12), with the constant risk free interest rate r replaced by the timevarying and stochastic interest rate rt and σ equal to the implied volatility of the current market option prices. Using Itô‟s lemma, the process followed by the log of the stock price can be expressed in a discrete environment as ∆ ln S t + ∆t = rt − ςs 2 2 ∆t + ςs εs ∆t εs ~ N 0,1 (24) where εs a random sample from the standard normal distribution. Equation (24) can be rewritten to a similar equation (Hull, 2006), but with a stochastic interest rate T S T = S 0 e0 r u du − ςs2 T + ςs ε T 2 (25) Using the above equation, it is possible to derive the value of an option, even when the payoff structure is complicated. Since the chosen guaranteed product has no analytical closed form solution, Monte Carlo simulation is needed, although it is computationally costly. It is important to note that when an option does have a closed form solution, using Monte Carlo simulations will generate equivalent values for the option. The answers from both calculation methods are theoretically equal, but in practice, with a finite number of simulations, the values can differ marginally. 4.3 Correlation between the interest rate and the stock price The BSHW model also incorporates a correlation between the interest rate and the stock price (as can be seen, for example, from the importance for stock markets of the decision of central banks Risk neutral versus real world valuation Page 15 of 38 on the interest rate). This means that an unexpected movement of the stock price or interest rate influences the other. Thus, the Brownian motions for the stock price and interest rate are correlated, with ρ equal to the correlation parameter: dWrQ t dWsQ t = ρ dt (26) The correlation between the two processes needs to be incorporated in the model. It can be captured by a Cholesky decomposition of the correlated random samples of both processes. The Cholesky decomposition states that a symmetric and positive semi-definite matrix (the correlation matrix possesses these properties) can be decomposed in a lower and upper triangular matrix. Using the Cholesky decomposition, the correlation matrix, Σ, can be decomposed in the following manner: 𝛴 = L LT (27) The correlation matrix is in this case a 2 x 2 matrix, since only the correlation between the interest rate and the stock price is included. Therefore, we can decompose Σ in the following way: Σ11 Σ21 Σ12 Σ22 = 1 ρ ρ 1 = l11 l21 0 l22 l11 0 l21 l22 (28) Where the equality can be solved by: l11 = l21 = l22 = Σ11 = 1 Σ21 = ρ l11 Σ22 − l221 = (29) 1 − ρ2 By using the Cholesky decomposition, the matrix L is found. Multiplying this matrix L with the generated random samples of both the stock price process and the interest rate process, results in correlated disturbances. In this way, the correlation between both processes is captured, which is necessary for a proper stochastic simulation model. Risk neutral versus real world valuation Page 16 of 38 4.4 BSHW model The two correlated Brownian motions can now be rewritten to: WrQ WSQ = 1 ρ 0 1 − ρ2 WrQ WπQ (30) Where WrQ and WπQ are independent and the correlation between the interest rate and the stock price is taken into account by the Cholesky decomposition. As can be seen from the results from equation (30), one of the two equations is unaffected by the Cholesky decomposition. In this case, the process for the interest rate is chosen to be the first process and the stock price the second process. This choice is arbitrary, it could be the other way around, but that is of no effect on the results, merely the way to get to the results. Now, the Cholesky decomposition can be used to account for the correlation between the interest rate and the stock price and the system of two equations can be written as: dr = θ t − ar dt + ςr dWrQ dSt = rt St dt + ςs St ρ dWrQ + ςs St (31) 1 − ρ2 dWπQ (32) It is assumed that both the interest rate as the stock price are ℱ-adapted random processes, just as was previously assumed. Since the interest rate process is unaffected by the Cholesky decomposition, the conditional expectation and variance given in equation (19) and (20) are also unaffected. However, the conditional expectation and the variance for the stock price are affected by the decomposition. On top of that, the results for the stock price in the first part of section 4.2 are not based on a stochastic interest rate, but on a deterministic interest rate. This assumption is loosened later on and a stochastic interest rate is accounted for in the BSHW model and this severely changes the expectation and variance of the stock price. Now, both processes are properly defined and so is the model. The interest rate process can be simulated, but it is still necessary to calculate the expectation and variance of the stock price. It is possible to integrate equation (32) (just as equation (15) and thus equation (15) was integrated) to obtain (with time steps t an T): t+∆t S t + ∆t = S t exp r u du − t ∆t 2 ς + ςs ρ WrQ t + ∆t − WrQ t 2 s + ςs 1 − ρ2 WπQ t + ∆t − (WπQ t (33) Risk neutral versus real world valuation Page 17 of 38 Where r(t) can be split up just as in equation(21), with x(t) satisfying the following property (Brigo & Mercurio, 2001): t+∆t 1 − e−a ∆t ς x t + a a x u du = t t+∆t 1 − e−a (t+∆t−u) dWrQ u (34) t Thus, equation 33 can also be written as: 𝑆 𝑡 + ∆𝑡 = 𝑆 𝑡 exp 1 − 𝑒 −𝑎 ∆𝑡 1 𝑥 𝑡 − 𝜎 2 ∆𝑡 + 𝜎𝜌 𝑊 𝑟𝑄 𝑡 + ∆𝑡 − 𝑊 𝑟𝑄 𝑡 𝑎 2 ∗ exp 𝜎 1 − 𝜌2 𝑊 𝑠 𝑡 + ∆𝑡 − 𝑊 𝑠 𝑡 𝑡+∆𝑡 ∗ exp 𝑓 𝑀 0, 𝑢 𝑑𝑢 + 𝑡 2 𝜎 2𝑎2 𝑡+∆𝑡 + 𝜎 𝑎 𝑡+∆𝑡 [ 1 − 𝑒 −𝑎 𝑡+∆𝑡−𝑢 ] 𝑑𝑊 𝑟𝑄 𝑢 𝑡 1 − 𝑒 − 𝑎𝑢 2 (35) 𝑑𝑢 𝑡 Rewriting equation (35), it is possible to find a solution for the expectation of the log-stock price under the risk neutral probability: E ℚ [ ln S T 1 − e−a ∆t 1 Fs = x t − ςs 2 ∆t S t a 2 + ln (36) f M 0,T ς2r 2 −aT 1 + 2 ∆t + e − e−at − e−2aT − e−2at M f 0, t 2a a 2a where x(t) can be split up in: x t = r t − f M 0, t − ςr 2 ( 1 − e− at )2 2a2 (37) The variance for the stock price can also be found from equation (35) under the risk neutral probability: Varℚ [ ln S T S t Fs = ςr2 a2 ∆t + 2 a e−a ∆t − 1 2a e−2a ∆t − 3 2a + ςs 2 ∆t + 2ρ ς s ς r a ∆t − 1 a 1 − e−a ∆t (38) At this time, the expectation and variance of the interest rate and stock price are known. These can be used to produce Monte Carlo simulations of the model, where the interest rate, stock price and the correlation between them are taken into account. However, for the stock price process an expectation and variance is given, but the distribution is not. Risk neutral versus real world valuation Page 18 of 38 With the use of these Monte Carlo simulations, the risk neutral value and (for example) the Value at Risk (VaR) at the present time can be determined for the product described in chapter 3. These results will be described in chapter 5. To be able to calculate the uncertainty of the risk neutral value in the future, some extra tools are required, which will be discussed in chapter 6. Risk neutral versus real world valuation Page 19 of 38 5 Results of the risk neutral valuation of the product In this chapter, the results of the risk neutral valuation in the BSHW model will be presented. Also, some measures for risk, for example VaR, will be presented. Results will be presented for different values of δ, the extra investment return based on the performance of the AEX index. The date at which the product is valued is chosen to be June 26, 2006. 5.1 Results At the 26th of June 2006, the AEX-index denoted 427,26 and had an implied ATM (at the money) volatility of 12,27% based on a maturity of one year. The mean reversion parameter in the HW model was calibrated at 2,2% and the volatility in the HW model at 0,67%, which will be further explained in chapter 8. Also, the zero interest rate curve is given as input for the model. Using the results obtained in chapter 4, 10,000 simulations are run to simulate the (1-month) interest rate and the stock price. In figure 5.1, the history and a forecast for the next 10 years, including a lower- (LB) and an upper bound (UB), of the interest rate is shown. The expectation of the interest follows the zero curve that was given as input, just as expected. The lower- and upper bound for the interest rate shows that the HW model suggests that the interest rate can fluctuate severely. The uncertainty in the interest rate becomes larger as the time (of the forecast period) increases, which implies that the impact of the interest rate on an interest rate-linked product increases along with the maturity. 1-month interest rate 97.5% UB. RNV 9% Average estimate RNV 2,5% LB RNV 8% 7% 6% 5% 4% 3% 2% 1% 0% 1998 2000 2002 2004 2006 2008 2010 2012 2014 Figure 5.1 The history and forecast, including a confidence interval of the 1-month interest rate. Risk neutral versus real world valuation Page 20 of 38 As can be seen from figure 5.1, the 2.5% lower bound approaches zero as time increases. Normally, the interest rate does not become negative in reality. In the HW model, it is possible for the interest rate to become negative, which is considered to be a drawback of the HW model (Levin, 2004). In figure 5.2, the history and a forecast for the next 10 years of the stock price, including a upperand lower bound are shown. 1800 AEX-index 2,5% LB RNV 97,5% UB RNV Average estimate RNV 1600 1400 1200 1000 800 600 400 200 0 1995 1997 1999 2001 2003 2005 2007 2009 2011 2013 2015 2017 Figure 5.2 The history and forecast, including a upper- and lower bound, of the stock price. By running the simulations, 10.000 scenarios of possible payouts of the product are generated. In this way, the average value and confidence interval boundaries can be extracted from the simulations. As mentioned in chapter 3, δ was defined as part of the return of the call option on the AEX index. In table 2, the results of the value of the product and a lower- and upper bound are given for different values of δ. δ 0,25 0,50 0,75 average 11,7 28,2 42,4 2.5% LB -35,2 -30,2 -28,6 97.5% UB 100,2 131,0 160,6 Table 2 The average value of the product and confidence interval for different values of δ As can be seen from table 2, as δ increases the value of the product increases as well. This makes sense, since a larger part of the investment in a call option on the AEX-index, generates a higher return on average than receiving the 1-month swap interest rate. Risk neutral versus real world valuation Page 21 of 38 Furthermore, the width between the upper- and lower bound is striking. This seems counterintuitive, however, the product has a term of 30 years, which makes the current value more volatile and that explains the width of the confidence interval. In table 1, the values of the product at the 26th of June in 2006 are given. Comparing these values to the values of the product at the 29th of August in 2008 shows that the market conditions in August 2008 lead to a higher value of the product. These values are given below in table 3. δ 0,25 0,50 0,75 average 30,0 57,2 84,0 5% C.I. -26,0 -19,0 -14,2 95% C.I. 130,6 188,1 237,7 Table 3 The average value of the product for different values of δ at August 29, 2008 The higher average value of the product, when valued at the 29th of August in 2008, results from the rise of the volatility of the stock price. The recent increase in the implied volatility can be related to the „credit crunch‟. Risk neutral versus real world valuation Page 22 of 38 6 The stochastic discount factor In this chapter, the use of and theory about the stochastic discount factor will be discussed. It is explained how the stochastic discount factor (SDF) is a bridge between the real world and the risk neutral world. Furthermore, a formula for the SDF will be derived in the BSHW model and it will be justified how the SDF can be used for the valuation of the product described in chapter 3. 6.1 Theory on the stochastic discount factor Assume once more a probability space (Ω, F, ℱ, ℙ), where Ω is the sample space, ℙ is the probability measure and ℱ is the natural filtration Ft 0≤t≤T. Suppose X is a ℱt -measurable random variable, the risk neutral probability measure is ℚ and the real world probability measure is ℙ. L, the Radon-Nikodym derivative of ℚ with respect to ℙ (Etheridge, 2002), is given by L= dℚ dℙ (39) and Lt = E ℙ L | Ft (40) For equivalent probability measures4 ℚ and ℙ, given the Radon-Nikodym derivative from equation (39), the following equation holds for the random variable X (Duffie, 1996) E ℚ X = E ℙ LX (41) and E ℚ X | Ft = E ℙ X LT | Ft Lt (42) It can be seen from the above equation that the expectation of X under the probability measure ℚ is equal to the expectation of L times X under the probability measure ℙ. Suppose Wtℚ is a bivariate ℚ-Brownian motion with the natural filtration that was given above as Ft . 4 ℚ and ℙ are equivalent probability measures when it is provided that ℚ(A) > 0 if and only if ℙ(A) > 0, for any event A (Duffie, 1996). Risk neutral versus real world valuation Page 23 of 38 Define: t Lt = exp − 0 θ′s dWsP − 1 2 t 0 θ′s θs ds (43) And assume that the following equation holds, with θt being ℱ-adapted E exp( 1 2 T 0 θ′t θt dt ) < ∞ (44) Where the probability measure ℙ is defined in such a way that L = Lt is the Radon-Nikodym derivative of ℙ with respect to ℚ. Now it is possible to use the preceding to rewrite equations (41) and (42) to catch the link between risk neutral pricing and pricing under a „real world‟ probability measure T E ℚ exp − t rs ds X T | ℱt = E ℙ exp − T t T rs ds − t θ′s dWs − 1 2 T t θ′s θs ds X T | Ft (45) Combining the above equations and using Girsanov‟s theorem states that the process Q Wt = WtP + t 0 θs ds (46) is a standard Brownian motion under the probability measure ℙ. A useful feature of this theorem is that when changing the probability measure from risk neutral to real world, the volatility of the random variable X is invariant to the process. In changing from a risk neutral to a real world probability measure, it is essential to make WtP a standard Brownian motion. By Girsanov‟s theorem, this is reduced to finding the correct θs in equation (46). With this knowledge, it is possible to return to the BSHW model and find a proper SDF. 6.2 The stochastic discount factor in the BSHW model Now processes for the interest rate and the stock price under the real world probability measure ℙ are assumed. Returning to the stochastic differential equations (SDE) for the stock price and the interest rate, bearing in mind the results of the Cholesky decomposition in chapter 4, it is possible to obtain the SDE for the stock price and the interest rate under a real world probability measure. Under this measure, it is necessary to account for the market price of risk for both the stock price and the interest rate. Risk neutral versus real world valuation Page 24 of 38 The SDE for the interest rate and the stock price, both similar to the SDE under the risk neutral probability measure in equations 31 and 32, can be written as: drt = μr − art dt + ςr dWtrP (47) dSt = μt St dt + ςs St ρ dWrP t + ς s St (48) 1 − ρ2 dWtπP Where μr is the real world expectation for the interest rate, based on historical data for the one month interest rate and μt is the expected growth of the stock price, which is equal to the expected growth rate under a risk neutral probability measure plus a constant market risk premium. Now, according to the theory of section 5.1, to rewrite from a probability measure ℙ to a probability measure ℚ, it is sufficient to find θs from equation 46. This leaves the following two equations: WtrQ = WtrP + WtπQ t 0 = WtπP + θrs ds t 0 (49) θss ds By choosing a proper value for θrs , the substitution of the first part of equation 49 into equation 31 should be equal to equation 47. By solving this inequality, θrs is found to be: θrt = μr − θ(t) ςr (50) Something similar can be done to compute θst . With this knowledge, substituting the second part of equation 49 into equation 32 and solving yields: πQ dWt = dWtπP + πs ςs 1− ρ2 dt − ρ 1 − ρ2 rQ dWt − dWtrP (51) Which results in: θss = θrs 1 1 − ρ2 0 −ρ 1 − ρ2 1 πs ςs μr − θ(t) ςr (52) Assuming that equation 44 holds, which is a requirement, the proper stochastic discount factor in the BSHW model can be written as: Risk neutral versus real world valuation Page 25 of 38 T SDF t, T = exp − t T rs ds − t θss dWπℙ − 1 2 T t θss 2 ds − T t θrs dWrℙ − 1 2 T t θrs 2 ds (53) In equation 53, the stochastic discount factor is stated. To be able to use Monte Carlo simulations to value the guaranteed product under the real world probability measure, the expectation and variance of the stock price and the interest rate under the real world probability measure need to be found. In this case, it is assumed that the stock price grows with μt , which equals the risk free interest rate plus the market price of risk, as described in equation 51. Furthermore, it is assumed that the market price of risk is a constant. This means that μt can be split up in the interest rate plus a constant, πs , and so the expectation and variance can be calculated similarly to the calculations under the risk neutral measure. Also, for the interest rate a (negative) risk premium is added to the equation. This is done, since the expectation of the interest rate is different under the real world probability measure. Under the risk neutral probability measure, the interest rate process follows the initial forward curve, which is not a realistic assumption, since under a „normal‟ initial forward curve, the interest rate increases as the maturity increases. At this point, the SDF, the interest rate and the stock price processes can all be simulated. Using real world simulations, it can be shown that the value of the product at the present time is equal to the value of the product using risk neutral simulations, i.e. equations 54 and 55 can be proved to hold. 30 ℙ V0 = E [ SDF 0, T ∗ CF T | F0 ] (54) SDF 1, T ∗ CF T | F1 ] (55) T=1 30 V1 = E ℙ [ T=2 Through the tower property of conditional expectations it must hold that 30 E ℙ [V1 |F0 ] = E ℙ [ SDF 1, T ∗ CF T | F0 ] (56) T=2 The above formulas state that the market value of the product can be seen as the sum of its discounted cash flows. Since the interest rate and the stock price are known, the payout (and thus the cash flow) of the product are known per month. The discount factor is known, both at t=0 and at t=1, so the present as well as the value of the product in one year can be calculated using this method. Risk neutral versus real world valuation Page 26 of 38 Only, using real world simulations, comments on the uncertainty in the future value of the product are useful and valid, so also the conditional expectation is necessary in order to calculate the 1year VaR. The 95% 1-year VaR is defined as that particular value, which in 95% of the simulations the change in the value of the guaranteed product lies under. When calculating the VaR, it must be taken into account that the value of the guaranteed product is expected to change. So, the change in the value is defined as V1 − V0 = E ℙ [SDF 0,1 ∗ CF 1 | F0 ] + Eℙ [ 30 T=2 SDF 1, T ∗ CF T F1 − E ℙ [ 30 T=2 SDF 1, T ∗ CF T | F0 ] (57) It holds that the first part is the expected cash flow in the first period and adding both expectations can be seen as the uncertainty in the value due to a possible change in market conditions. To be able to determine the VaR of the product, the interest rate process, stock price process and the stochastic discount factor are simulated. With these instruments, the result of formula (57) can be calculated for all simulations. By taking the 95th percentile of the simulation process for formula (57), the 1-year 95% VaR of the guaranteed product can be extracted. In the next chapter, the results will be presented. However, a drawback of the model must be noted here. The conditional expectations in formulas (55) and (57) assume that the parameters of the model are known at t=1 (..|F1 ). Obviously, the (known) market conditions at t=0 will be different from the (unknown) market conditions at t=1. This parameter risk is neglected in these calculations, which is a drawback of the model. Risk neutral versus real world valuation Page 27 of 38 7 Results of the real world valuation of the product This chapter presents the results of the real world valuation by means of the BSHW model. Also, some measures for risk, for example VaR, will be presented. Results will be presented for different values of δ, the extra investment return based on the performance of the AEX index. The date at which the product is valued is chosen to be the 26th of June in 2006. 7.1 Results Just as in chapter 5, the AEX-index denoted 427,26 had an implied ATM (at the money) volatility of 12,27% based on a maturity of one year. The mean reversion parameter in the HW model was calibrated at 2,2% and the volatility at 0,67%. Also, the zero interest rate curve is given as input for the model. The average 1-month interest rate μr is chosen to be 4,27% based on historical data. Furthermore, the risk premium, πs , is fixed at 3%. Using the results obtained in chapter 6, 10,000 simulations are run to simulate the (1-month) interest rate and the stock price. In figure 3, the history and a forecast for the next 10 years, including a lower- and upper bound, of the interest rate are shown. Just as under the risk neutral probability measure, the uncertainty in the interest rate becomes larger as the time (of the forecast period) increases, which implies that the impact of the interest rate on an interest ratelinked product increases along with the maturity. 8% 7% 1-month interest rate RW Average estimate RW 97.5% UB RW 2.5% LB RW 6% 5% 4% 3% 2% 1% 0% 1999 2001 2003 2005 2007 2009 2011 2013 2015 Figure 7.1 The history, forecast and its lower- and upper bound interval of the 1-month interest rate. Figure 7.1 depicts the development of the 1-month interest rate under the real world probability measure compared to the development under the risk neutral probability measure. Clearly, under the risk neutral probability measure the interest rate increases faster. The latter can be explained by the fact that it follows the current zero curve very closely, as can also be seen from figure 7.1. Risk neutral versus real world valuation Page 28 of 38 The interest rate tends to the long-term average under the real world probability measure, but it must be noted that this is a slow process. If a longer forecast period would have been chosen, this would be clearer. 6.0% 1-mo nth inte re s t ra te RW Ave ra g e e s tima te RW Ave ra g e e s tima te RNV The fo rw a rd z e ro curve 5.0% 4.0% 3.0% 2.0% 1.0% 0.0% 1999 2001 2003 2005 2007 2009 2011 2013 2015 Figure 7.2 The history, both forecasts and zero curve of the 1-month interest rate In figure 7.2, the history, the zero curve and both forecasts are shown for the next 10 years. Figure 7.3 shows the history, forecast and a lower- and upper bound of the AEX-index under both probability measures. As can be seen, the average predicted value of the AEX-index has a smooth course, but the width of the confidence interval shows that the predicted values of the index are in fact rather volatile. As expected, under the real world probability measure the index increases faster on average. This is due to the addition of the risk premium of 3% under this real world measure. Risk neutral versus real world valuation Page 29 of 38 1000 800 AEX-index RN 98% C.I. RW 98% C.I. RN average estimate RW average estimate 600 400 200 0 Jan-04 The level of the AEX-index (412,84) at August 31, 2008 is used as a starting point for the analysis. The current level lies outside the 98% confidence interval. Jan-05 Jan-06 Jan-07 Jan-08 Jan-09 Jan-10 Jan-11 Figure 7.3 The history and forecast, including a lower- and upper bound of the stock price Table 4 shows the average value and lower- and upper bound for the guaranteed product under investigation for different values of δ. On the left hand side, the average value and lower- and upper bound are reproduced for calculations under the real world probability measure. On the right hand side, the results from chapter 5 are shown again. 30/6/2006 Real World δ average 2,5% LB 0,25 13,3 -33,6 0,5 27,5 -30,5 0,75 42,7 -33,6 97.5% UB 100,7 127,2 153,5 δ 0,25 0,5 0,75 average 11,7 28,2 42,4 Risk Neutral 2,5% LB -35,2 -30,2 -28,6 97.5% UB 100,2 131,0 160,6 Table 4 The average value of the product and the C.I. for different values of δ. Clearly, the results for both calculation methods are very similar. This actually makes sense, since the theory discussed in chapters 4 and 6 states that the average results should be the same. The (minor) differences can first be explained by the fact that a different set of simulations is run for both methods. Second, a discrete approximation for a part of the stochastic discount factor had to be made in order to use it in the stochastic simulation model. When taking these reasons into account, the differences between the estimated values under a risk neutral and a real world probability measure can be surmountable. Also, the difference in the estimated values of the product at the 29th of August 2008 are reasonable. These values are given in table 5. Risk neutral versus real world valuation Page 30 of 38 29/8/2008 Real World δ average 2,5% LB 0,25 32,1 -23,9 0,5 54,4 -17,8 0,75 82,5 -17,3 97.5% UB 133,3 181,6 225,1 δ 0,25 0,5 0,75 average 30,0 57,2 84,0 Risk Neutral 2,5% LB -26,0 -19,0 -14,2 97.5% UB 130,6 188,1 237,7 Table 5 The average value of the product for different values of δ at August 29, 2008. Just as under the risk neutral probability measure, the average values are much higher due to the rise in the implied volatility. Now, the present value of the product is calculated under a real world probability measure. Using these real world simulations, the value of the product in one year and the uncertainty in this value can be calculated. Also, the 95% VaR of the product can be calculated. By taking the difference between the value of the product at the present time and the expected market value at t=1, one can simulate the change in the value of the product. This change should be corrected for the actual expected change, since the time to maturity of the product has declined at t=1. The VaR is defined by the difference between the average at t=1 minus its 95% boundary. The results are given in table 6 and 7: Risk Neutral Date Expected market value in 1 year 30/6/2006 28,1 34,5 29/8/2008 57,4 67,2 5% LB Real World VaR Expected market value in 1 year 20,2 6,4 28,4 35,6 21,2 7,2 44,1 9,8 55,7 66,4 45,0 10,7 95% UB 5% LB 95% UB VaR Table 6 The market values and the VaR of the guaranteed product Whether the value of the product in one year is estimated correctly can be tested by using the method of backtesting. This method is described in the next section. 7.2 Backtest To examine the forecast capabilities of the model, the results can be tested by the method of backtesting. Backtesting is a method where the model is used to calculate the present value of the product on several different moments in the past and after that compares the results to the actual (already known) outcomes. The model under the real world probability measure is used to predict the value of the product in one year. However, it is difficult to create enough observations and therefore, a rolling window of one year is used to create more observations. Since the dataset starts in May 2003, still 51 observations are available for the backtest. In all of these 51 observations, it will be tested whether the actual value of the product lies outside the 90% confidence interval of the predicted value, generated in the model under the real world probability measure. The results for the backtest are shown in figure 7.4. Risk neutral versus real world valuation Page 31 of 38 20 90% CI boundaries RW 90% CI boundaries RN Realised market value 15 10 5 0 -5 -10 jun-04 dec-04 jun-05 dec-05 jun-06 dec-06 jun-07 dec-07 jun-08 Figure 7.4 Backtest results for the prediction of the value of the product in one year. What can be concluded from the above figure, is that in particular the observations in the last year of the dataset fall outside the predicted confidence interval. In total 15 of the 51 observations, lie outside the predicted 90% confidence interval. These results can be mainly attributed to the rise in the implied volatility due to the turbulent market conditions from May 2007 on, which can be seen in figure 7.5. 30,0 Implied volatility AEX-index 22,5 15,0 7,5 0,0 jun-04 dec-04 jun-05 dec-05 jun-06 dec-06 jun-07 dec-07 jun-08 Figure 7.5 A graph of the implied volatility of the AEX-index Whether the model passes the backtest can be calculated in a likelihood ratio testing framework (Christoffersen, 1998). In this framework, suppose that It T t=1 is the indicator variable for the interval forecast given by the model under the real world probability measure, which means that whenever It =1 the actual value lies in the interval. The conditional coverage can be tested by Risk neutral versus real world valuation Page 32 of 38 comparing the null hypothesis that E It = p with the alternative hypothesis that E It ≠ p, given independence. The likelihoods under the null hypothesis and under the alternative hypothesis are given by: L p; I1 , I2 , … , I52 = px 1 − p n−x L π; I1 , I2 , … , I52 = πx 1 − π n−x (58) x Where the maximum likelihood estimate of π is n , the number of values outside the interval forecast divided by n, the total number of observations. Using these likelihoods, a likelihood ratio test for the test of the conditional coverage can be formulated LR cc = −2 ∗ log L p; I1 , I2 , … , I52 / L π ; I1 , I2 , … , I52 ~ χ2 2 (59) Where the test statistic is actually asymptotically Chi-Squared distributed with s(s-1) degrees of freedom, with s=2 as the number of possible outcomes. Using this statistic it is difficult to take the autocorrelation (due to the use of the rolling window) into account. So, the resulting conclusions are less powerful. In this case, the LR-test statistic is 14,4, which is higher than the 5,99 from the (5%) confidence level of the Chi-squared distribution, what justifies the conclusion that the model is inaccurate. However, the recent crisis is a very unexpected event. If this point would have been left out of consideration, the backtest would have a totally different outcome. These results are presented in figure 7.6. 20 90% CI boundaries RW 90% CI boundaries RN Realised market value 15 10 5 0 -5 -10 jun-04 dec-04 jun-05 dec-05 jun-06 dec-06 Figure 7.6 Results of the backtest until may 2007 The LR test statistic for this dataset is 0,048, which would lead to not rejecting the model, as opposed to a rejection taking the data from May 2007 until August 2008 into account. Risk neutral versus real world valuation Page 33 of 38 When these tests are performed for the model under a risk neutral probability measure, both tests perform similar as in the model under the real world probability measure, see table 7. Date Until August 2008 Until May 2007 Real world Risk neutral 1% critical value 5% critical value 14,4 0,05 23,6 1,67 9,21 5,99 Table 7 Results of the backtest for both models So, when the data until May 2007 are used to backtest the model under a real world probability measure, it is not rejected. This is similar to the model under the risk neutral probability measure. Risk neutral versus real world valuation Page 34 of 38 8 Summary and conclusions The objective of this article was to determine the uncertainty in the future market value of the product. To be able to determine this, a stochastic discount factor (SDF) was used. The SDF, also called deflator, was needed for proper valuation using real world simulations. For the HWBS model an SDF could be constructed, which was used to value the guaranteed product. The most important conclusions that can be drawn from the results and the backtest are: Valuation under the real world probability measure calculates the market value of a product with the use of a stochastic discount factor. The main advantage of using real world simulations is that the simulations can also be used for a „realistic‟ simulation of random variables. Combining the real world simulations with a stochastic discount factor is very useful for banks and insurers. They can use this method to estimate the current value of their products and, more importantly, estimate the uncertainty in this value in one year in a consistent way. This can be used in regulatory (e.g. Basel II or Solvency II) and economic capital calculations. Capital calculations are typically based on a one year 99% VaR. When using real world simulations and a standard discount factor, estimated average values are inaccurate, therefore, resulting VaR calculations can be as well. When using risk neutral valuation to estimate the VaR, only current market conditions are taken into account. Current market conditions are not necessarily a good measure for future outcomes, which could also lead to inaccurate VaR estimations. However, some drawbacks of the model must be noted. The model under the real world probability measure, using the SDF, did not pass the backtest. The null hypothesis that the model correctly predicts the uncertainty in the future value is rejected. The failure of the model in the backtest needs to be taken seriously. However, as already mentioned, the market conditions in the last period of the sample, are quite unusual. Whenever the dataset is cut off at May 2007, the model passes the backtest unlike the model under a risk neutral probability measure. Of course, doing this would be a case of data mining, but it does not alter the fact that the current market conditions are difficult to take into account. It could be defined as an outlier, some theories state that the recent crisis is comparable to the crisis in the twenties. Two variables, the stock price and interest rate, are modelled stochastically. When more variables are modelled stochastically, the SDF becomes more complicated. For banks and insurers, who also model variables like exchange rates and volatility stochastically, several more random variables enter the model. As the results have shown, the value of the product greatly depends on this input and modelling this input as a random variable could help to improve the Risk neutral versus real world valuation Page 35 of 38 forecasting qualities of the model. However, this would make the model and the SDF more complicated and less practical. Risk neutral versus real world valuation Page 36 of 38 9 References Bingham, N., & Kiesel, R. (2000). Risk-neutral valuation. Springer Finance. Black, F., & Scholes, M. (1973). The pricing of options and corporate liabilities. Journal of Political Economy , 637-654. Brigo, D., & Mercurio, F. (2001). Interest rate models theory and practice. Heidelberg: Springer Finance. Broadie, M., & Detemple, J. B. (2004). Option pricing: valuation models and applications. Management Science , 1145–1177. Christoffersen, P. F. (1998). Evaluating interval forecasts. Washington: International Monetary Fund. Duffie, D. (1996). Dynamic asset pricing theory . Princeton University Press. Etheridge, A. (2002). A course in financial calculus. Cambridge: Cambridge University Press. Gibbs, S., & McNamara, E. (2007). Practical issues in ALM and stochastic modelling for actuaries. Christchurch: Trowbridge Deloitte. Hull, J. (2006). Options, Futures and other derivatives. New Jersey: Prentice Hall. Hull, J., & White, A. (1990). Pricing interest-rate-derivative securities. The Review of Financial Studies , 573-592. Ito, K. (1951). On stochastic differential equations. Memoirs, American Mathematical Society , 151. Levin, A. (2004). Interest rate model selection. New York: Andrew Davidson & Co., Inc. . Nielsen, J., & Sandmann, K. (1995). Equity-linked life insurance: a model with stochastic interest rates. Insurance: Mathematics and Economics , 225-253. Rebonato, R. (2000). Interest-rate option models. Chichester: John Wiley & Sons. Rendleman, R. J., & Bartter, B. J. (1979). Two-state option pricing. The journal of finance , 10931110. Tamba, Y. (2005). Pricing a bermudan swaption with a short rate lattice method. Osaka: 21st Century Center of Excellence Program (COE). Risk neutral versus real world valuation Page 37 of 38 Appendix explanation used data This appendix discusses the assumptions that were made about the parameters needed for the input in the model. Also, some estimation techniques for certain parameters needed in the model will be elaborated on. Used data First, the estimation of the parameters in the HW model will be discussed. This estimation is obtained by calibration. When calibrating the mean reversion speed and the volatility (Tamba, 2005), a specific interest rate derivative has to be chosen. Here, the instruments that are used for calibration are at the money swaption premiums5 (Levin, 2004). The prices of the swaption premiums are extracted for different maturities and expiry dates. Then, in the HW model the swaption premiums can be calculated for a fixed mean reversion speed and volatility. Of course, using several different swaption premiums (i.e. swaption premiums for swaptions with different maturities etc), no exact fit can be found. So, a specific goodness of fit measure has to be chosen, in this case the absolute difference between the prices of the swaption premiums generated from the HW model and the actual (market priced) swaption premiums is minimized. In this model, the mean reversion parameter and volatility are chosen to be constant. This means that there are two volatility parameters (Hull J. , 2006). Using the absolute difference between both (above described) prices as the goodness of fit measure, the best fitting values of the mean reversion speed and the volatility can be calculated with an optimization approach. Here, the calibrated mean reversion speed equals 2,2% and the calibrated volatility 0,67%. When backtesting the model in section 7.2, the calibration procedure was used every month. This was necessary to exclude arbitrage opportunities at the start of every month used in the backtest procedure. In the model described in chapter 4 and extended in chapter 6, a constant volatility of the stock price is assumed. The volatility can be estimated by the implied volatility, which is determined by option prices observed in the market. In an ordinary option pricing model, such as Black-Scholes, the price of an option can be written as a function of the volatility. This suggests that the other parameters in the model are kept constant. If an option price is observed in the market, with a known maturity and strike price (normally chosen to be close to the present stock price), the same price can be obtained via the BlackScholes model by changing the volatility. The latter can be achieved by an iterative procedure. This vested volatility is also known as the implied volatility (ATM and with a maturity of one year). Furthermore, the risk premium is given as input for the model. A proper estimation for the risk premium is very difficult. An estimation based on historical data is not very plausible, since the 5 All swaption premiums are extracted from Bloomberg. Risk neutral versus real world valuation Page 38 of 38 value of the risk premium is highly dependable on the chosen length of the historical data. A risk premium of 3% for the AEX-index does not seem fairly unreasonable and is chosen as constant risk premium. Also, some adjustments have to be made for the interest rate. Under a risk neutral probability measure, the interest rate follows the zero curve, which is not a realistic assumption. It is more realistic to expect the interest rate to drift around some long-term average, which was defined to be μr . This long-term average can be estimated using historical data of the 1-month Euribor interest rate and equals 3,8% for our sample period.