1.

Introduction Option markets have grown spectacularly over the last three decades throughout the world, and with this growth in the market came growth in demand for accurate pricing methods. As noted by Bakshi et al (1997), the sheer number of models, even a decade ago, was overwhelming. Choosing the best model for option pricing today, therefore, is a very difficult task. A model should be judged on a number of criteria, including fit to the data, internal consistency, and analytic tractability. While most complicated models have had good success in improving fit and some success in achieving consistency, tractability has become a serious issue in option pricing (Garcia et al, 2002).

One particular class of models that have succeeded in also achieving analytic tractability is the class of affine jump diffusion option pricing models. Models in this class may be defined with complex processes for the underlying price and option pricing is provided through the transform and inversion methods of Duffie et al (2000) and Bakshi and Madan (2000). This results in semiclosed form formulas that require only univariate numeric integration of a well-behaved function. Models in this class include the stochastic volatility models stemming from the work of Heston (1993), the jump-diffusion models beginning with Merton (1976) and Bates (1991) as well as combinations of the two such as in Bates (1996 and 2000) and Bakshi et al (1997). Some of the most complex models of recent times, with stochastic volatility and jumps in both prices and volatility such as Eraker et al (2004) and Chernov (2006) are also members of this class.

Testing the fit of these models has been a fruitful area of research for many. While diverse methods are variously employed to evaluate these models, as noted by Chernov and Ghysels (2000) and Garcia (2002), the estimation of the implicit distribution from option prices can be an especially informative way to learn about option pricing models and hence decide among

alternatives. This avoids the complicated problem of estimating the risk premiums that must be added to a model based on objective parameters to produce option prices

The relative empirical desirability of some of these models for the FTSE/JSE Top 40 Index options market has been studied by many researchers. Most of these advanced models, however, have not been studied in the South African context. The double jump model considered in this paper has, to the authors best knowledge, not been studied empirically in any context previously. The need for the investigation of these models is confirmed by Twite (1996) and Brown (1999), who look at the implicit volatility of FTSE/JSE Top 40 Index futures options. They each document consistent biases, concluding that there need for a more complex option pricing model more consistent with the underlying distribution.

The aim of this paper, therefore, is to introduce a more complex class of models to the South African context, and study their utility and the information they convey about the implicit underlying distribution. As such, the purpose of this paper is two-fold. Firstly, it aims to test the empirical desirability of each of these models in the South African context and look at the information about the implicit distribution provided. Secondly, it aims to study the internal consistency of the models used and provide evidence on possible misspecification.

The research reveals a number of interesting findings on the Johannesburg index options market. We find that the distribution implicit in call options is substantially different to that of put options. It is difficult to determine, however, whether the difference is a true one or a reflection of the different structure of the call and put option data samples. We find that the best model for fitting call options is a stochastic volatility with jumps in price and volatility while for put options the best model is a stochastic volatility with jumps in price but not volatility. Under the assumption of near constant parameters through time a more parsimonious model is the best choice, with a plain stochastic volatility model performing best for call options and a jumpdiffusion or stochastic volatility model for put options. The choice, however, is shown to be a decision between significantly misspecified models through analysis of model internal consistency and implicit volatilities.

The remainder of this paper is structured as follows. Section 2 reviews some of the relevant literature. Section 3 introduces and explains the data sources. Following this, Section 4 formally introduces the models to be studied and discusses all relevant methodological issues. Section 5 presents the empirical results of the paper while Section 6 concludes.

Review of Relevant Literature An option can be perfectly priced if we know the expected future value of the underlying and the appropriate discounting kernel. The Black-Scholes model finds the expected value of the underlying using a log-normal assumption for price changes. This approach, however, has been shown by many to produce a very poor fit empirically. Obvious biases, often characterised as smiles or skews in the implicit volatility of that model, have been documented extensively. In the South African market, Twite (1996) and Brown (1999) look at the implicit volatility of S&P/ASX200 index futures options and find significant biases. Brown (1999) also concludes explicitly that there are risk premiums implicit in option prices that should be incorporated in an appropriate option pricing model, highlighting the need for the exploration of more advanced models, something we aim to deal with here.

Options are not redundant securities as assumed by the Black-Scholes model, but in fact span relevant risk factors. This implies not only that the risk free rate cannot be used in discounting the expected value of the underlying, but also a model that accounts for these risk factors must be used in finding this expected value. Risk factors such as volatility risk (Pan, 2002) and jump risk (Chernov, 2006) have been investigated and shown to be relevant as well as general unparameterised risks (Ait-Sahalia and Duarte, 2002).

This paper concentrates on parametric forms for volatility and jump risk in modelling the underlying since this provides more informative conclusions about the distribution of underlying price changes and is well supported empirically (Garcia et al, 2002). Generally, the modelling of these risk factors is best done with fully stochastic specifications, often involving a further state

variable for the level of volatility, such as models from the class of affine models studied in this paper (Dumas et al, 1998).

These stochastic models aim to capture the common empirical features of returns, such as time changing and mean reverting variance whose innovations are correlated with price changes (Engle and Patton, 2001) and sudden, large, discontinuous changes in prices (Bates, 1991) and volatility (Naik, 1993). Models capable of addressing some or all of these features include the stochastic volatility model of Heston (1993), the jump-diffusion model of Merton (1976), the variance gamma model of Madan et al (1998), the constant elasticity of variance model of Jones (2003) and the combined stochastic volatility and jumps models of Bates (1996), Bakshi et al (1997) and Duffie et al (2000).

Option pricing using these complicated stochastic models can be very difficult to apply, especially in the case of the variance gamma and constant elasticity of variance models (Bakshi et al, 2006). The stochastic volatility and jump-diffusion models, and their combinations, being part of the class of affine models, however, have semi-closed form solutions available due to the work of Bakshi and Madan (2000) and Duffie et al (2000), extending Heston (1993) and Stein and Stein (1991). These solutions use transform analysis and inversions to reduce option pricing to simple univariate integration, which we address with a speedy method of non-adaptive quadrature.

The use of these solutions, however, requires knowing not only the level of the latent volatility state variable, but also the parameters that describe the model. Risk premiums associated with these newly defined risks must also be dealt with. These risk premiums mean that the objective parameters that describe the actual changes in the underlying (under the objective probability measure P) cannot be used to price options using the risk free rate. There exists an equivalent martingale measure, the risk neutral probability measure Q, under which the risk neutral parameters can be used to price options using the risk free rate for discounting. The risk neutral parameters, however, cannot be estimated from the time series of returns due to the transformation to the risk neutral probability measure.

Implicit parameter estimation is a way to recover the risk neutral parameters using option price data. Under the assumption of accurately measured option prices, an appropriate cross-section of option prices can be used to effectively invert an option pricing model and exactly recover the parameters of the underlying risk neutral returns distribution (Bates, 2003). Knowing these risk neutral parameters allows us to learn about an option pricing model and the implicit view an option market has of the risk neutral return distribution, as well as measure the fit and internal consistency of a model. This type of estimation is used here to study not only the option pricing models under consideration, but also the implicit distributions of call and put options in South Africa.

This approach has been used by many researchers to study models from the class of Affine option pricing models, such as Bakshi et al (1997), Bates (1996 and 2000) and Pan (2002). These models have had success in improving over the standard Black-Scholes model, often used as the straw man to measure improvement against (Bates, 2003). Bakshi et al (1997) and Bates (1996 and 2000) all find that stochastic volatility offers the greatest first order improvement in fit, and that jumps in prices are also important. They do, however, find that the process is still significantly misspecified and call for the inclusion of jumps in volatility.

The results regarding the utility of a jump in volatility process is mixed. Eraker et al (2003) find that it is very relevant for fitting the time series of returns, though Eraker (2004) finds that that a simultaneous jump process where jumps in price and volatility are correlated and occur at the same time is not relevant in fitting options data (through implicit estimation). An independent jumps in volatility process is used here, similar to that specified by Chernov (2006), though he does not test this process with regard to fit but rather uses it to investigate volatility forecasting. This paper is the first study of such an independent jump in volatility model in the context of options data.

3. Data and Sample Statistics

3.1. Data

This paper uses data on FTSE/JSE Top 40 option prices, the underlying FTSE/JSE Top 40 index levels and the BBSW (Bank Bill Swap) interest rate, a proxy for the risk free rate of return. The data cover the period from 1 January 2006 to 31 December 2010. The options data are provided by the Johannesburg Stock Exchange. FTSE/JSE Top 40 index options are European in style and cash-settled with quarterly expirations. They are generally available over a wide variety of exercise prices and several maturities, though only a subset of these are traded on any given day.
st st

The risk-free interest rate data are sourced from the Bloomberg data service. The Bank-Bill Swap Rates (BBSW), as calculated by the South African Financial Markets Association (SAFMA), for various maturities are the primary series. The BBSW however, does not cover the entire sample for all maturities. In the earlier part of the sample, the predecessor to the BBSW 1year rate is used. We use one, two, three, four, five and six month rates and a one year rate. The interest rate to be used with each option is matched to the options maturity by linearly interpolating the two nearest interest rate series. The index levels are also obtained from the Bloomberg data service. The data are the daily level of the index underlying FTSE/JSE Top 40.

In some studies, such as Bates (1996 and 2000), the joint data set of call and put options is used for implicit parameter estimation. In other implicit parameter estimation studies, for example, Bakshi, Cao and Chen (1997), only call options are used to simplify estimation. We will use both put and call options to study the differences that may appear between the two. Since FTSE/JSE Top 40 options are European style, put-call-parity may be used to calculate put prices from the associated call price under the assumption of no transaction costs (Stoll, 1969). We use the put-call parity formula:

( )

( )

(1)

Where ( ) is the price of a put maturing in periods (years) and ( ) is the price of a call maturing in periods (years). X is the exercise price of the option, S0 is the initial stock price and r is the risk free rate of return.

To prepare the data for estimation, a number of offending daily option prices are removed for varying reasons. Following Bakshi et al (1997), all observations that do not satisfy the following minimum value arbitrage constraints are removed: ( ) max [0, S0 XB( )], ( ) max [0, XB( ) S0]

(2) (3)

Where B( ) is the current price of a R1 zero coupon bond with the same maturity as the option.

Imposing this condition results in the removal of about 8% of observations at this stage, much larger than the 1.3% removed in Bakshi et al (1997). Due to much lower liquidity in the options market in South Africa, it was expected that there would be a larger number of observations violating this condition in the South African market compared to the American S&P500 index options market used more commonly in studies.

Certain FTSE/JSE Top 40 index option series have an exercise price of zero. These are all calls, and they are nicknamed LEPOs (Low Exercise Price Options). Since these options require no payment on exercise and are always in-the-money (hence always rationally exercised), they behave more like a forward contract. Because of this unusual behaviour, they too are removed from the sample. Very short-term options (those with less than 5 days to maturity) may also introduce bias. In line with Bakshi et al (1997) (who remove options with less than 6 days to maturity) we remove all observations from the sample that have less than 5 days to maturity.

3.3. Sample Statistics

Table 1 summarises some of the properties of the sample call and put data. The data are divided into nine categories using moneyness and term to expiration. The most common options are medium-term and at-the-money. Long-term and in-the-money options appear least frequently. Put options tend to be much more likely to be out of the money than call options, though for both options types the average moneyness is out of, rather than in-the-money. These differing characteristics of the sample of put and call data is relevant in discussions in later sections. The average option maturity, daily volume and open interest are similar between calls and puts, though the number of actively traded series is greater for puts, indicating greater overall activity in this type of option. This allows the more complex models to be estimated on more days for put options than for call options. Average market prices for options are also reported and, as expected, value increases, on average, with term-to-expiration and moneyness.

Table 1. Sample Properties of FTSE/JSE Top 40 Index Options

Panel I: Call options

Moneyness OTM M< 0.97 ATM 0.97 < M < 1.03 ITM M> 1.03 Totals

maturity < 60
(short) 9.51 pts (744) 42.36 pts (3181) 287.31pts (276) 52.64 pts (4201)

60 Maturity

Maturity (days)

180

maturity 180
(long) Totals 51.76 pts (1001) 126.12 pts (553) 260.85 pts (17) 80.20 pts (1571) 29.85 pts (4490) 67.51 pts (7805) 261.71 pts (523) 62.26 pts (12818)

(medium) 27.37 pts (2745) 79.62 pts (4071) 231.04 pts (230) 64.00 pts (7046) Sample Averages Open Interest

S/X 0.98

Maturity (days) 97.55 days

Volume

Series actively traded per day 10.56 series

76.17 contracts

833.10 contracts

Panel II: Put options Maturity (days) Moneyness OTM M> 1.03 ATM 0.97 < M < 1.03 ITM M< 0.97 Totals

maturity < 60
(short) 11.99 pts (1953) 47.16 pts (3148) 223.03 pts (475) 49.82 pts (5576)

60 = maturity <
(medium) 30.69pts (4610) 81.62 pts (4019) 242.20 pts (809) 70.51 pts (9438)

180

maturity = 180
(long) Totals 58.08 pts (1827) 122.06 pts (822) 334.69 pts (176) 93.93 pts (2852) 32.30 pts (8390) 72.20 pts (7989) 247.11 pts (1460) 67.75 pts (17839)

Table 1 reports the average prices and number of observations (in brackets) that fall into each category. In addition, some sample averages are also reported. The sample covers the years 2006 to 2010. S denotes the index level and X the exercise price. OTM, ATM and ITM denote out-ofthe-money, at-the-money and in-the-money options, respectively. Moneyness is defined as M=S/X. Panel I shows the statistics for call options, while Panel II shows those for put options.

4. Models and Methodology In this section, we will introduce the specification of the models to be studied and describe the methodology for the estimation of model parameters and option pricing fit. The five models studied are nested within the general framework of the Duffie et al (2000) derived double jump model (SVJJ). The method of implicitly estimating option pricing model parameters each day from option prices is described and the sum squared error objective function and its advantages and disadvantages are discussed. The method for summarising these daily parameter estimates to obtain whole-sample estimates and a method for efficiently evaluating the integrals associated with the option pricing formulas are also presented.

4.1. Option Pricing Models

The double jump model (SVJJ) is the most complex of the models to be used in this paper. This specification is derived from the general specification of Duffie et al (2000). A double jump specification with simultaneous price and volatility jumps, also derived from the Duffie et al (2000) framework, has been studied by Eraker et al (2003) and Eraker (2004). Our specification, with independent price and volatility jumps has not been studied previously in this manner. We define the underlying price process in terms of risk neutral parameters as follows:
( ) ( )

=(

( )

( )

( )

( )

(4)

Where V(t) is itself stochastic and obeys the following process ( ) [ ( )] ( ) ( )

( )

( )

(5)

J is modeled as

( )]

(6)

J(t) can be modeled as ( )

( )

(7)

As shown by Duffie et al (2000), under risk neutrality, the following must hold:

( )

(8)

One should notice that the mean volatility jump size can be only positive (and hence all volatility jumps must be positive) otherwise the exponential distribution would not be defined. The process

is defined in this way intentionally by Duffie et al (2000) for the following reasons. Firstly, negative jumps in volatility suggest the possibility of volatility becoming negative in the case of a large jump. This would destroy the reflecting barrier at zero that forces the square root volatility model to maintain positive volatility, and produce negative option prices. Secondly, negative jumps in volatility are not a commonly observed phenomenon.

4.2. Empirical Methodology This sub-section introduces the methods used in implicit parameter estimation and discusses the issues involved in their use. The method used in computing option pricing integrals is described and the method used to calculate implicit volatilities is also discussed.

4.2.1. Implicit Parameter Estimation Implicit parameter estimation is the method used to estimate the parameters of the option pricing models in this paper. Garcia et al (2002) suggest that this type of study of the risk neutral parameters is a more informative exercise compared to the use of time series methods and objective parameters. As found by Chernov and Ghysels (2000), using only options data is the preferred method for recovering risk neutral parameters, rather than using a combination of options data and underlying index data. Recovering the risk neutral parameters implicitly involves selecting a basket of options, at least as many as the number of variables to be estimated, and selecting an objective function to measure fit. We follow many other studies that use this method and choose an objective function that minimises the sum of squared absolute pricing errors. ( ( )) ) Where the ith option in the basket, ( ) is the level of implicit volatility and [ ( ( ) )]

(9) is the model price of represents the parameter

vector, whose length is determined by the model being estimated.

Table 2: Whole-Sample Implicit Risk Neutral Parameters

The minimisation of the objective function (9) could be viewed as a non-linear least squares
.

problem and solved using some non-linear least squares algorithm (Bates, 2000). We formulate the problem as the following non-linear least squares regression: ( ( ) )

(10)

An important consideration before performing the non-linear least squares minimisation in (10) is the form of the errors, i. As noted by Bates (2003), there is no established theory for the form of option pricing errors. Bates (1996 and 2003) notes that these errors must be at least heteroskedastic with respect to exercise price and maturity and are generally contemporaneously correlated across different exercise price-maturity categories and serially correlated over time. Since the heteroskedasticity and correlation problems should theoretically only affect standard errors of the daily estimates, we should not expect bias in the parameter and state variable estimates due to this.

5. Empirical Results In analysing the five option pricing models (BS, JD, SV, SVJ and SVJJ), we invert each days sets of option prices to obtain the implicit risk neutral parameters and volatility level for each different model. Each day has two sets of options, the first set contains all call options actively traded on the day and the second set contains all actively traded put options.

The median of the parameters and volatility level over all days under consideration is taken for each model and option-type pair and their 90% confidence intervals are calculated. In judging fit, we will consider all days that the options implicit parameters were successfully estimated in calculating the median. While this affords a bias to smaller models that are able to easily fit days with fewer options due to less variation in exercise prices and maturities, this is balanced by the larger models having more free parameters. The results are reported in Table 2.

Table 2: Whole-Sample Implicit Risk Neutral Parameters

Model BS JD Days 1182 1031 8.447% Imp. 8.156% Vol. 8.758% Call Options SV SVJ 976 744 6.960% 9.472% 6.516% 8.813% 10.126% 7.353% 5.295 5.242 5.337 7.021% 6.649% 7.307% 15.316% 13.879% 16.750% 0.463 0.327 0.609 1.062 1.034 1.105 SVJJ 561 9.090% 8.470% 9.715% 5.765 5.673 5.840 5.321% 4.850% 5.774% 11.902% 10.769% 13.050% 0.389 0.219 0.564 0.670 0.648 0.688 BS 1171 10.295% 9.510% 11.231% 4.818 4.523 5.095 6.446% 5.777% 7.161% 29.312% 26.738% 32.024% -0.397 -0.610 -0.261 0.558 0.487 0.646 0.610 0.571 0.645 2.843% 2.435% 3.457% 2.908% 2.616% 3.756% 5.276% 4.541% 6.233% 6.614% 3.017% 8.904% 0.026% 0.021% 0.029% 0.129% 0.020% 0.166% 8.610% 7.349% 9.804% 21.129% 20.045% 22.584% 16.659% 15.548% 18.077% -0.008% -0.713% 0.701% 5.236% 4.173% 6.332% 0.574 0.516 0.637 JD 1075 15.301% 14.901% 15.629% SV 1035 9.202% 8.847% 9.506% Put Options SVJ SVJJ 909 801 13.958% 13.379% 14.463% 3.820 3.651 3.979 22.917% 22.357% 23.370% 141.775% 135.643% 150.883% 0.085 0.059 0.103 10.711% 10.401% 11.212% 3.256 2.862 3.574 16.205% 15.409% 17.029% 51.533% 46.765% 55.409% -0.229 -0.280 -0.154 0.446 0.380 0.532

5.1.1. Fit to Call Options Let us first look at each models performance on call options. Comparing the results for BS and JD models, we can see that the addition of jumps in price adds relatively little explanatory power. Looking at the parameters of the jumps in price process for the JD model, we see that they are very small and relatively insignificant.

Looking at the performance of the SV, SVJ and SVJJ models on call options, it is clear that the SVJ model performs worse than the SV model, though the SVJJ model performs the best among all models considered. The higher mean squared error per option of the SVJ model indicates that the additional feature of the SVJ model (jumps in prices) over the SV model does not add any explanatory power. This further confirms that jumps in price are not very relevant for call option pricing. This is again strengthened by the observation of small and relatively insignificant jump in price process parameters for the SVJ model.

Turning to the SVJJ model, it can be seen that adding jumps in volatility does seem to add explanatory power, with a substantial decrease in mean-squared error over the SV model (and SVJ model). It should also be noted that the characteristics of the jumps in price process have changed, with the mean jump size and jump size volatility both increasing and becoming more stable. Thus this may suggest that jumps in price may add some explanatory power when combined with jumps in volatility. These results, while they only show improvement for call options, suggest that a model with independent jumps in prices and volatility may be a better characterisation than the simultaneous jump process used in Eraker (2004).

In general, we can summarise the results of the fit of each model to call options as follows. Firstly, it is important to model stochastic volatility in order to improve upon the BS model. Secondly, jumps in prices seem to add very little explanatory power for call options. Lastly, the best fit can be achieved by a stochastic volatility model that includes both jumps in price and volatility, though the degree of the explanatory power of this model provided by jumps in prices cannot be determined.

5.1.2. Fit to Put Options Looking at the results for put options, we can see immediately the behaviour of the put option implicit distribution is quite different to that of call options. It should also be noted that the best performing model for put options (the SVJ model) actually performs better than the best performing model for call options (the SVJJ model). This indicates that while put option implicit distributions seem to depart further from (log) normality than call option implicit distributions, their features are better dealt with by the models studied here.

If we first note the performance on put options of the JD model, compared to the BS model, we find that there is a substantial increase in fit. If we then compare this improvement across option types we see that the JD model reduces mean squared error per option by four times for put options, compared to only about 10% for call options. This suggests the relevance of jumps in prices for put option pricing.

For put options, the JD model also has larger, significant parameters suggesting the better fit of jumps in prices model. This indicates that there are important features of the implicit distribution of put options that are effectively dealt with using the JD model that dont appear to exist in implicit call option distributions. Considering the SV models performance on put options, the conclusion that modelling stochastic volatility is of first order importance is confirmed, with the SV model having a substantially lower mean squared error per option than the JD model. If we then compare the SV model to the SVJ model we see a further increase in fit, contrary to what was observed for call options. This increase in fit however, is expected since we have observed that jumps in prices appear to be very relevant in fitting the distribution implicit in put options through the performance of the JD model.

Looking finally at the SVJJ model, we again observe a model with more free parameters performing worse than one of its smaller nested models (the SVJ model). We can then say that this poorer performance compared to the SVJ model suggests that jumps in volatility are not relevant for fitting put option implicit distributions. This result is in line with what is found by

Eraker (2004) for his simultaneous double jump model using a combination of call and put option. This could be explained by the fact that put options are generally traded more actively and may have dominated the sample used by Eraker (2004).

5.1.3. Fit of the Models Assuming Constant Parameters

Table 3 shows the mean squared error per option for each model for each data set. It is immediately apparent, due to the much higher errors, that all the models significantly violate the assumption of constant or near-constant parameter values. By this measure, the SV model is the best choice for call options and the JD model is best for put options.

Considering the relevance of different features for call and put options, we see that stochastic volatility appears to be most relevant for call options, with the SV model having a much lower error than the JD model. Jumps and stochastic volatility both appear important for put options with similarly low errors, though when combined it appears there is some degree of over fitting. Studying jumps in volatility, we see that the SVJJ model error for puts is much larger than for calls, suggesting the relevance of jumps in volatility for call options but not put options. Their much higher errors suggest, however, that there is a serious problem of over-fitting a changing implicit distribution. One may note that these conclusions are generally the same as reached earlier using different parameters for each day.

We see that while the fit of all the models has decreased substantially, the decrease becomes worse as the number of parameters increases. The SVJJ model, the best performing model for call options and second best performing model for put options using daily estimation, is now the worst performing model for both data sets (performing even worse than the BS model). This result could be interpreted either as over fitting by models with more free parameters or as clear evidence of parameters that are non-constant through time. The distinction between the two is difficult since the two hypotheses are hard to separate. The important result to note is that this is clear evidence of misspecification in all models.

Table 4. Near-Constant Parameter Option Pricing Model Fit This table shows the mean squared error per option for each model for each data set (puts and calls). The mean squared error for a model is calculated as the total (over all days) sum squared error divided by the total (over all days) number of options used in the calculations. BS Calls - MSE Puts - MSE 36.81 pts 88.97 pts JD 35.4 pts 34.09 pts SV 20.99 pts 34.62 pts SVJ 24.82 pts 39.95 pts SVJJ 51.28 pts 90.56 pts

5.1.4. Summary of Fit The ranking, from poorest fit to best fit for call options is: BS, JD, SVJ, SV and SVJJ. For put options, the ordering is different, being BS, JD, SV, SVJJ and SVJ. Overall, it appears that put option implicit distributions are best modelled by a process that allows firstly for stochastic volatility, and secondly, but also importantly, jumps in price. Jumps in volatility do not appear to be relevant.

This contrasts with call option implicit distributions, which are best modelled with a process that includes stochastic volatility and jumps in prices and volatility. These differences in fit underline the fundamental differences in the implicit distributions of call and put options, which will be discussed in more detail in the following section.

Under the assumption of constant (or slow-changing) parameters, an SV model is best for call options, while the either an SV or JD model perform equally well for put options. Significant differences again exist between call and put options and there is evidence of over-fitting by the more complex models.

While the implicit distributions of call and put options are clearly different, this is not positive evidence that call and put options behave differently. The characteristics of the data set are such that the average moneyness and maturity is different for call and put options. Due to this, any

differences between put and call implicit distributions could be due to some biases of the models used that are related to these contract-specific variables.

Simple regression analysis of the errors of the models, similar to Bakshi et al (1997), indicates exercise price and a maturity related biases for all models under consideration. As such, we cannot distinguish between whether the implicit distributions of calls and puts differ or the models are significantly misspecified.

6. Concluding Remarks This paper has studied four option pricing models from the class of affine models, as well as the Black-Scholes model. With the exception of the BS model, which is available in closed-form, we used tractable pricing formulas for each of the models in semi-closed form. These formulas are based on transform and inversion techniques such as those developed by Bakshi and Madan (2000) and Duffie et al (2000). This paper is, to the authors best knowledge, the first application of an SVJJ model with independent jumpsxxxvi. We further presented an efficient method for numerically evaluating the integrals associated with such option pricing formulas.

We take the approach of estimating the risk neutral model parameters and level of implicit volatility using the daily cross-section of traded FTSE/JSE Top 40 index options as has been done inter alia by Bakshi et al (1997) and Pan (2002). This not only allows us to summarise the risk neutral distribution of underlying price changes (and hence the option pricing formula) for each model by taking an average of the daily estimates, but also to study changes in the implicit distribution of underlying returns over time.

Studying the internal consistency of the models by looking at the time series of implicit volatilities and implicit volatility smile, we find that all the models are still significantly misspecified with regard to both option exercise price and maturity, and inconsistency through time. This misspecification, however, decreases as the model under consideration incorporates more features. We also conclude that the mean reverting square-root diffusion process for stochastic volatility is significantly misspecified.

Even given this inherent misspecification in all option pricing models, a choice among available alternatives must be made. In studying the utility of each extra feature of an option pricing model we find that it is stochastic volatility that is of first order importance in improving over the BS model. We find that the best fitting model depends on the type of option being priced (either put or call) and that the risk neutral distributions implicit in call options are substantially different to that of put options. We find that jumps in price are especially relevant for put options, with the best model being the SVJ model, while jumps in volatility are required to fit call options with the best choice being an SVJJ.