Appendix B. Prior Sensitivity Analysis in MCMC Method
Sensitivity is examined by considering four possible prior distributions—namely, normal, beta, exponential, and gamma—by either modifying the mean or/and increasing the variances relative to our initial prior. These prior conditions are listed in
Table A2, including narrowed variation and more variation. Following
Nugroho and Morimoto (
2015), to sample the
b parameter, which has a range of
to 1, we define
, where
; hence, the prior on
b is
with a prior mean of
.
The selection of normal and beta priors follows established Bayesian practice for financial volatility models. Normal distributions serve as weakly informative priors for unbounded parameters (e.g.,
in H
and RJP
), where large variances (e.g.,
) ensure the data dominate posterior inference. For bounded parameters (e.g.,
in A
), beta priors provide natural conjugacy and flexibility:
represents a uniform prior, while
gently centres mass near
without being restrictive. These choices align with recommendations for the RealG model (
Nakajima & Omori, 2012) and ensure priors are sufficiently diffuse to avoid unintended information while preventing numerical instability in tail regions (e.g.,
).
Table A2.
Alternativeprior distributions, other than the normal distribution, with hyperparameters on ST parameters.
Table A2.
Alternativeprior distributions, other than the normal distribution, with hyperparameters on ST parameters.
| | | | Exp | Exp |
Mean | 0.5 | 0.5 | | 4 | 10 |
SD | 0.1 | 0.29 | 0.14 | 4 | 14.1 |
| Exp | | | | |
Mean | 10 | 20 | 40 | 1 | 20 |
SD | 141.4 | 5 | 8.2 | 10 | 25.8 |
Table A3 and
Table A4 present the parameter estimates (posterior mean and standard deviation), IAT for the ST parameters, and log-likelihood values. Notice that given the specified priors, the estimates for the other parameters turn out to be nearly identical, except
. In all cases, the estimates for the skewness parameter (
) and shape parameters
are not affected by changing the prior for the parameter itself or the others. As a consequence, changing the priors for those parameters does not affect the other parameter(s) in each distribution, and vice-versa.
The degree-of-freedom parameters of the A distribution; degree-of-freedom parameter () of E, H, MP, and RJP; zero-location and unit-scale parameters of JF distribution; and parameter of EHL can be greatly influenced by the type of prior itself, as well as by the amount of variance. Under the same prior, the posterior standard deviation becomes larger as the dispersion of the prior increases.
Excluding the JF
and EHL
specifications, the above results indicate that the skewness and heavy-tailed parameters are independent of each other. This finding is consistent with the findings of an earlier study. For example,
Leisen et al. (
2017) argues the heaviness of the left (or right) tail of A
is controlled by
(or
) only. For EHL
and JF
distributions, notice that the behaviour of the tail is simultaneously reflected by the
and
parameters, respectively. Thus, our finding on parameters the
parameters is consistent with the result reported by
Adubisi et al. (
2022), who argue that the skewness and kurtosis values of EHL
are not affected by
.
Table A3.
Prior sensitivity analysis for the ST parameters in A, EHL, EST, and HST distributions.
Table A3.
Prior sensitivity analysis for the ST parameters in A, EHL, EST, and HST distributions.
A |
∼ | | | | | | |
∼ | | | Exp | Exp | | |
| 0.517 | 0.516 | 0.515 | 0.514 | 0.518 | 0.522 |
| 0.01 | 0.01 | 0.01 | 0.01 | 0.01 | 0.01 |
| 8.4 | 14.1 | 9.9 | 16.6 | 8.7 | 10.0 |
| 7.6 | 8.8 | 8.4 | 9.3 | 12.1 | 21.7 |
| 1.1 | 1.6 | 1.4 | 1.7 | 2.4 | 6.0 |
| 96.6 | 57.9 | 49.3 | 40.6 | 30.9 | 39.7 |
| 12.4 | 21.5 | 20.9 | 39.2 | 24.2 | 39.7 |
| 1.6 | 4.6 | 5.3 | 14.4 | 4.5 | 7.9 |
| 11.5 | 23.4 | 29.1 | 76.7 | 19.2 | 31.3 |
| 4288.4 | 4284.2 | 4284.0 | 4283.3 | 4284.7 | 4290.5 |
E |
∼ | | | | | | |
∼ | | | Exp | Exp | | |
| 0.044 | 0.046 | 0.051 | 0.053 | 0.047 | 0.049 |
| 0.009 | 0.009 | 0.009 | 0.009 | 0.009 | 0.009 |
| 5.6 | 5.3 | 8.9 | 7.6 | 6.5 | 6.2 |
| 10.7 | 14.2 | 14.1 | 16.0 | 16.4 | 27.2 |
| 1.2 | 2.5 | 2.5 | 3.4 | 2.9 | 6.1 |
| 8.1 | 17.0 | 10.9 | 11.9 | 16.8 | 33.2 |
| 4292.0 | 4290.2 | 4290.0 | 4290.4 | 4289.9 | 4293.8 |
EHL |
∼ | | | Exp | Exp | | |
∼ | | | Exp | Exp | | Exp |
∼ | | | Exp | Exp | | |
| 3.2 | 3.2 | 3.2 | 3.2 | 3.5 | 3.2 |
| 0.099 | 0.144 | 0.121 | 0.109 | 0.112 | 0.096 |
| 120.3 | 192.5 | 151.5 | 107.5 | 190.5 | 161.9 |
| 3.0 | 3.0 | 3.0 | 3.0 | 3.7 | 3.0 |
| 0.192 | 0.285 | 0.234 | 0.215 | 0.275 | 0.183 |
| 122.5 | 193.2 | 153.6 | 105.2 | 201.7 | 154.7 |
| 5.2 | 9.3 | 6.0 | 5.1 | 8.2 | 5.0 |
| 0.626 | 1.777 | 0.632 | 0.491 | 0.810 | 0.517 |
| 283.4 | 250.7 | 160.1 | 110.5 | 193.5 | 218.9 |
| 4314.6 | 4314.0 | 4314.7 | 4314.9 | 4317.1 | 4314.5 |
H |
∼ | | | | | | |
∼ | | | Exp | Exp | | |
| | | | | | |
| 0.02 | 0.02 | 0.02 | 0.02 | 0.02 | 0.02 |
| 5.3 | 7.2 | 5.8 | 5.5 | 5.2 | 10.2 |
| 10.8 | 13.7 | 13.3 | 15.5 | 16.5 | 25.7 |
| 1.3 | 2.5 | 2.2 | 3.2 | 2.9 | 5.4 |
| 7.9 | 17.2 | 9.2 | 15.1 | 12.4 | 21.9 |
| 4285.9 | 4284.1 | 4284.2 | 4284.0 | 4284.3 | 4287.2 |
Next, we asses the impact of a change in prior on the log likelihood. Excluding the EHL and JF distributions, the smallest log-likelihood value is obtained when the prior of is or , which has a standard deviation greater than 8. In these cases, the estimate of has a larger mean and standard deviation. When the prior of is , which has a standard deviation of 5, the log-likelihood values are competitive with those produced by other priors. For the MP and RJP distributions, all priors yield almost the same log likelihoods. Overall, very close log-likelihood values are produced by priors , Exp, Exp, and . In these cases, the maximum log likelihood may be achieved.
Table A4.
Prior sensitivity analysis for the ST parameters in JF, MP, and RJP distributions.
Table A4.
Prior sensitivity analysis for the ST parameters in JF, MP, and RJP distributions.
JF |
∼ | | | Exp | Exp | | |
∼ | | | Exp | Exp | | |
| 4.2 | 5.3 | 5.7 | 5.9 | 7.5 | 5.6 |
| 0.28 | 0.703 | 0.587 | 0.697 | 0.469 | 0.511 |
| 179.9 | 237.7 | 260.2 | 281.8 | 243.7 | 245.7 |
| 5.4 | 6.4 | 5.8 | 6.1 | 7.7 | 5.8 |
| 0.286 | 0.716 | 0.597 | 0.709 | 0.476 | 0.520 |
| 177.5 | 237.2 | 259.5 | 282.2 | 244.9 | 245.6 |
| 4298.0 | 4295.2 | 4294.4 | 4294.7 | 4294.8 | 4294.5 |
MP |
∼ | | | Exp | Exp | | |
∼ | | | Exp | Exp | | |
| 0.956 | 0.956 | 0.956 | 0.955 | 0.955 | 0.955 |
| 0.009 | 0.009 | 0.009 | 0.009 | 0.009 | 0.009 |
| 7.1 | 6.2 | 6.8 | 7.9 | 6.8 | 5.8 |
d | 2.0 | 1.9 | 1.9 | 1.7 | 1.8 | 1.8 |
| 0.099 | 0.102 | 0.105 | 0.078 | 0.047 | 0.092 |
| 198.3 | 215.2 | 127.2 | 247.8 | 115.6 | 168.9 |
| 6.6 | 11.5 | 8.9 | 43.1 | 19.5 | 20.2 |
| 1.7 | 4.7 | 3.5 | 30.5 | 4.9 | 14.9 |
| 58.9 | 63.6 | 16.3 | 319.3 | 27.8 | 229.8 |
| 4292.0 | 4291.7 | 4291.3 | 4291.8 | 4290.95 | 4291.9 |
RJP |
∼ | | | | | | |
∼ | | | Exp | Exp | | |
| | | | | | |
| 0.01 | 0.01 | 0.01 | 0.01 | 0.01 | 0.01 |
| 7.3 | 7.9 | 6.7 | 6.3 | 7.4 | 8.9 |
| 11.1 | 15.0 | 14.7 | 17.1 | 17.4 | 28.1 |
| 1.3 | 2.9 | 2.6 | 4.4 | 3.1 | 6.5 |
| 6.9 | 14.5 | 15.5 | 30.7 | 14.4 | 22.6 |
| 4290.4 | 4289.2 | 4289.3 | 4289.0 | 4288.9 | 4291.3 |
Appendix C. Sensitivity Analysis of Initial Values in GRG Method
An illustrative example of the suitability of the Excel Solver’s GRG in applying the LogRealG model with H distribution is outlined as follows.
Initializing the parameter values. The initial values of the model parameters are stored in D3:D14. These are changing variable cells. The value of each parameter cell will be changed by Solver interatively, at the same time, until the optimum has been found.
Defining the objective cells. We assign cell D17 to the objective function (total log likelihood in this case). Then, in cells D18:D21, we enter the function of AIC, ABIC, BIC, and CAIC.
Defining data cells. The historical data series (called return and log realized ) are stored in cells B26:D4528.
Initializing the volatility and measurement process. To implement volatility and measurement equations, the initial value () is set to be equal to the log of the variance of returns, and is set to be equal zero. We calculate , , , and in cells D25:G25 by placing the formulas expressed as “=LN(VAR(B26:B4528))”, “=EXP(D25)”, “0”, and “=$D$5*F25+$D$6*(F252-1)”, respectively.
Calculating the volatility and measurement process. At time point , processes (not including the exogenous variable), , , , and in cells D26:H26 are calculated by the formulas expressed as “=$D$3 + $D$4*D25 + G25”, “=EXP(D26)”, “=B26/SQRT(E26)”, “=$D$5*F26 + $D$6*(F262-1)”, and “=$D$10*F26 + $D$11*(F262-1)”, respectively. The formulas in E26:G26 can now be copied down to compute the complete processes at each time point. Meanwhile, process , for includes an exogenous variable. Starting from in cell D26 with the formula expressed as “=$D$3 + $D$4*D26 + G26 + $D$7*C26”; then, this formula can be copied down at each time point.
Calculating the log density function. We can first calculate the constant parts of the H
density function in Equation (
5)—namely,
,
a and
b. These parts are calculated in cells J16:J18 by the formulas expressed as “=GAMMA((D14+1)/2)/(GAMMA(D14/2)* SQRT(PI()*D14))”, “=4*D13*J17*(D14-2)/(D14-1)”, and “=SQRT(1 + 3*D13
2-J18
2)”, respectively. The part containing the condition is calculated in cell I26 by the formula expressed as “=-0.5*(
$D
$14+1)*IF(F26<-
$J
$18/
$J
$19,LN(1+1/
$D
$14*((
$J
$19* F26+
$J
$18)/(1-
$D
$13))
2),LN(1+1/
$D
$14*((
$J
$19*F26+
$J
$18)/(1+
$D
$13))
2))”; then, this formula can be copied down at each time point.
Calculating the log likelihood. Based on model (
1), the log-likelihood of return and measurement at time point
are calculated in cells J26:K26 by the formulas expressed as “=-0.5*LN(E26) + LN(
$J
$19) + LN(
$J
$17) + I26” and “=-0.5*LN(2*PI()*
$D
$12)-0.5*(C26-
$D
$8-
$D
$9* D26-H26)
2/
$D
$12”, respectively. The log likelihood for the fitted model is the sum of the two log likelihoods and is calculated in cell L26 by the formula expressed as “=J26 + K26”. Those formulas can now be copied down at each time point to obtain the total log likelihood for the fitted model stored in cell D17.
Applying the Solver. Open Solver window and define the problem. Applying the GRG method, the iterative calculation takes less than one minute.
Figure A1 displays an Excel screenshot showing the above process.
Figure A1.
Workbook for modelling of LogRealG with H distribution.
Figure A1.
Workbook for modelling of LogRealG with H distribution.
In this example, we illustrate the process of sensitivity analysis to see how various initial values of the ST distribution parameters would affect the estimated cells (for parameters and log likelihood). According to the parameter constraints, the analysis is performed by changing the initial value of one parameter and fixing the initial value of the other(s). Sensitivity is calculated by dividing the percentage change in the output variable by the percentage change in the input variable. In
Table A5 we can observe the results of our experiment. Observations are made on the existence of solutions, changes in parameter estimates, and changes in estimated log-likelihood values. This analysis helps determine the combination of a set of initial values on all ST parameters in the model.
First, considering the model with the A distribution, the initial values of , , and range from 0.1 to 0.9 at an increment of 0.1, those of range from 2 to 18 at an increment of 0.2, and those of range from 5 to 40 at an increment of 5. The result shows that the Solver could not find a feasible solution when the value of was initially set to 0.1, 0.2, 0.3, and 0.8 for fixed initial values of and . The estimates of were found to depend on its initial value and tend to increase as the initial value increases. In this case, the sensitivity of increased by 31%. The estimates of , and log likelihood do not appear to be affected by the initial values of each parameter. Hence, the initial values of can be used to compare the estimated values obtained by the ARWM method.
Second, the initial values of and in the model with the E distribution were set in increments of 0.1 using and in increments of 4 using . The result shows that the Solver could not find a feasible solution when the value of was initially set to 0.9 for the fixed initial values . The different initial values of had less of affect on the estimation of , with a sensitivity of 2.9%, and a great affect on the estimation of and log likelihood, with sensitivity greater than 100%. When the initial value of is fixed at 0, different initial values of have less of an affect on the estimation of and greatly affect the estimation of and log likelihood, with sensitivity values of greater than 35%. In addition, the maximum log likelihood was achieved with initial values of .
Table A5.
Sensitivity analysis for Excel’s Solver.
Table A5.
Sensitivity analysis for Excel’s Solver.
Initial Value | Estimated Value |
---|
Fixed Value | Changing Value | | | |
---|
A | | | | |
| | 0.514 | (9, 35) | |
| | 0.514 | (9, 26 to 37) | |
| | 0.514 | (9, 33 to 35) | |
E | | | | |
| | 0.038 to 0.087 | 14 to 24 | to |
| | 0.046 to 0.049 | 14 to 28 | to |
EHL | | | | |
| | (3.2, 2.9) | 3.8 | |
| | (3.2, 2.9) | 3.7 to 4.1 | to |
| | (3.2, 2.9) | 2.1 to 7.7 | to |
H | | | | |
| to 0.5 | | 14 | |
| to 40 | | 12 to 16 | to |
JF | | | | |
| | 6.2 | 6.4 | |
| | 6.0 to 6.3 | 6.2 to 6.5 | |
MP | | | | |
| | (1.8, 0.955) | 21 | |
| | (1.8, 0.955) | 15 to 22 | |
| | | | |
| | (1.6 to 2.0, 0.955) | 7 to 40 | to |
RJP | | | | |
| | | 14 to 16 | |
| | | 15 | |
Third, the initial value of each parameter in the model with the EHL distribution was varied from 1 to 10 in increments of 1. The Solver results are very sensitive to the initial values of , as indicated by the small number of choices of initial values for both parameters that provide feasible solutions. When the optimal solution is achieved, the estimates of the parameters are not affected by different initial values of each parameter. For the parameter, the estimated value is slightly less affected by different initial values of , with a sensitivity of 20%, and greatly affected by its own different initial values, with a sensitivity of 62%. Meanwhile, the estimated log likelihood is greatly affected by different initial values of , with sensitivity greater than 60%. In this case, the estimates of and log likelihood tend to grow as the initial value of rises, with initial values of providing the maximum log likelihood.
Fourth, regarding the parameters of the model with the H distribution, taking initial values of ranging from to in increments of 0.1 results in considerable sensitivity to feasible solutions, since these are only provided by values within the range of to . Such values do not affect the estimation of the parameters and log likelihood. Taking initial values of ranging from to in increments of 0.1 results in considerable sensitivity to feasible solutions, as these are only provided by values ranging from to 0.5. Such values do not affect on the estimation of the parameters and log likelihood. Meanwhile, changing the initial values of in increments of 4 in the range of does not affect on the estimation of and log likelihood, but it has less of an effect on the estimation of log likelihood (), with sensitivity values of less than 12%. Hence, we can choose initial values of to maximize the log-likelihood value.
Fifth, in the case of the model with the JF distribution, there are several combinations of initial values of that do not yields feasible solutions—namely (6, 1), (2, 7), and (2, 8)—when the initial values are set in increments of 1 using . These different values do not appear to affect the parameter and log-likelihood estimates. Thus, we can choose any combination that produces a feasible solution, such as .
Sixth, the optimal solution for the model with the MP distribution is not found for initial values of when changing by increments of 0.1 in the range of [0.1, 2]. The same situation also occurs when changing the initial value of d within the range 0.1 to 3 in increments of 0.1. When it has a solution, the parameter and log-likelihood estimates are not affected by different initial values of each parameter. Meanwhile, the estimated value of is greatly affected by different initial values of d and , where the initial value of changes in increments of 4 within the range of . It was found that the estimated value of increases as its initial value increases. Here, the maximizer of log likelihood is provided by the initial values of .
Finally, the model with RJP is considered. When the initial values of are changed by increments of 1 within the range of , the feasible solution is found in the range of to . These different values affect the estimation of only, with a sensitivity of 0.25%. Meanwhile, different initial values of in increments of 1 do not appear to affect the parameter and log-likelihood estimates. The two conditions produce very similar log-likelihood values, so we can choose as the initial values for .