Are VIX futures prices predictable?

An empirical investigation*

Eirini Konstantinidia and George Skiadopoulosb **

Department of Banking and Financial Management, University of Piraeus, Greece

Department of Banking and Financial Management, University of Piraeus, Greece, and Financial Options Research Centre, Warwick Business School, University of Warwick, UK

Abstract
This paper investigates whether volatility futures prices per se can be forecasted by studying the fast growing VIX futures market. To this end, alternative model specifications are employed. Point and interval out-of sample forecasts are constructed and evaluated under various statistical metrics. Next, the economic significance of the obtained forecasts is also assessed by performing trading strategies. Only weak evidence of statistically predictable patterns in the evolution of volatility futures prices is found. No trading strategy yields economically significant profits. Hence, the hypothesis that the VIX volatility futures market is informationally efficient cannot be rejected.

JEL Classification: C53, G10, G13, G14. Keywords: Interval forecasts, Market efficiency, Predictability, VIX, Volatility futures.

We would like to thank Ales Cerny, Alexandros Kostakis, Jolle Miffre, Allan Timmermann and Halbert White for useful discussions and comments. Any remaining errors are our responsibility alone. ** Corresponding author. Tel: +30-210-4142363; fax: +30-210-4142341. E-mail addresses: [email protected] (E. Konstantinidi), [email protected] (G. Skiadopoulos).

1.

Introduction

Volatility derivatives have attracted much attention over the past years since they enable trading and hedging against changes in volatility. Brenner and Galai (1989, 1993) first suggested derivatives written on some measure of volatility that would serve as the underlying asset. Since then, a number of volatility derivatives have been trading in the overthe-counter market. In March 26, 2004, volatility futures on the implied volatility index VIX were introduced by the Chicago Board Options Exchange (CBOE)1. Volatility futures on a number of other implied volatility indices have been also introduced since then. The liquidity of volatility futures markets is steadily growing, with the VIX futures market being the most liquid one2. This paper focuses on the VIX futures market and addresses for the first time the question whether VIX futures prices per se can be predicted3. Answering the question whether volatility futures prices can be predicted is of importance to both academics and practitioners. This is because it contributes to understanding whether volatility futures markets are efficient and helps market participants to develop profitable volatility trading strategies and set successful hedging schemes. There is already some extensive literature that has investigated whether the prices of stock index, interest rate, currency, and commodity futures can be forecasted. The significance of the results has been evaluated under either a statistical or economic (trading profits) metric. A number of studies have documented a statistically predictable pattern in futures returns. In particular, Bessembinder and Chan (1992) found that the monthly nearest maturity commodity and currency futures returns can be forecasted within sample in a statistical sense. They concluded that this predictability could be attributed to an asset pricing model with time-varying risk-premia. Similar findings were documented by Miffre (2001a) for the FTSE 100 futures and by Miffre (2001b) for commodity and financial futures. On the other hand, the empirical evidence on the predictability in futures markets

under an economic metric is mixed. For instance, Hartzmark (1987) found that in aggregate, speculators do not earn significant profits in commodity and interest rate futures markets; daily data of all contract maturities were employed. On the other hand, Yoo and Maddala (1991) studied commodity and currency futures and found that speculators tend to be profitable; daily data for a number of futures maturities were considered. Similar findings were reported by Taylor (1992), Kho (1996), Wang (2004) and Kearns and Manners (2004). In particular, all these studies found that economically significant profits can be obtained by employing various trading rules in currency futures markets; daily data were used by Taylor (1992), and weekly by Kho (1996), Wang (2004) and Kearns and Manners (2004). A number of futures maturities were examined by Taylor (1992) and Kearns and Manners (2004), while Kho (1996) and Wang (2004) focused on the shortest maturity series. Significant profits were also reported in Hartzmark (1991) and Miffre (2002) who examined the commodity and financial futures markets; the latter study focused only on the shortest maturity contracts. Regarding the source of the identified trading profits, Taylor (1992) and Kearns and Manners (2004) attributed them to the inefficiency of the currency futures market. On the other hand, Yoo and Maddala (1991), Kho (1996), Wang (2004) and Miffre (2002) found that the reported profits were not abnormal and Hartzmark (1991) found that profitability is determined by luck rather than superior forecast ability; hence, the considered markets were efficient la Jensen (1978). In contrast to the number of papers devoted to the topic of predictability in the previously mentioned futures markets, the research on whether there exist predictable patterns in the evolution of volatility futures prices is still at its infancy. The literature on volatility futures has primarily focused on developing pricing models (see e.g. Grnbichler & Longstaff, 1996; Zhang & Zhu, 2006; Dotsis et al., 2007; Brenner et al., 2008; Lin, 2008) and assessing their hedging performance (see e.g. Jiang & Oomen, 2001). On the other hand,

to the best of our knowledge, Konstantinidi et al. (2008) is the only related study that has explored the issue of predictability of volatility futures prices. However, this has been done indirectly and only under a financial metric. The authors developed trading strategies with VIX and VXD volatility futures based on point and interval forecasts that were formed for the corresponding underling implied volatility indices. They found that the obtained Sharpe ratios were not statistically different from zero and hence the volatility futures markets are efficient. This study extends the literature on whether the evolution of volatility futures prices can be forecasted. In contrast to Konstantinidi et al. (2008), we investigate the predictability of the VIX volatility futures prices per se without resorting to the underlying implied volatility index4. To this end, both point and interval out-of-sample forecasts are considered. This is because interval forecasts have been found to be useful for volatility trading purposes; Poon and Pope (2000) found that profitable volatility spread trades can be developed in the S&P 100 and S&P 500 index option markets by constructing certain intervals. We test the statistical significance of the obtained forecasts by a number of tests and criteria. In addition, their economic significance is investigated by means of trading strategies. This is the ultimate test to conclude whether the recently inaugurated volatility futures market is efficient. To check the robustness of our results, the analysis is performed across various maturity futures series and by employing a number of alternative model specifications. The latter is necessary since the question of predictability is tested inevitably jointly with the assumed forecasting model. The remainder of this paper is structured as follows. Section 2 describes the data set. Section 3 presents the forecasting models to be used. Section 4 discusses the results concerning the in-sample performance of the models under consideration. Next, the out-ofsample predictive performance of the various models is evaluated in statistical and economic

terms in Sections 5 and 6, respectively. The last section concludes.

2.

The Data Set

Daily settlement prices of CBOE VIX volatility futures and a set of economic variables are used. The sample period under consideration is from March 26, 2004 to March 13, 2008. The subset from March 18, 2005 to March 13, 2008 is used for the out-of-sample evaluation. VIX futures were listed in March 26, 2004 by the CBOE. These are exchange-traded futures contracts on volatility and may be used to trade and hedge volatility. The underlying asset of these contracts is VIX. The contract size is $1,000 times the VIX5. On any day, the CBOE Futures Exchange (CFE) may list for trading up to six near-term serial months and five months on the February quarterly cycle for the VIX futures contract. The VIX futures contracts are cash settled. The final settlement date is the Wednesday that is thirty days prior to the third Friday of the calendar month immediately following the month in which the contract expires. The VIX futures prices are obtained from the CBOE website. By ranking the data based on their time to expiration, three time series of futures prices are constructed; namely, the shortest, second shortest, and third shortest maturity series. To minimize the impact of noisy data, we roll over to the next maturity contract five days before the contract expires. For the same reason, settlement prices corresponding to a trading volume less than five contracts are excluded. The data set of economic variables consists of the return on the S&P 500 stock index, the Libor one-month continuously compounded interbank interest rate, and the slope of the yield curve, calculated as the difference between the prices of the ten-year U.S. government bond and the one-month continuously compounded interbank rate. These data are obtained from Datastream. 5

Figure 1 shows the evolution of the three maturity VIX futures series and the VIX index over the period from March 26, 2004 to March 17, 2005. The term structure of futures prices appears to be upward sloping, with prices being higher for longer maturities (see also Brenner et al., 2008). Table I shows the summary statistics for the three series of futures prices and the economic variables in levels and first differences (Panel A and B, respectively). All variables measured in levels are positively (first order) autocorrelated; this is not the case when they are measured in first differences. The ADF test indicates that most of the variables are non-stationary in the levels, but stationary in the first differences. The average volume decreases for longer maturities.

[Insert Figure 1 about here] [Insert Table I about here]

3.
3.1

The Forecasting Models

We use a set of lagged economic variables to forecast the evolution of futures prices. This model specification tests the semi-strong form efficiency (Fama, 1970, 1990) of the volatility futures market (see also Bessembinder & Chan, 1992; Miffre, 2001a, 2001b, 2002; Kearns & Manners, 2004; Konstantinidi et al., 2008, for applications of similar predictive specifications to futures markets). Based on the BIC criterion, the following regression is estimated:

Ft ,T = c + a1Ft 1,T + a2 Ft 2,T + a3 Rt 1 + a4 Rt 2 + a5it 1 + a6it 2 + a7 yst 1 + a8 yst 2 + t ,T (1)

yst the changes of the slope of the yield curve. The employed variables have been shown to have forecast power in equity markets (see e.g., Goyal & Welch, 2007) and hence they may also have predictive power in futures markets.

3.2

Univariate Autoregressive, ARIMA and VAR Models

Univariate autoregressive, ARIMA and VAR models are employed to investigate the extent to which past volatility futures prices can be exploited for predictive purposes and to examine whether there are spillovers between the three futures series. These model specifications set up tests of weak form market efficiency (Fama, 1970, 1991). The number of lags employed are chosen on the basis of the BIC criterion and to avoid over-fitting the data. The following AR(2) model is estimated: Ft ,T = c + 1Ft 1,T + 2 Ft 2,T + t ,T We estimate also an ARIMA(1,1,1) model: Ft ,T = c + 1Ft 1,T + 1 t 1 + t ,T Furthermore, the following VAR(1) model is estimated: Ft = C + 1Ft 1 + t (4) (3) (2)

where Ft is the ( 3 1) vector of changes in the three futures prices series that are assumed to be jointly determined, C is a

( 3 1)

vector of constants, 1 is a

( 3 3)

matrix of

coefficients and t is a ( 3 1) vector of residuals.

3.3

Principal Components Model

Principal components (PCs) extracted from Principal Component Analysis (PCA) are used as predictors in a regression setting. This model specification has been employed for forecasting purposes in various settings, such as to predict macroeconomic variables (see e.g., Stock &

Watson, 2002b; Artis et al., 2005, among others), the term structure of petroleum futures (Chantziara & Skiadopoulos, 2008) and implied volatility indices (Konstantinidi et al., 2008). Stock and Watson (2002a) showed that the PCs are consistent estimators of the true latent factors and that the forecasts based on PCs converge to those that would be obtained if the latent factors were known. Forecasting is performed in two steps. In the first step, we apply PCA to the block of daily changes in the three futures series and the economic variables. The first four principal components are retained; they explain 94% of the total variability of the system of variables employed. In the second step, the retained principal components are used as predictors in a regression setting and the following model specification is estimated:
Ft ,T = c1 + r1 j PC1t j + r2 j PC 2t j + r3 j PC 3t j + r4 j PC 4t j + t ,T
j =1 j =1 j =1 j =1 2 2 2 2

(5)

where rkj denotes the coefficient of the i-th PC lagged j times (k=1, 2, 3, 4 and j = 1,2). Note that this model specification has implications for the semi-strong form of efficiency of volatility futures markets, since volatility futures prices and economic variables are both employed to extract the PCs. Figure 2 shows the correlation loadings of the retained principal components. Focusing on the correlation loadings that show the impact of each PC to the daily changes of each one of the three futures series, we can see that almost all PCs affect the changes of futures prices by a similar magnitude across maturities. This implies that the respective shocks move the term structure of futures prices in a parallel way. The only exception is PC2 that seems to affect the steepness of the term structure of VIX futures.

[Insert Figure 2 about here]

3.4

Combination Forecasts

Apart from model based forecasts, we also consider combination forecasts. Combination forecasts aggregate the information used by the individual forecasting models. They have been found to be more accurate than individual forecasts (see e.g., Bates & Granger, 1969; Clemen, 1989, for a review). Two alternative linear combination forecasts are considered. First, an equally weighted combination forecast is employed:

1 FtU1,T = Ft|it 1,T |t 5 i =1

(6)

where Ft|it 1,T is the forecasted futures price constructed at time t-1 for t for a given maturity T by using the i-th model specification [i = 1 (economic variables model), 2 (AR(2) model), 3 (ARIMA(1,1,1) model), 4 (VAR model), 5 (PCA regression model)], and FtU1,T denotes the |t equally weighted combination forecast of the futures price constructed at time t-1 for t for a given maturity T. This is a simple average of all model based forecasts and is hereinafter referred to as unweighted combination forecast; there is evidence that simple combinations frequently outperform more sophisticated ones (see e.g., Clemen, 1989). Second, a weighted average of the individual forecasts is used; the weights are chosen so as to minimize the mean squared forecast error (see Granger & Ramanathan, 1984). To fix ideas, the weights are obtained by estimating the following OLS regression recursively: Ft ,T = c + ai Ft|it 1,T + t
i =1 5

(7)

where Ft ,T is the realized futures price change between time t-1 and t for a given maturity T, c is a constant, Ft|it 1,T = Ft|it 1,T Ft 1,T is the forecasted futures price change between time t-1 and t for a given maturity T and t is the error term. Then:

 FtW1|t ,T = Ft ,T + c + ai Ft i+1|t ,T +
i =1

(8)

where FtW1|t ,T denotes the weighted combination forecast of the futures price constructed at t + for t+1 for a given maturity T. To start constructing the weighted combination forecasts recursively, one needs an initial time series of individual forecasts to estimate regression (7). To this end, the insample data (from March 26, 2004 to March 17, 2005) are divided into a genuine insample period (from March 26, 2004 to September 24, 2004) and a pseudo out-of-sample period (from September 29, 2004 to March 17, 2005). First, the genuine in-sample data are used to estimate the model specifications described in Section 3.1-3.3 [equations (1), (2), (3), (4), and (5)]. Then, forecasts are formed recursively over the pseudo out-of-sample period by adding each observation of the pseudo out-of-sample data set to the genuine insample data set as it becomes available. Finally, the individual forecasts over the pseudo out-of-sample period are used to estimate regression (7) and the first out-of-sample weighted combination forecast (corresponding to March 18, 2005) is constructed as the fitted value of regression (7). To form the remaining out-of-sample combination forecasts equation (7) is estimated recursively by adding each individual forecast to the sample as it becomes available.

4.

In-Sample Evidence

Tables II, III and IV show the insample performance of the economic variables, AR(1)/ARIMA(1,1,1)/VAR and PCA regression models, respectively. The estimated coefficients, the t-statistics within parentheses and the adjusted R2 are reported for each one of the implied volatility indices, respectively. One and two asterisks indicate that the estimated parameters are statistically significant at 1% and 5% level, respectively. 10

In the case of the economic variables model [Table II], we can see that the adjusted R2 ranges from 1% to 3.2% across the three futures series. This is similar to the values of R2 documented by the previous related literature in various futures markets (see Bessembinder & Chan, 1992; Miffre, 2001a, 2001b, 2002; Konstantinidi et al., 2008). In the case of AR(2) and VAR models [Table III, Panel A and Panel B, respectively], we can see that the largest values of the adjusted R2 are obtained for the second shortest series (1.7% and 3%, respectively). The application of the ARIMA model [Table III, Panel C], reveals a strong predictable pattern in the case of the shortest futures series (adjusted R2 equals 5.1%). Finally, in the case of the PCA regression model [Table IV], the strongest pattern appears in the third shortest series (adjusted R2 equals 3.8%). To sum up, the in-sample goodness-of-fit depends on the model specification and the maturity of the futures series under consideration. Next, the out-of-sample performance is assessed so as to provide a firm answer to the question whether volatility futures prices can be forecasted. [Insert Table II about here] [Insert Table III about here] [Insert Table IV about here]

5.

Out-Of-Sample Evidence: Statistical Significance

Point and interval forecasts are used to assess the out-of-sample performance of the models described in Section 3. The out-of-sample period is from March 18, 2005 to March 13, 2008. To form the point forecasts, the models are initially estimated over the in-sample period (from March 26, 2004 to March 17, 2005) and the first out-of-sample point forecast is obtained (corresponding to March 18, 2005). To construct the remaining out-of-sample point forecasts, the models are re-estimated recursively by adding each observation to the in11

sample data set as it becomes available. The interval forecasts are formed by generating 10,000 simulation runs on each time step (i.e., day).

5.1

Point Forecasts: Statistical Testing

To assess the statistical significance of the obtained out-of-sample point forecasts, three alternative metrics are employed. The first metric is the root mean squared prediction error (RMSE) calculated as the square root of the average squared deviations of the actual volatility futures prices from the model based forecast, averaged over the number of observations. The second metric is the mean absolute prediction error (MAE) calculated as the average of the absolute differences between the actual volatility futures price and the model based forecast, averaged over the number of observations. The third metric is the mean correct prediction (MCP) of the direction of volatility futures price changes calculated as the average frequency (percentage of observations) for which the predicted by the model change in the volatility futures price has the same sign as the realized change. The forecasts are compared to those obtained from the random walk that is used as the benchmark model. To this end, we perform pairwise comparisons based on the modified Diebold-Mariano test (see Diebold & Mariano, 1995; Harvey et al., 1997) and a ratio test for the RMSE/MAE and MCP metrics, respectively. The null hypothesis is that the model under consideration and the random walk perform equally well. Moreover, we use Whites (2000) test (also termed reality check) to compare jointly all forecasts to the benchmark model under the RMSE and MAE metrics6. In this case, the null hypothesis is that no model outperforms the random walk. To fix ideas, the two tests are described as follows. Let

{F }
i t |t 1,T

t =1

and

{F }
RW t |t 1,T

t =1

denote the forecasted futures price based on the i-th model [i = 1 (economic

variables model), 2 (AR(2) model), 3 (VAR model), 4 (ARIMA(1,1,1) model), 5 (PCA

12

regression model), 6 (unweighted combination forecast), 7 (weighted combination forecast)] and the random walk, respectively. Define a loss function g ( eti,T ) and g ( etRW ) and the loss ,T
W differential dti,T = g ( eti,T ) g ( etRT ) , with {eti,T } ,
n

t =1

and {etRW } ,T

t =1

being the respective forecast

errors for the ith model specification and the T-maturity futures series. In the case of the modified Diebold-Mariano test, the null hypothesis is

H 0 : E ( dti,T ) = 0 . We test this against two alternative hypotheses. The first alternative hypothesis is that the random walk outperforms the respective model, i.e. H1 : E ( dti,T ) > 0 . The second alternative hypothesis is that the model under consideration outperforms the random walk, i.e. H 2 : E ( dti,T ) < 0 7. In the case of one-step ahead forecasts, the modified
i Diebold-Mariano test statistic MDM T for the ith model specification and the T-maturity

futures series is given by:

dTi var(dTi )

(9)

with dTi =

dti,T
t =1

and var(dTi ) =

(d
t =1

i t ,T

dTi )

n 1

. Following Harvey et al. (1997), we make

accept/reject decisions by comparing the calculated test statistic to the critical values from the Students t distribution with (n-1) degrees of freedom. The idea of Whites (2000) test is as follows see also Sullivan et al., 1999). At any point in time t the performance measure fi ,t ,T is defined for the i-th model and for a given maturity T : fi ,t ,T = dti,T (10)

So, the null hypothesis is H 0 : max E ( f i ,T ) 0 . The test statistic for the observed sample is:
i =1,..., k

13

V = max
n

i =1,..., k

{ n ( f )}
i ,T

(11)

where fi ,T = fi ,t ,T n , and n is the number of forecasts made. White (2000) suggests that
t =1

the null hypothesis can be evaluated by applying the stationary bootstrap of Politis and Romano (1994) to the observed values of fi ,t ,T 8. In particular, let B generated bootstrapped samples of fi ,t ,T . For each bootstrap sample, the following statistic is calculated:
V j = max

i =1,..., k

{ n( f

* i ,T , j

fi ,T )

(12)

where j = 1, 2, B. We choose B = 10,000. Whites (2000) reality check p-value is then obtained by comparing V and the obtained V j for j = 1, 2, B.

5.2

Interval Forecasts: Statistical Testing

Christoffersens (1998) likelihood ratio test of unconditional coverage is used to evaluate the constructed interval forecasts. The test can be applied for any assumed underlying stochastic process, since is not model dependent (Christoffersen, 1998). The idea of the test is as follows. A sample path, { Ft ,T }t =1 , of futures prices for a given maturity is observed and a
n

series of interval forecasts {( Lit / t 1,T (1 a ) , U ti/ t 1,T (1 a )}

T t =1

is constructed.

Lit / t 1,T (1 a )

and U ti/ t 1,T (1 a ) denote the lower and upper bound of an (1-a)%-interval forecast for time t constructed at t-1 for a given maturity contract based on the i-th model, respectively. We test whether the (1-a)%-interval forecast is efficient, i.e. whether the percentage of times that the realized future price at time t falls outside the interval forecast for time t constructed at time t-1 is a% for a given maturity. To this end, an indicator function I ti,T is defined:

14

i t ,T

0, if Ft,T Lit / t 1,T (1 a ) ,U ti/ t 1,T (1 a ) = i i 1, if Ft,T Lt / t 1,T (1 a ) ,U t / t 1,T (1 a )

(13)

Thus, the null hypothesis of an efficient (1-a)% interval forecast 0: E( I ti,T ) = is tested against the alternative 1: E( I ti,T ) . Under the null hypothesis, Christoffersens test statistic is given by a likelihood ratio test:

LRunc = 2 ln[ L(a) / L(a)] X 2 (1)

(14)

where n0 is the number of times that the futures prices falls within the constructed interval forecast (i.e. I t = 0 ), n1 is the number of times that the futures price falls outside the interval
forecast (i.e. I t = 1 ), a = n1 ( n0 + n1 ) is one minus the observed coverage probability,

L ( a ) = (1 a ) 0 a n1 is the likelihood under the null hypothesis and L ( a ) = (1 a ) 0 a n1 is the

likelihood under the alternative hypothesis. However, the power of this test may be sensitive to sample size. Hence, Monte Carlo simulated p-values are generated to assess the statistical significance of our results. We construct 99% and 95% interval forecasts to assess the robustness of the obtained results across different levels of significance.

5.3

Point and Interval Forecasts: Results

In the case of point forecasts, Table V shows the RMSE, MAE and MCP obtained for point forecasts based on the random, walk model (Panel A), the economic variables model (Panel B), the AR(2) model (Panel C), the VAR model (Panel D), the ARIMA(1,1,1) model (Panel E) and the PCA regression model (Panel F). Results for the unweighted and weighted combination of point forecasts (Panel G and H, respectively) are also reported. One and two asterisks (crosses) denote rejection of the null hypothesis in favor of the alternative H1 (H2) at significance levels 1% and 5%, respectively. We can see that there are 15 (out of 63 possible combinations in total) combinations of futures series and predictability metrics in which the
15

random walk beats one of the models (i.e., 24% of the cases). On the other hand, in 5 out of 63 cases (i.e., 8%) the model under consideration outperforms the random walk. All of these occur under the MCP measure and most of them are observed for the shortest series (4 out of 5). Note that under the assumption of independence of accept/reject decisions, one would expect the models to beat the random walk only in roughly 3 out of 63 cases (i.e., 5% of the cases) at a 5% significance level. Thus, there is weak evidence of a statistically predictable pattern in the evolution of the shortest futures series. In the case of Whites (2000) test, the reality check p-value for the RMSE (MAE) is 0.999 (0.916), 0.963 (0.998) and 0.888 (0.9532) for the shortest, second shortest and third shortest series, respectively. Thus, we accept the null hypothesis in all cases. This implies that even the best performing model specification under the RMSE (MAE) metric does not outperform the random walk.

[Insert Table V about here]

Regarding interval forecasts, Table VI shows the percentage of observations that fall outside the constructed 99% and 95%-interval forecasts, and Christoffersens (1998) test statistic value obtained by the economic variables model, the AR(2) model, the VAR model, the ARIMA(1,1,1) model, the PCA regression model, the unweighted combination interval forecasts and the weighted combination interval forecasts (Panels A, B, C, D, E, F and G, respectively); results are reported for each one of the three futures series. One and two asterisks denote rejection of the null hypothesis at 1% and 5% significance levels, respectively. We can see that the null hypothesis of efficient interval forecasts is rejected in all instances. This holds for both the 99% and 95% interval forecasts.

16

[Insert Table VI about here]

6.

Out-of-Sample Evidence: Economic Significance

The previously reported results on point forecasts suggest that there is a weak evidence of a statistically predictable pattern in the evolution of the shortest futures series based on the modified Diebold-Mariano test. Moreover, none of the constructed 99% and 95%-interval forecasts were found to be efficient. To provide a definite answer on the issue of predictability in volatility futures markets, the economic significance of the obtained forecasts is assessed by performing trading strategies based on point and interval forecasts. The trading strategies are performed despite the fact that there is no evidence of a statistically predictable pattern. This is because the statistical evidence does not always corroborate a financial criterion (see e.g. Ferson et al., 2003). The trading strategies involve a single volatility futures contract. Transaction costs have been taken into account; the standard transaction fee in the VIX futures market is 0.5$ per transaction.

6.1

Testing for Economic Significance: Measures

The profitability of the trading strategies is evaluated in terms of the Sharpe Ratio (SR), Lelands (1999) alpha (Ap) and their bootstrapped 95% confidence intervals9. The bootstrap samples have been generated under the null hypothesis of a zero SR and Ap, respectively. The continuously compounded one month Libor rate is used as the risk free rate in the calculation of both measures of performance. Lelands (1999) alpha is employed in order to account for the presence of nonnormality in the distribution of the trading strategys returns. It is based on the asset pricing model of Rubinstein (1976) and Breeden and Litzenberger (1978), who assumed that the

17

returns on the market portfolio are i.i.d. at any point in time and that markets are perfect. Under these assumptions, the equilibrium expected returns satisfy the following single period relationship:

E ( rp ) = rf + B p E ( rmkt ) rf

(15)

where rp is the return on the trading strategy, rf is the risk-free rate of interest, rmkt is the

return on the market portfolio, B p

( = cov ( r

cov rp , (1 + rmkt )
mkt

, (1 + rmkt )

) is a measure of risk similar to the )

CAPMs beta and =

ln E (1 + rmkt ) ln (1 + rf ) var ln (1 + rmkt )

is a measure of risk aversion. Then,

Lelands (1999) alpha is defined as:

Ap = E ( rp ) B p E ( rmkt ) rf rf

(16)

A two step procedure is employed to calculate Lelands (1999) alpha. First, and

Bstr are computed for each time step. We use the one month continuously compounded Libor
rate and the return on the S&P 500 as proxies for the rf and rmkt , respectively. Second, the following regression is estimated:

rpi ,t B ip ,t rmkt ,t rf ,t rf ,t = Aip + t

(17)

where rpi ,t and Aip are the return on the trading strategy and Lelands (1999) alpha, respectively, that are based on the forecasts from the i-th model [i = 1 (economic variables model), 2 (AR(2) model), 3 (VAR model), 4 (ARIMA(1,1,1) model), 5 (PCA regression model), 6 (unweighted combination forecast), 7 (weighted combination forecast)]. If Aip > 0 then we conclude that the trading strategy offers an expected return in excess of its equilibrium risk adjusted level.

18

6.2

Trading Strategy and Results Based on Point Forecasts

The economic significance of the constructed point forecasts is evaluated in terms of the following trading rule: If Ft 1,T < (>) Ft|it 1,T , then go long (short). If Ft 1,T = Ft|it 1,T , then do nothing. The rational of this trading rule is as follows: If the current futures price is higher (lower) than the forecasted futures price, then the price is anticipated to decrease (increase) and the investor goes short (long). If the current futures price is equal to the forecasted futures price, then the investor takes no action and maintains his/her position. Table VII shows the annualised SR, Ap, and their respective bootstrapped 95% confidence intervals (95% CI) for the three VIX futures series. Results are reported for trading strategy based on point forecasts derived by the economic variables model (Panel A), the AR(2) model (Panel B), the VAR model (Panel C), the ARIMA(1,1,1) model (Panel D), the PCA regression model (Panel E) and the weighted (Panel F) and unweighted (Panel G) combination point forecasts. We can see that the SR and Ap are insignificant in all cases. This implies that all trading strategies based on point forecasts do not yield economically significant profits. The results are similar to those obtained for a nave buy and hold strategy in VIX futures that yields a SR equal to 0.0178 [95% CI = (-0.06, 0.09)] for the shortest series, 0.0421 [95% CI = (-0.03, 0.11)] for the second shortest series, and 0.0532 [95% CI = 0.02, 0.13)] for the third shortest series.

[Insert Table VII about here]

6.3

Trading Strategy and Results Based on Interval Forecasts

19

The economic significance of the constructed interval forecasts is evaluated in terms of the following trading rule: If Ft 1,T < (>)

U ti|t 1,T (1 ) + Lit|t 1,T (1 )

, then go long (short).

If Ft 1,T =

U ti|t 1,T (1 ) + Lit|t 1,T (1 )

, then do nothing.

The rational behind this trading rule is as follows: If the futures price is closer to the lower (upper) bound of next days interval forecasts, then we anticipate the index price to increase (decrease) and as a result the investor should go long (short). Table VIII shows the annualised SR, Ap, and their respective bootstrapped 95% confidence intervals (95% CI) for the three VIX futures series. Results are reported for the trading strategy based on 99% and 95%-interval forecasts derived by the economic variables model (Panel A), the AR(2) model (Panel B), the VAR model (Panel C), the ARIMA model (Panel D), the PCA regression model (Panel E) and the weighted (Panel F) and unweighted (Panel G) combination interval forecasts. We can see that the results are similar for the strategies based on the 99% and 95%-interval forecasts. In particular, the SR and Ap are insignificant in all but one cases. This means that overall, the trading strategies based on interval forecasts do not yield significant profits, just as was the case with the trading strategies based on point forecasts. The results are similar to those obtained for a nave buy and hold strategy in VIX volatility futures. In particular ,the SR equals 0.0178 [95% CI = (0.06, 0.09)] for the shortest series, 0.0421 [95% CI = (-0.03, 0.11)] for the second shortest series, and 0.0532 [95% CI = -0.02, 0.13)] for the third shortest series.

[Insert Table VIII about here]

20

7. Conclusions
This paper has investigated for the first time whether the volatility futures prices per se can be forecasted. To this end, the most liquid volatility futures market (futures on VIX) has been considered. A number of alternative model specifications have been employed: the economic variables model, the AR(2) model, the VAR model, the ARIMA(1,1,1) model and the PCA regression model. Weighted and unweighted combination forecasts have also been considered. Point and interval forecasts have been constructed and their statistical and economic significance has been evaluated. The latter is assessed by means of trading strategies using the VIX futures. This has implications for the efficiency of the VIX volatility futures market. Regarding the statistical significance of the obtained forecasts, in the case of point forecasts, we found weak evidence of a statistically predictable pattern in the evolution of the shortest futures series. In the case of interval forecasts no model specification had predictive power. Regarding the economic significance of the obtained forecasts, the constructed forecasts did not yield economically significant profits. Overall, our results imply that one cannot reject the hypothesis that the VIX volatility futures market is informationally efficient. These findings are consistent with Konstantinidi et al. (2008), who had studied the efficiency of the VIX futures market indirectly. On the other hand, our results are in contrast to those found about the efficiency of other futures markets (stock, currency, interest rate and commodities) where predictability in either statistical or economic terms has been documented. The fact that the VIX futures market is found to be efficient does not invalidate the trading of VIX futures though. This is because VIX futures can also be used for hedging against changes in volatility. After all, this was the main motivation for their introduction (see Brenner & Galai, 1989, 1993).

21

Future research should investigate the issue of predictability in volatility futures markets at longer horizons. It has been well documented that the predictability in asset returns increases as the horizon increases (see e.g., Fama & French, 1988a, 1988b; Poterba & Summers, 1988). However, a longer horizon study is beyond the scope of this paper due to data limitations, as the VIX market operates only since 2004. Inta-day data should also be used to test whether any predictable patterns may be detected within the day; this will be particularly useful for scalpers. Finally, it may be worth considering more complex model specifications given that the answer on the predictability question always depends on the assumed specification of the predictive regression.

Bibliography
Artis, M.J., Banerjee, A. & Marcellino, M. (2005). Factor forecasts for the UK. Journal of Forecasting, 24, 279-298. Bates, J.M. & Granger, C.W.J. (1969). The combination of forecasts. Operational Research Quarterly, 20, 451-468. Bessembinder, H. & Chan, K. (1992). Time varying risk premia and forecastable returns in futures markets. Journal of Financial Economics, 32, 169-193. Breeden, D. & Litzenberger, R. (1978). Prices of state contingent claims implicit in option Price. Journal of Business, 51, 621-652. Brenner, M., & Galai, D. (1989). New financial instruments for hedging changes in volatility, Financial Analysts Journal, 45, 61-65. Brenner, M., & Galai, D. (1993). Hedging Volatility in Foreign Currencies. Journal of Derivatives, 1, 53-59. Brenner, M., Shu, J. & Zhang, J.E. (2008). The market for volatility trading; VIX futures. Working Paper, New York University, Stern School of Business. 22

Chantziara, T. & Skiadopoulos, G. (2008). Can the dynamics of the term structure of petroleum futures be forecasted? Evidence from major markets. Energy Economics, 30, 962-985. Christoffersen, P. F. (1998). Evaluating interval forecasts. International Economic Review, 39, 841-862. Clemen, R.T. (1989). Combining forecasts: A review and annotated bibliography. International Journal of Forecasting, 5, 559-583. Coppola, A. (2008). Forecasting oil price movements: Exploiting the information in the futures market. Journal of Futures Markets, 28, 34-56. Diebold, F.X. & Mariano, R.S. (1995). Comparing predictive accuracy. Journal of Business and Economic Statistics, 13, 253-263. Dotsis, G., Psychyios, D. & Skiadopoulos, G. (2007). An empirical comparison of continuous-time models of implied volatility indices. Journal of Banking and Finance, 31, 35843603. Fama, E.F. & French, K.R. (1988a). Dividend yields and expected stock returns. Journal of Financial Economics, 22, 3-25. Fama, E.F. & French, K.R. (1988b). Permanent and temporary components of stock prices. Journal of Political Economy, 96, 246-273. Goyal, A. & Welch, I., (2007). A comprehensive look at the empirical performance of equity premium prediction. Review of Financial Studies, forthcoming. Granger, C.W.J. & Ramanathan, R. (1984). Improved methods of combining forecasts. Journal of Forecasting, 3, 197-204. Grnbichler, A. & Longstaff, F.A. (1996). Valuing futures and options on volatility. Journal of Banking and Finance, 20, 985-1001. Hartzmark, M.L. (1987). Returns to individual traders of futures: Aggregate results. Journal

23

of Political Economy, 95, 1292-1306. Hartzmark, M.L. (1991). Luck versus forecast ability: Determinants of trader performance in futures markets. Journal of Business, 64, 49-74. Harvey, D.I., Leybourne, S.J., Newbold, P. (1997). Testing the equality of prediction mean squared errors. International Journal of Forecasting, 13, 281-291. Jensen, M.C. (1978). Some anomalous evidence regarding market efficiency. Journal of Financial Economics, 6, 95-101. Jiang, G. & Oomen, R. (2001). Hedging derivatives risks: A simulation study. Working Paper, University of Arizona. Jobson, J.D. & Korkie, B.M. (1981). Performance hypothesis testing with the Sharpe and Treynor measures. Journal of Finance, 36, 889-908. Kearns, J. & Manners, P. (2004). The profitability of speculators in currency futures markets. Working Paper, Reserve Bank of Australia. Kellard, N., Newbold, P., Rayner, T. & Ennew, C. (1999). The relative efficiency of commodity futures markets. Journal of Futures Markets, 19, 413-432. Kho, B.C. (1996). Time-varying risk premia, volatility, and technical trading rule profits: Evidence from foreign currency futures markets. Journal of Financial Economics, 41, 249-290. Konstantinidi , E., Skiadopoulos, G. & Tzagkaraki, A. (2008). Can the evolution of implied volatility be forecasted? Evidence from European and U.S. implied volatility indices. Journal of Banking and Finance, forthcoming. Leland, H.E., (1999). Beyond mean-variance: Performance measurement in a

nonsymmetrical world. Financial Analysts Journal, 55, 27-35. Lin, Y. (2007). Pricing VIX futures: Evidence from integrated physical and risk-neutral probability measures. Journal of Futures Markets, 27, 1175-1217.

24

Miffre, J. (2001a). Efficiency in the pricing of the FTSE 100 futures contracts. European Financial Management, 7, 9-22. Miffre, J. (2001b). Economic activity and time variation in expected futures returns. Economic Letters, 73, 73-79. Miffre, J. (2002). Economic significance of the predictable movements in futures returns. Economic Notes by Banca Monte dei Paschi di Siena SpA, 31, 125-142. Nossman, M. & Vilhelmsson, A. (2008). Is the VIX futures market able to predict the VIX index? A test of the expectations hypothesis. Working Paper, Lund University. Politis, D. & Romano, J. (1994). The stationary bootstrap. Journal of the American Statistical Association, 89, 1303-1313. Poon, S. H. & Pope, P.F., (2000). Trading volatility spreads: A test of index option market efficiency. European Financial Management Journal, 6, 235-260. Poterba, J. & L. Summers (1988). Mean reversion in stock returns: Evidence and implications. Journal of Financial Economics, 22, 27-60. Rubinstein, M. (1976). The valuation of uncertain income streams and the pricing of options. Bell Journal of Economics, 7, 407-425. Stock, J.H. & Watson, M.W. (2002a). Forecasting using principal component from a large number of predictors. Journal of the American Statistical Association, 97, 1167-1179. Stock, J.H. & Watson, M.W. (2002b). Macroeconomic forecasting using diffusion indexes. Journal of Business Economics and Statistics, 20, 147-162. Sullivan, R., Timmermann, A. & While, H. (1999). Data-snooping, technical trading rule performance, and the bootstrap. Journal of Finance, 54, 1647-1691. Taylor, S.J. (1985). The behavior of futures prices over time. Applied Economics, 17, 713734. Taylor, S.J. (1992). Rewards available to currency futures speculators: Compensation for risk

25

or evidence of inefficient pricing?. Economic Record, 68 (Supplement), 105-116. Wang, C. (2004). Futures trading activity and predictable foreign exchange market movements. Journal of Banking and Finance, 28, 1023-1041. Whaley, R.E. (2000). The investor fear gauge. Journal of Portfolio Management, 26, 1217. White, H. (2000). A reality check for data snooping. Econometrica, 68, 1097-1126. Yoo, J. & Maddala, G.S. (1991). Risk premia and price volatility in futures markets. Journal of Futures Markets, 11, 165-177. Zhang, J.E. & Zhu, Y. (2006). VIX futures. Journal of Futures Markets, 26, 521-531.

26

250

25

200 VIX Futures Prices

20

100

10

50

0
26/3/2004 17/5/2004 7/7/2004 25/8/2004 29/9/2004 11/11/2004 27/12/2004 VIX Index 11/2/2005 Shortest VIX Futures Series 3rd Shortest VIX Futures Series 2nd Shortest VIX Futures Series

Figure 1: Evolution of the three shortest maturity VIX futures series (left axis) and the VIX index (right axis) over the period March 26, 2004 to March 17, 2005.

0.8 0.6 0.4 0.2 0 Shortest -0.2 -0.4 -0.6 -0.8 -1 2nd Shortest 3rd Shortest i ys R

PC1

PC2

PC3

PC4

Figure 2: Correlation loadings of the first four principal components. Principal component analysis has been applied to the block of daily changes in the three futures series and the explanatory economic variables employed in equation (1) over the period March 26, 2004 to March 17, 2005.

27

VIX (%)

150

15

Panel A: Summary Statistics of VIX futures and Economic Variables (Levels): Mar 26, 2004 to Mar 17, 2005 Shortest # Observations Mean Std. Deviation Skewness Kurtosis Jarque-Bera 1 ADF Average Volume (min-max) 240 158.17 22.46 0.05 2.01 10.00* 0.921* -3.67** 186.17 (5 1,218) 2nd Shortest 3rd Shortest 234 169.59 24.13 -0.08 1.88 12.51* 0.912* -3.07 135.03 (5 865) 194 178.70 24.34 -0.23 1.86 12.28* 0.792* -2.75 104.16 (58 974) 1M interest rate Slope of yield curve 249 1.69 0.49 0.11 1.64 19.71* 0.966* -2.67 249 2.49 0.70 0.26 1.67 21.19* 0.973* -3.63** S&P 500 254 1143.79 41.18 0.26 1.82 17.47* 0.972* -1.41

Panel B: Summary Statistics of VIX futures and Economic Variables (Daily differences): Mar 26, 2004 to Mar 17, 2005

Shortest Mean Std. Deviation Skewness Kurtosis Jarque-Bera 1 ADF -0.243805 5.04 1.87 11.64 835.22* -0.01 -14.87*

2nd Shortest 3rd Shortest -0.261751 4.20 0.66 8.40 278.87* 0.08 -13.56* -0.461538 3.32 0.56 5.10 36.58* 0.04 -10.71*

1M interest rate Slope of yield curve 0.003561 0.00 1.26 7.18 239.89* 0.25* -6.77* -0.004924 0.05 0.18 5.53 65.88* -0.01 -15.63*

S&P 500 0.000284 0.01 -0.13 2.93 0.80 0.04 -15.18*

Table I: Summary statistics. Entries report the summary statistics for each VIX futures series and the economic variables. The economic variables under consideration are the one-month Libor interbank interest rate, and the slope of the yield curve (calculated as the difference between the prices of a ten-year U.S. government bond and the one-month interbank rate), and the S&P 500 stock index. The first order autocorrelation 1, the Jarque-Bera and the Augmented Dickey Fuller (ADF) test values are also reported. One asterisk denotes rejection of the null hypothesis at the 1% level. The null hypothesis for the Jarque-Bera and the ADF tests is that the series is normally distributed and has a unit root, respectively.

28

Included Obs. c Ft-1 Ft-2 Rt-1 Rt-2 it-1 it-2 yst-1 yst-2 Adj. R2

Dependent Variable: Shortest 194 Coeff. (t-stat) -0.605 (-1.071) -0.022 (-0.404) 0.013 (0.277) -74.828 (-1.493) 128.078 (1.930) 104.662 (1.000) -33.730 (-0.536) 5.745 (0.919) -3.175 (-0.578)
0.012

Dependent Variable: 2nd Shortest 181 Coeff. (t-stat) -0.593 (-1.286) 0.085 (1.136) -0.019 (-0.223) -31.932 (-0.782) 84.341 (1.548) 98.723 (1.211) -24.736 (-0.403) 3.595 (0.563) 1.312 (0.233)
0.010

Dependent Variable: 3rd Shortest 113 Coeff. (t-stat) -0.535 (-0.843) -0.158 (-1.466) -0.033 (-0.262) -132.828** (-2.244) 39.788 (0.694) 67.946 (0.783) 0.514 (0.006) 12.446 (1.738) -7.385 (-1.099)
0.032

Table II: Forecasting with the economic variables model: In-sample analysis. The entries report results from the regression of each VIX futures series on a set of lagged economic variables, augmented by an AR(2) term. The following specification is estimated: where

Ft ,T = c + a1Ft 1,T + a2 Ft 2,T + a3 Rt 1 + a4 Rt 2 + a5it 1 + a6it 2 + a7 yst 1 + a8 yst 2 + t ,T ,

Ft ,T : daily changes in the futures prices between time t-1 and t for a given maturity T, c: a constant,

Rt : the log-return on the S&P 500 stock index between time t-1 and t,

it : the one month

continuously compounded Libor rate in log-differences and yst : the changes of the slope of the yield curve calculated as the difference between the prices of the ten year U.S. government bond and the one month continuously compounded interbank rate. The estimated coefficients, Newey-West tstatistics in parentheses, and the adjusted R2 are reported. One and two asterisks denote rejection of the null hypothesis of a zero coefficient at the 1% and 5% level, respectively. The model has been estimated for the period March 26, 2004 to March 17, 2005.

29

Included Obs. c 1 2 Adj. R2 C F1t-1 F2t-1 F3t-1 Adj. R2 Included Obs. c 1 1 Adj. R2

Dependent Dependent Variable: Variable: Shortest 2nd Shortest Coeff. Coeff. (t-stat) (t-stat) Panel A: AR(2) Model 200 187 -0.396 -0.413 (-1.262) (-1.517) 0.028 0.116 (0.609) (1.926) -0.092 -0.119 (-1.692) (-1.465) 0.002 0.017 Panel B: VAR Model -0.464 -0.469 (-1.175) (-1.600) -0.162 -0.136 (-0.966) (-1.093) 0.415 0.307 (1.933) (1.928) -0.062 0.013 (-0.245) (0.070) 0.018 0.030 Panel C: ARIMA(1,1,1) Model 226 217 -0.373 -0.034* (-3.517) (-0.933) 0.864* -0.533** (26.370) (-2.002) -0.992* 0.643** (-161.249) (2.537) 0.051 0.009

Dependent Variable: 3rd Shortest Coeff. (t-stat) 118 -0.497 (-1.544) 0.015 (0.196) -0.079 (-0.904) -0.010 -0.285 (-0.907) -0.121 (-0.906) 0.352 (2.064) -0.148 (-0.739) 0.012 156 -0.068** (-2.238) 0.895* (18.308) -0.984* (-58.883) 0.021

Table III: Forecasting with the univariate autoregressive, ARIMA and VAR models: In-sample analysis. Panel A: The entries report results from the estimation of a univariate AR(2) specification for the daily changes of each VIX futures series, namely: Ft ,T = c + 1Ft 1,T + 2 Ft 2,T + t ,T . Panel B: The entries report the estimated coefficients of a VAR, for the three VIX futures prices series:

Ft = C + 1Ft 1 + t ,T , where Ft is the (7x1) vector of changes in the three futures prices series,
C is a (7x1) vector of constants, 1 is the (7x7) matrix of coefficients to be estimated, and t,T is a (7x1) vector of errors. Panel C: The entries report the estimated coefficients of a ARIMA(1,1,1) model: Ft ,T = c + 1Ft 1,T + 1 t 1 + t ,T . The estimated coefficients, Newey-West t-statistics in parentheses and the adjusted R2 are reported. One and two asterisks denote rejection of the null hypothesis of a zero coefficient at the 1% and 5% level, respectively. The models have been estimated for the period March 26, 2004 to March 17, 2005.

30

Included Obs. c PC1t-1 PC1t-2 PC2t-1 PC2t-2 PC3t-1 PC3t-2 PC4t-1 PC4t-2 Adj. R2

Dependent Variable: Shortest 116 Coeff. (t-stat) -0.005 (-0.011) 0.222 (0.606) -0.360 (-0.974) (-0.636) (-1.391) 0.121 (0.385) 0.445 (1.157) -0.234 (-0.584) 0.571 (1.632) 0.001 (0.003)
-0.007

Dependent Variable: 2nd Shortest 116 Coeff. (t-stat) -0.177 (-0.572) 0.352 (1.392) -0.372067 (-1.225) -0.181 (-0.602) -0.109772 (-0.385) 0.409 (1.182) -0.275 (-0.879) 0.421 (1.669) -0.160 (-0.520)
0.008

Dependent Variable: 3rd Shortest 107 Coeff. (t-stat) -0.278 (-0.777) 0.228 (1.038) -0.364211 (-1.169) -0.342 (-0.975) 0.322149 (1.090) 0.246 (0.656) -0.211 (-0.598) 0.797** (2.489) -0.116 (-0.356)
0.038

Table IV: Forecasting with the principal components analysis model: In-sample analysis. The entries report results from the regression of the futures price changes on the lagged first four principal components PC1, PC2, PC3 and PC4 derived from the set of the three VIX futures prices series and the economic variables: Ft ,T = c1 + r1 j PC1t j + r2 j PC 2t j + r3 j PC 3t j + r4 j PC 4t j + t ,T . The
j =1 j =1 j =1 j =1 2 2 2 2

estimated coefficients, Newey-West t-statistics in parentheses, and the adjusted R2 are reported. One and two asterisks denote rejection of the null hypothesis of a zero coefficient at the 1% and 5% level, respectively. The model has been estimated for the period March 26, 2004 to March 17, 2005.

31

Shortest 2nd Shortest 3rd Shortest Panel A: Random Walk 7.01 5.16 4.55 RMSE 4.32 3.31 2.89 MAE Panel B: Economic Variables Model Point Forecasts RMSE 7.40* 5.42* 4.90* 3.50 MAE 4.60* 3.17* 50.09% 50.75% MCP 54.07%+ Panel C: AR(2) Model Point Forecasts 7.26 5.37 4.81 RMSE 4.46 3.46 3.09 MAE 51.16% 52.45% MCP 54.78%++ Panel D: VAR Model Point Forecasts 7.60 5.49 4.77 RMSE 4.74 3.56 3.05 MAE 52.41% 50.26% MCP 53.71%+ Panel E: ARIMA(1,1,1) Model Point Forecasts 5.23 4.68 RMSE 7.16* 4.33 3.35 2.97 MAE 51.49% 52.26% MCP 55.29%++ Panel F: PCA Regression Model Point Forecasts 5.06 RMSE 8.16** 5.76* MAE 5.26** 3.86** 3.35** 49.71% 49.80% 49.70% MCP Panel G: Unweighted Combination Point Forecasts 7.13 5.24 4.68 RMSE 4.35 3.36 2.98 MAE + 51.33% 51.59% MCP 54.41% Panel H: Weighted Combination Point Forecasts RMSE 8.05** 5.76* 5.04** MAE 5.14 3.25 3.83** MCP 51.08% 47.42% 50.71%
Table V: Out-of-sample performance of the model specifications for each one of the VIX futures prices series. The root mean squared prediction error (RMSE), the mean absolute prediction error (MAE), and the mean correct prediction (MCP) of the direction of change in the value of each VIX futures price series. The random walk model (Panel A), the economic variables model (Panel B), the AR(2) model (Panel C), the VAR model (Panel D), the ARIMA(1,1,1) model (Panel E), and the PCA regression model (Panel F) have been implemented. Results for the unweighted and weighted combination point forecasts (Panel G and H, respectively) are also reported. The Modified DieboldMariano test (for RMSE and MAE) and the ratio test (for MCP) are employed, to test the null hypothesis that the random walk and the model under consideration perform equally well. Two alternative hypotheses H1 and H2 are considered. Namely H1: the random walk outperforms the model and H2: the model outperforms the random walk. One and two asterisks (crosses) denote rejection of the null hypothesis in favour of the alternative H1 (H2) at significance levels 1% and 5%, respectively. The models have been estimated recursively for the period March 18, 2005 to March 13, 2008.

32

Interval Forecasts # Violations LRunc # Violations LRunc # Violations LRunc # Violations LRunc # Violations LRunc # Violations LRunc # Violations LRunc

Shortest 99% 95%

2nd Shortest 99% 95%

3rd Shortest 99% 95%

Panel A: Economic Variables Model Interval Forecasts 6.06% 11.80% 6.67% 11.28% 6.93% 75.13* 44.94* 83.61* 36.40* 81.86* Panel B: AR(2) Model Interval Forecasts 6.02% 11.11% 6.95% 11.42% 7.44% 76.64* 38.39* 93.19* 39.11* 95.96* Panel C: VAR Model Interval Forecasts 5.50% 10.65% 7.33% 11.69% 7.59% 57.93* 29.99* 97.15* 39.91* 102.25* Panel D: ARIMA(1,1,1) Model Interval Forecasts 17.94% 23.24% 16.12% 22.07% 20.44% 495.00* 262.94* 395.03* 221.43* 528.59* Panel E: PCA Regression Model Interval Forecasts 5.89% 12.18% 7.74% 12.70% 6.87% 57.87* 40.15* 94.05* 44.93* 74.69* Panel F: Unweighted Combination Interval Forecasts 8.25% 14.54% 9.52% 14.68% 9.49% 65.91* 106.23* 134.25* 131.19* 66.98* Panel G: Weighted Combination Interval Forecasts 5.70% 11.00% 6.94% 13.10% 6.87% 54.25* 29.20* 77.57* 49.07* 74.69*

13.11% 52.14* 11.80% 39.44* 11.31% 35.48* 25.96% 290.25* 13.13% 48.57* 15.96%
81.39*

11.52%
32.87*

Table VI: Statistical efficiency of the constructed interval forecasts. Entries report the percentage of the observations that fall outside the constructed intervals, and the values of Christoffersens (1998) likelihood ratio test of unconditional coverage (LRunc) for each VIX futures price series. The null hypothesis is that the percentage of times that the actually realized futures price falls outside the constructed (1-)%-interval forecasts is a%. One and two asterisks denote rejection of the null hypothesis at 1% and 5% significance levels, respectively. The results are reported for daily 99% and 95%-interval forecasts generated by the economic variables model (Panel A), the AR(2) model (Panel B), the VAR model (Panel C), the ARIMA(1,1,1) model (Panel D), and the PCA regression model (Panel E). Results for the unweighted and weighted combination 99% and 95%-interval forecasts (Panel F and G, respectively) are also presented. The models have been estimated recursively for the period March 18, 2005 to March 13, 2008.

33

Shortest 2nd Shortest 3rd Shortest Panel A: Economic Variables Model Point Forecasts SR 0.064 -0.030 0.016 95% CI (-0.01, 0.14) (-0.10, 0.05) (-0.06, 0.09) Ap 0.648 -0.198 0.110 95% CI (-0.11, 1.39) (-0.74, 0.35) (-0.38, 0.60) Panel B: AR(2) Model Point Forecasts SR 0.014 -0.023 -0.007 95% CI (-0.06, 0.09) (-0.10, 0.06) (-0.09, 0.07) Ap 0.146 -0.147 -0.025 95% CI (-0.59, 0.86) (-0.69, 0.42) (-0.51, 0.46) Panel C: VAR Model Point Forecasts SR -0.013 -0.026 0.038 95% CI (-0.09, 0.06) (-0.10, 0.05) (-0.04, 0.11) Ap -0.128 -0.179 0.238 95% CI (-0.90, 0.60) (-0.72, 0.36) (-0.26, 0.72) Panel D: ARIMA(1,1,1) Model Point Forecasts SR 0.009 -0.008 -0.016 95% CI (-0.06, 0.09) (-0.08, 0.07) (-0.09, 0.06) Ap 0.099 -0.039 -0.091 95% CI (-0.63, 0.82) (-0.59, 0.51) (-0.56, 0.39) Panel E: PCA Regression Model Point Forecasts SR 0.007 0.040 0.057 95% CI (-0.07, 0.08) (-0.04, 0.12) (-0.02, 0.14) Ap 0.072 0.296 0.363 95% CI (-0.69, 0.83) (-0.27, 0.85) (-0.12, 0.85) Panel F: Unweighted Combination Point Forecasts SR 0.024 -0.015 0.001 95% CI (-0.05, 0.10) (-0.09, 0.06) (-0.08, 0.08) Ap 0.240 -0.098 0.025 95% CI (-0.51, 0.98) (-0.65, 0.46) (-0.47, 0.50) Panel G: Weighted Combination Point Forecasts 0.011 -0.034 -0.064 SR (-0.06, 0.09) (-0.11, 0.04) (-0.14,0.01) 95% CI
Ap

95% CI

0.111 (-0.64, 0.86)

-0.231 (-0.78, 0.31)

-0.415 (-0.89, 0.08)

Table VII: Trading strategy with VIX futures based on point forecasts from March 18, 2005 to March 13, 2008. The entries show the annualised Sharpe ratio (SR) and Lelands (1999) alpha (Ap) and their respective bootstrapped 95% confidence intervals (95% CI) within parentheses. The strategy is based on point forecasts obtained from the economic variables model (Panel A), the AR(2) model (Panel B), the VAR model (Panel C), the ARIMA(1,1,1) model (Panel D) and the PCA regression model (Panel E). Results for the unweighted and weighted combination point forecasts (Panel F and G respectively) are also reported. The SR for a nave buy and hold strategy in VIX volatility futures is 0.0178 [95% CI = (-0.06, 0.09)] for the shortest maturity series, 0.0421 [95% CI = (-0.03, 0.11)] for the second shortest maturity series and 0.0532 [95% CI = -0.02, 0.13)] for the third shortest maturity series.

34

2nd Shortest 3rd Shortest Interval Forecasts 99% 95% 99% 95% 99% 95% Panel A: Macro Regression Point Forecasts SR 0.061 0.078 -0.011 -0.017 0.002 0.026 95% CI (-0.01, 0.13) (0.00, 0.15) (-0.09, 0.07) (-0.09, 0.06) (-0.07, 0.08) (-0.05, 0.10) 0.607 0.777 -0.060 -0.106 0.029 0.165 Ap 95% CI (-0.13, 1.37) (0.05, 1.52) (-0.62, 0.48) (-0.66, 0.46) (-0.47, 0.52) (-0.33, 0.64) Panel B: AR(2) Model Interval Forecasts SR 0.036 0.044 0.061 0.006 -0.006 0.012 95% CI (-0.04, 0.11) (-0.03, 0.12) (-0.02, 0.14) (-0.07, 0.08) (-0.08, 0.07) (-0.06, 0.09) 0.363 0.441 0.436 0.062 -0.032 0.090 Ap 95% CI (-0.36, 1.11) (-0.29, 1.16) (-0.11, 0.99) (-0.49, 0.61) (-0.52, 0.46) (-0.40, 0.58) Panel C: VAR Model Interval Forecasts SR 0.005 0.002 0.004 -0.004 0.037 0.064 95% CI (-0.07, 0.08) (-0.07, 0.08) (-0.07, 0.08) (-0.08, 0.07) (-0.04, 0.12) (-0.01, 0.14) 0.058 0.023 0.037 -0.029 0.228 0.391 Ap 95% CI (-0.72, 0.81) (-0.73, 0.77) (-0.51, 0.59) (-0.57, 0.51) (-0.25, 0.71) (-0.09, 0.89) Panel D: ARIMA(1,1,1) Model Interval Forecasts SR 0.006 0.009 0.004 -0.015 -0.019 -0.018 95% CI (-0.07, 0.08) (-0.06, 0.08) (-0.07, 0.08) (-0.09, 0.06) (-0.10, 0.06) (-0.10, 0.06) 0.069 0.093 0.044 -0.087 -0.112 -0.103 Ap 95% CI (-0.65, 0.80) (-0.64, 0.83) (-0.49, 0.59) (-0.64, 0.47) (-0.57, 0.37) (-0.56, 0.37) Panel E: PCA Regression Model Interval Forecasts SR 0.045 0.031 0.033 0.017 0.029 0.052 95% CI (-0.03, 0.12) (-0.04, 0.11) (-0.04, 0.11) (-0.06, 0.09) (-0.05, 0.11) (-0.02, 0.13) 0.446 0.314 0.253 0.130 0.195 0.330 Ap 95% CI (-0.32, 1.18) (-0.42, 1.07) (-0.30, 0.80) (-0.42, 0.69) (-0.29, 0.67) (-0.15, 0.83) Panel F: Unweighted Combination Interval Forecasts SR 0.016 0.019 -0.005 -0.014 0.009 0.003 95% CI (-0.06, 0.09) (-0.05, 0.09) (-0.08, 0.07) (-0.09, 0.06) (-0.07, 0.09) (-0.08, 0.08) 0.158 0.187 -0.017 -0.092 0.073 0.038 Ap 95% CI (-0.57, 0.87) (-0.55, 0.93) (-0.57, 0.55) (-0.66, 0.44) (-0.41, 0.56) (-0.45, 0.52) Panel G: Weighted Combination Interval Forecasts SR 0.021 0.006 -0.021 -0.047 -0.038 -0.058 95% CI (-0.05, 0.10) (-0.07, 0.08) (-0.10, 0.05) (-0.12, 0.03) (-0.11, 0.04) (-0.13, 0.02) -0.14 -0.3274 -0.24 -0.37879 Ap 0.212 0.058 95% CI (-0.55, 0.95) (-0.69, 0.82) (-0.70, 0.41) (-0.88, 0.22) (-0.71, 0.26) (-0.85, 0.12) Shortest Table VIII: Trading strategy with VIX futures based on interval forecasts from March 18, 2005 to March 13, 2008. The entries show the annualised Sharpe ratio (SR), Lelands (1999) alpha (Ap) and their respective bootstrapped 95% confidence intervals (95% CI) within parentheses. The strategy is based on 99% and 95%-interval forecasts obtained from the economic variables model (Panel A), the AR(2) model (Panel B), the VAR model (Panel C), the ARIMA(1,1,1) model (Panel D) and the PCA regression model (Panel E). Results for the unweighted and weighted combination point forecasts (Panel F and G respectively) are also reported. The SR for a nave buy and hold strategy in VIX volatility futures is 0.0178 [95% CI = (-0.06, 0.09)] for the shortest maturity series, 0.0421 [95% CI = (-0.03, 0.11)] for the second shortest maturity series and 0.0532 [95% CI = -0.02, 0.13)] for the third shortest maturity series.

35

Footnotes
1

VIX is an implied volatility index that tracks the implied volatility of a synthetic option on the S&P500 with constant time to maturity (thirty days). It reflects the market participants consensus view about expected futures stock market volatility and can be considered as an investor fear gauge (Whaley, 2000).

The CBOE launched the VXD and VXN volatility futures in April 25, 2005 and July 6, 2007, respectively. The VXD and VXN are implied volatility indices that track the implied volatility of a synthetic option on the Dow Jones Industrial Average and the Nasdaq 100, respectively, with constant time to maturity (thirty days). Regarding the liquidity of volatility futures, on January 2, 2008 the open interest for VIX futures was 55,792 contracts or $1.3 billion in terms of market value; this corresponds to a 59% increase from January 3, 2007. The trading volume was 2,481 contracts or $57 million in terms of market value. On the same date, the open interest of VXD and VXN futures was $19 and $4 million, respectively.

This question is distinct from the question whether futures markets are efficient in the sense that the futures price is an optimal forecast of the underlying spot price to be realized on the contract expiry date (see e.g., Coppola, 2008; Kellard et al., 1999, and the references therein, and Nossman and Vilhelmsson, 2008, for a study using VIX futures). In our study, Jensens (1978) definition of futures market efficiency is adopted; a market is efficient with respect to the information set It, in the case where it is impossible to make economic profits by trading on the basis of this information set.

Efficiency in the underlying implied volatility index market does not necessarily imply an efficient volatility futures market; there may be other factors/information flow that affect volatility futures markets, as well.

On March 26, 2007, the VIX futures were rescaled in two ways. First, VIX futures were based directly on the underlying VIX volatility index instead on the Increased-Value index (VBI=10*VIX). Second, the multiplier was increased from $100 to $1,000. As a result, the traded futures prices were reduced by a factor of 10, but the $ value of each contract did not change. We adjusted the VIX futures series accordingly.

Note that the MCP cannot be calculated for the random walk model. However, we proxy the random walk with the nave rule that the predicted change in the futures prices has a 50% chance to be positive and a 50% to be negative. This is to say that the random walk case corresponds to an MCP equal to 50%. In addition, Whites (2000) test is not applied to the MCP metric. This is because the corresponding loss function cannot be defined for the benchmark model.

7 In the case of the MCP, the H1 and H2 hypotheses are stated as H1: MCP < 50% and H2: MCP > 50%.

36

The stationary bootstrap is applicable to weakly dependent stationary time series; stationarity has been found
for fi ,t ,T in our case. It involves re-sampling blocks of random size from the original time series to form a

pseudo time series (or a bootstrapped sample). The block size follows a geometric distribution with mean block length 1/q. The main feature of this procedure is that the re-sampled pseudo time series retain the stationarity property of the original series. Following Sullivan et al. (1999), we choose q = 0.1 that
corresponds to a mean block size of 10. This is a reasonable block size given the low autocorrelation in fi ,t ,T

(results not reported).

We do not employ Jobson and Korkies (1981) testing procedure to investigate the statistical significance of the SR. This is because this test is based on the assumption that the returns of the strategy under consideration are distributed asymptotically normally. Instead, we use bootstrapped confidence intervals that are robust to non-normality. In our case, the trading strategies returns exhibit excess kurtosis and skewness; kurtosis and skewness range from 6 to 13 and from 1 to 2 respectively, across the three futures maturities. The nonnormality of volatility futures returns is consistent with previous findings in the related literature for other futures markets (see e.g., Taylor, 1985, and the references therein).

37