Estimating the evolution of elasticities of natural gas demand: the case of Istanbul, Turkey

Abstract

Much of the existing literature on demand for natural gas assumes constant and single-value elasticities, overlooking the possibility of dynamic responses to the changing conditions. We aim to fill this gap by providing individual time series of short-run elasticity estimates based on maximum entropy resampling in a fixed-width rolling window framework. This approach does not only enable taking the variability of the elasticities into account, but also helps obtain more efficient and robust results in small samples in comparison with conventional inferences based on asymptotic distribution theory. To illustrate the methodology, we employ monthly time-series data between 2004 and 2012 and analyze the dynamics of residential natural gas demand in Istanbul, the largest metropolitan area in Turkey. Our findings reveal that the elasticities of the demand model do not remain constant and they are sensitive to the economic situation as well as weather fluctuations.

Appendix: ARDL estimation results

In order to assess the results of a more conventional parametric method on the whole sample, we also employ an autoregressive distributed lag (ARDL) model of the form

$$\begin{aligned} \text {QNG}_t= & {} \beta _0 + \beta _1 \sum _{j=1}^J \text {QNG}_{t-j} + \beta _2 \sum _{k=0}^K \text {GDP}_{t-k} + \beta _3 \sum _{l=0}^L \text {PNG}_{t-l} + \beta _4 \sum _{m=0}^M \text {PE}_{t-m} \nonumber \\&+ \beta _5 \sum _{n=0}^N \text {TAV}_{t-n} + \beta _6 \sum _{o=0}^O \text {QU}_{t-o} + v_t. \end{aligned}$$

(2)

where the QNG, GDP, PNG, PE, and QU variables are in natural logs. Table 2 shows the results of the estimated ARDL(2,0,1,0,0,0) model, which was determined by the Akaike information criterion. The CUSUM and CUSUMSQ plots of the full-sample estimates of the ARDL model are also shown in Fig. 4. As can be seen in the table, the income, the real price of electricity as well as the number of customers variables do not seem to be significant at the 5 % level in the short run. In the long run, only the average temperature variable is significant; moreover, the sign of the own price elasticity is positive. The poor results provide additional support for the moving window meboot approach that we employ in this study.

Table 2 Estimation results for ARDL(2,0,1,0,0,0) model selected based on Akaike information criterion

Fig. 4

CUSUM and CUSUMSQ plots of the full-sample estimates of the ARDL model with 95 % asymptotic intervals