Happy new year, I hope (wonkish)

Might Kazakhstan’s economy pick up steam in 2016?

I forecast a growth rate of real gross domestic product in 2016 of roughly 4%, comparable to 2013. This is higher than most other forecasts, which are in the range of 1% to 2%, because I expect the annual price of oil on spot markets to rise next year.

For January through October, the average monthly price of Brent crude on the spot market fell nearly 45% compared to the same period in 2014, from $104.65 to $54.60, according to data from the US Energy Information Administration.  This decline slowed the Kazakhstani economy this year to a pace between 1% and 2% in terms of output. 

I attribute the drop in oil prices mainly to speculation, since oil supply and demand did not change enough in 2014-5 to justify such a large fall in price.  I anticipate that the overshooting of the long-run price – that is, of the price that is based on market fundamentals -- will be corrected in 2016 with a 9% increase to $61 or $62 per barrel.  A simple econometric model relating the annual growth rate of real GDP in Kazakhstan to the annual growth rate of global oil prices suggests that GDP will rise by about 4.1% in 2016.

A time-series approach reaches roughly the same conclusion as the structural one:  Real GDP will rise moderately in 2016. The model forecasts 7% growth with a 95% confidence interval of [4.3%, 9.7%].  I prefer the lower bound (4.3%) as the forecast rather than 7%.  The lower bound is consistent with the prediction from the structural model, and it also performed fairly well in the 2015 forecast, where the mean was 3.1% and the 95% confidence interval was [1.8%, 4.3%]. –Leon Taylor tayloralmaty@gmail.com

Notes

Dependent Variable: D_RGDP
Method: Least Squares
Date: 11/20/15   Time: 15:30
Sample (adjusted): 2000 2014
Included observations: 15 after adjustments
Variable	Coefficient	Std. Error	t-Statistic	Prob.
C	0.039527	0.012513	3.158817	0.0091
SE_MA2_16	-0.630262	0.184061	-3.424204	0.0057
SE_MA8_16	-0.536826	0.163803	-3.277271	0.0074
YEAR2000	0.048905	0.023816	2.053428	0.0646
R-squared	0.714206	Mean dependent var		-0.000533
Adjusted R-squared	0.636262	S.D. dependent var		0.035869
S.E. of regression	0.021633	Akaike info criterion		-4.606055
Sum squared resid	0.005148	Schwarz criterion		-4.417241
Log likelihood	38.54541	Hannan-Quinn criter.		-4.608066
F-statistic	9.163076	Durbin-Watson stat		2.818151
Prob(F-statistic)	0.002504

Table 1:  Forecast model for 2016.  The dependent variable is the first difference in the annual growth rate of real GDP (measured as a fraction).  In this particular model, the intercept C is the forecasted first difference for 2016 as compared to 2015; the standard error for C is the forecast error.  SE_MA2_16 is the second lag of the moving-average term (that is, of residuals) obtained by regressing the annual growth rate of GDP upon its lag; this variable captures short-run shocks.  I adjusted the variable for the moving-average value assumed for the forecast.  Similarly, SE_MA8_16 is the eighth lag of the moving-average term, adjusted for the forecast.  Year2000 is a dummy variable for the year 2000, which was a turning point for the economy since it was recovering from the ruble crisis of the late 1990s.  All three independent variables are statistically significant at the 7% level of significance.   

Dependent Variable: D_RGDP
Method: Least Squares
Date: 11/20/15   Time: 15:17
Sample (adjusted): 2000 2014
Included observations: 15 after adjustments
Variable	Coefficient	Std. Error	t-Statistic	Prob.
C	0.001471	0.006115	0.240652	0.8143
SE_MA2_15	-0.630262	0.184061	-3.424204	0.0057
SE_MA8_15	-0.536826	0.163803	-3.277271	0.0074
YEAR2000	0.048905	0.023816	2.053428	0.0646
R-squared	0.714206	Mean dependent var		-0.000533
Adjusted R-squared	0.636262	S.D. dependent var		0.035869
S.E. of regression	0.021633	Akaike info criterion		-4.606055
Sum squared resid	0.005148	Schwarz criterion		-4.417241
Log likelihood	38.54541	Hannan-Quinn criter.		-4.608066
F-statistic	9.163076	Durbin-Watson stat		2.818151
Prob(F-statistic)	0.002504

Table 2:  Model for 2015 forecast.  Similar to Table 1.

Date: 11/18/15   Time: 14:13
Sample: 1991 2014
Included observations: 24
Autocorrelation	Partial Correlation		AC	PAC	Q-Stat	Prob
.  |***** |	.  |***** |	1	0.756	0.756	15.520	0.000
.  |****  |	. *|  .   |	2	0.543	-0.068	23.882	0.000
.  |***   |	.  |* .   |	3	0.427	0.095	29.311	0.000
.  |**.   |	. *|  .   |	4	0.293	-0.113	31.997	0.000
.  |* .   |	.**|  .   |	5	0.095	-0.226	32.293	0.000
. *|  .   |	. *|  .   |	6	-0.077	-0.128	32.496	0.000
. *|  .   |	. *|  .   |	7	-0.200	-0.106	33.966	0.000
.**|  .   |	. *|  .   |	8	-0.322	-0.145	38.012	0.000
***|  .   |	.  |* .   |	9	-0.345	0.103	42.971	0.000
.**|  .   |	.  |  .   |	10	-0.309	0.055	47.230	0.000
.**|  .   |	.  |  .   |	11	-0.271	0.020	50.755	0.000
.**|  .   |	. *|  .   |	12	-0.287	-0.161	55.052	0.000

Figure 1:  Autocorrelation functions for the annual growth rate of real GDP.  The first column gives simple correlations of the variable with itself over time; for example, the correlation of the variable at time t with its one-year lag is .756. The second column gives the autocorrelation of the current variable with its k-period lag, holding intervening autocorrelations constant.  For example, the correlation of the variable at time t with its two-year lag, holding constant its correlation with the one-year lag, is -068. The vertical lines mark the 95% confidence interval for the null hypothesis of no autocorrelation.  The patterns suggest that the growth rate is nonstationary – i.e., it changes in fundamental ways over time. They also suggest that the first difference of the growth rate might be stationary and so may be suitable for regression.

Date: 11/18/15   Time: 14:17
Sample: 1991 2014
Included observations: 23
Autocorrelation	Partial Correlation		AC	PAC	Q-Stat	Prob
.  |* .   |	.  |* .   |	1	0.122	0.122	0.3864	0.534
****|  .   |	****|  .   |	2	-0.485	-0.507	6.8271	0.033
. *|  .   |	.  |* .   |	3	-0.072	0.107	6.9782	0.073
.  |***   |	.  |* .   |	4	0.390	0.192	11.585	0.021
.  |**.   |	.  |* .   |	5	0.254	0.208	13.643	0.018
. *|  .   |	.  |* .   |	6	-0.111	0.108	14.058	0.029
. *|  .   |	.  |  .   |	7	-0.172	0.001	15.125	0.034
. *|  .   |	***|  .   |	8	-0.186	-0.385	16.454	0.036
.  |  .   |	. *|  .   |	9	0.063	-0.084	16.618	0.055
.  |* .   |	. *|  .   |	10	0.134	-0.177	17.414	0.066
.  |  .   |	.  |* .   |	11	0.010	0.149	17.419	0.096
.**|  .   |	. *|  .   |	12	-0.286	-0.140	21.690	0.041

Figure 2:  Autocorrelation functions for the first difference of the annual growth rate of real GDP.  They suggest that the second and eighth lags of the moving-average terms might influence the first difference of the growth rate.  This was the basis for my forecast model in Table 1.

The Q-statistic for lag k relates to the probability that there is no autocorrelation up to that lag.  For example, the probability that there is no autocorrelation up to the second lag is 3.3% (in the far-right column).

All data in this post are from the statistical committee of the national economic ministry.