Re-examining the effect of oil prices on economic activity in Russia∗ Nikola Mirkov† Swiss National Bank

October 2015 Preliminary Draft

Abstract The new exchange rate regime announced by the Russian Central Bank in November 2014 is likely to weaken the cointegrating relationship between the Russian GDP and the Urals oil price. Using quarterly data from 1996 to 2015, I show that their co-integration became less pronounced towards the end of the sample, as the economy begins to adjust to changes in oil prices through higher import prices, greater inflation and tighter monetary policy rather than lower (oil) income. Nevertheless, the oil price level should remain an important determinant of investment activity in Russia given a relatively high reliance of oil exploration on imported technology. Along the lines of Hansen (2000), I estimate that such ’threshold’ effects of the oil price on activity are the strongest between 45 and 55 dollars per barrel. Keywords: oil price, GDP, co-integration, threshold JEL Classifications: E32, P28, E52

∗

I would like to thank Iikka Korhonen for his valuable comments. Views expressed in this paper belong to the author and do not necessarily reflect the views of the Swiss National Bank. † Nikola Nikodijevic Mirkov, Swiss National Bank, Borsenstrasse 15, 8001 Zurich, Switzerland, E-mail: [email protected], Tel: +41 79 512 7892

1

1

Introduction

In a convergence towards fully-fledged inflation targeting, the Russian Central Bank (CBR) decided in November 2014 to let the rouble freely float against the US dollar and the euro. As a result, the ruble was allowed to react to market forces, most notably to the changes in oil prices. By the end of the year, the Urals price of oil dropped by more than 35 percent. Lower export revenues resulted in a short supply of dollars with respect to roubles, causing a 31 percent depreciation of the rouble against the dollar during the same period. One important consequence of the new exchange rate policy was that the export revenues in roubles, particularly relevant to fiscal policy, barely changed. However, import prices reacted strongly to the rouble depreciation, forcing the CBR to raise domestic interest rates sharply. Does the new FX policy fundamentally alter the ways of how the economy adjusts to shocks in the oil price? And if that is the case, could the well-established cointegration between Russian GDP and oil prices, documented in Rautava (2013), Andreas Benedictow and Lofsnaes (2010) and Shafi and Hua (2014) among many others, weaken as a consequence. To answer the question, I estimate the Johansen’s statistic across sub-samples of expanding and rolling windows of data and find that the statistic drops considerably towards the end of the sample. If the Urals oil price in roubles is used in the estimation, the Johansen’s statistic for the full-sample is below any conventional level of significance. Although GDP and oil prices are likely to exhibit less co-movement in the future, the economy is not necessarily less exposed to oil price fluctuations. As already noticed, lower energy prices affect the Russian economy through weaker terms of trade, currency devaluation and higher import prices. The Central bank raises interest rates in order to counteract inflationary pressures and the economic activity drops as a consequence. On top of that, the level effect of oil price could still re-emerge whenever the Urals oil price falls below a certain ’threshold’ level. A particularly low oil price could discourage capital expenditure in the oil exploration and result in a permanently lower potential growth of the economy, because the Russian energy sector is highly dependent on imported technology, see for example Tokarev (2015). Following Hansen (2000), I estimate the ’particularly low’ price to be between 45 and 55 dollars per barrel. 2

The rest of the paper is organised as follows. Section 2 explains the data set. Section 3 describes the co-integration test and threshold regression model, and reports the main results. Section 4 concludes.

2

Data

I use quarterly data on the Russian GDP, the price of Urals oil and the rouble/US dollar exchange rate. Seasonally and working day adjusted GDP numbers are downloaded from CEIC. The price of Urals oil is downloaded from Oxford Economics database and rouble exchange rate from Datastream. The sample begins in 1996 for the reasons of comparability to other studies, most notably to Rautava (2013). Official GDP data are available only from the first quarter of 1995. The last GDP release regards the first quarter of 2015.

3 3.1

Long-run relationship Co-Integration test of Johansen

In order to assess the significance of co-integration between the Russian economic activity and the oil prices, I perform the likelihood ratio test proposed in Johansen (1991). Consider the following 2-variable Vector Error Correction (VEC) model

∆ X t = µ + Π X t− k +

kX −1

Γ i ∆ X t− i + ε t

i =1

where X t is a 2 × 1 vector containing log levels of GDP and oil price, µ is an intercept, Π and Γ i are feedback matrices and ε t is a vector of independent Gaussian variables with mean zero and variance matrix Σ. If the matrix Π has a rank of zero, the system can be expressed entirely in differences and there is no cointegration between the variables. I test this null hypothesis against the alternative that rank(Π) = 1, i.e. that there exists one arbitrary linear combination of GDP and the oil price which is stationary.

3

The test is performed on the full data sample, as well as on the sub-samples of expanding- and rolling-window data. The first sub-sample using expanding window subsamples ends in the first quarter of 2005. For the rolling window, I arbitrarily choose 40 quarters of data points. Each time the model is estimated, the lag k is automatically chosen from k = [1, 10] using the Bayesian Information Criterion. Figure 1 illustrates the results of the co-integration test. The reported statistics of Johansen test the null hypothesis of co-integration rank being equal to zero (no co-integration) against the alternative hypothesis of the existence of one linear combination of GDP and oil price (in logs) being stationary. The upper panel of the figure shows the results with the expanding sub-sample, whereas the lower panel reports the rolling-window results. As it can be noticed, the null hypothesis can be rejected at the level of significance of 0.1 throughout a large part of the sample. However, the Johansen’s statistic points to weakening co-integration between the two variables towards the end of the sample. The finding applies in particular to the link between GDP and the Urals price of oil in roubles.

3.2

Permanent effects

In order to assess the long-term effects of Urals oil price on the Russian economy, I run the following regression

gdp t = α + βt + γoil t + ε t

(1)

where gdp t and oils t are GDP and Urals oil price in logs, t is a trend variable and ε t ∼ iidN (0, σ). Similarly to Johansen co-integration test, I run the regression on the full sample of data, as well as on the sub-samples of expanding- and rolling-window data. The Figure 2 reports the slope coefficient γ together with a 99 percent confidence bound calculated using Newey-West standard errors.1 Judging by the Panel A, the slope coefficient γ was broadly stable across the sample with the coefficient on the rouble oil price constantly below the coefficient on 1

As a rule of thumb, the number of lags used in every iteration is set to N 1/4 where N is the

number of observations, see for example Greene (2003).

4

the dollar price. The full-sample estimate (with 1Q2015 as the end-observation) of γ implies a 1.2 or 1.7 percent permanent drop in output for a 10 percent drop in rouble or dollar oil price, respectively. The effect of the dollar oil price is broadly in line with the previous literature, see for example Rautava (2013). Using a 10-year rolling window to estimate the slope coefficient paints a considerably different picture. Both rouble and dollar price coefficient drop significantly towards the end of the sample implying a 0.8 and 1.0 percent of permanent loss in output as a result of a 10 percent drop in rouble and dollar price of Urals oil, respectively.

3.3

Threshold effects

Even though the level effect of oil prices on the economy seem to weaken due to the new exchange rate regime, those effects could still be important below a certain oil price level. Sufficiently low prices of oil on the global markets could put a considerable strain on revenues of Russian oil companies, which are heavily reliant on imported technology, see for example Tokarev (2015). It could even discourage capital expenditure in oil exploration and result in a permanently lower potential growth of the economy. Taking these considerations into account, this section explores the possibility that there might be a ’threshold’ oil price under which the effect of oil prices on the Russian GDP is even stronger than otherwise. In particular, consider a threshold regression model of the form

gdp t = α + βt + γ1 oil t + ε t

when

oil t ≤ θ

(2)

gdp t = α + βt + γ2 oil t + ε t

when

oil t > θ

(3)

where oil price is the threshold variable and θ is the threshold price. Following Hansen (2000), the model in (2) and (3) can be written in a single equation

gdp t = α + βt + γoil t + δoil t × I (oil t ≤ θ ) + ε t

(4)

where I (oil t ≤ θ ) is an indicator function, taking the value of one when the oil price 5

is lower than the threshold and zero otherwise, δ = γ2 − γ1 can be considered as a ’threshold effect’ and γ = γ2 . Conditional on θ , (4) is linear in all the parameters and can be estimated via OLS. Moreover, I include a dummy variable on the right hand side that takes the value of 1 in all the quarters of the sample before 2000 and 0 thereafter. The dummy should capture the ’transitional effects’ in the Russian economy that might have affected negatively the level of activity in that part of the sample.2 That said, I define a grid of oil prices in a range [20, 60] dollars and estimate a sequence of concentrated sum of squared errors S (θ ) for each point on the grid. Following Hansen (2000), I calculate the likelihood ratio statistics

LR (θ ) = T

S 0 − S (θ ) S (θ )

(5)

where T is the sample size and S 0 is the sum of squared errors from equation (1). In other words, the null hypothesis is that H0 : δ = 0 i.e. “no threshold effect”. Hansen (2000) shows that the asymptotic distribution of the likelihood ratio statistic is non-standard and reports selected critical values. The critical value for the confidence level of 0.99 is reported to be 10.59 and the upper panel of Figure 3 plots the obtained likelihood ratios against this critical value. The points at which the LR statistic crosses the horizontal line constitute a 99 percent confidence set [$27, $53]. Apparently, there seem to be a ’second regime’ in the relationship between the Urals oil price and Russian GDP beneath those levels. However, the confidence range is very bright and I now turn to threshold estimates that account for heteroscedasticity in the data.

3.3.1

Accounting for heteroscedasticity

Hansen (2000) argues that the likelihood ratio statistic in (5) leads to valid inference only if the error term in (4) exhibits homoscedasticity i.e. E (ε2t ) = σ2 . If the assumption proves to be too strong, the likelihood ratio statistic need to be scaled as follows 2

These transitional effects together with relatively low oil prices during the 90ties could produce

a downward bias in the estimate of threshold oil price. I thank to Iikka Korhonen for this comment.

6

LR ∗ (θ ) = T

S 0 − S (θ ) σˆ 2 ηˆ 2

(6)

where ηˆ 2 by a nuance parameter, which can be consistently estimated using a polynomial regression or the Nadaraya-Watson kernel, see Hansen (2000) for details. Once I obtain the normalised likelihood ratios, I estimate p-values for each LR ∗ (θ ) using a bootstrap method proposed in Hansen (1996), who shows that the method yields asymptotically correct p-values. I randomly draw data observations with replacement and construct 10,000 samples of 80 quarters. Then, I estimate LR ∗sim (θ ) along the grid of [20, 60] dollars on each simulated sample. The number of times

LR ∗ (θ ) exceeds LR ∗sim (θ ) for a given θ is considered to be the p-value for that particular likelihood ratio. The lower panel of Figure 3 plots the results. As it can be noticed, the confidence set is much narrower and the threshold level of oil price lies between 45 and 55 dollars per barrel. Below these levels, the effect of changes in the Urals oil price on Russian GDP is statistically different from the corresponding effect above the threshold. Strictly speaking, the p-values below the level of significance of 0.1 correspond to a range of Urals oil price equal to [$53.1 55.1$]. According to the figure though, the likelihood values from 45 to 55 dollars per barrel appear to be fairly close to each other.

3.3.2

What happens when the oil price drops below 45 dollars?

If the Urals oil price goes below 45 dollars a barrel, the estimated equation (4) that includes the ’pre-2000’ period dummy reads

gdp t = 8.261 + 0.005 × t − 0.059 d pre−2000 + 0.146 × oil t − 0.016 × oil t × I (oil t ≤ θ ) (0.072) (0.001) (0.018) (0.019) (0.006) where asymptotic standard errors of Hansen (2000) are reported in parenthesis. So beneath the threshold, a 10 percent drop in the oil price results in 1.62 (= δ - γ) percent permanent loss of output i.e. the effect on output increases by roughly 10 percent. Considering other coefficients, the trend growth of the Russian economy is roughly equal to 2 percent per year (0.005 × 4) and the estimate is broadly in line with Rautava (2013).

7

4

Conclusion

This paper argues that co-integration of Russian GDP and the oil price is likely to weaken significantly in the future. The new exchange rate regime in Russia allows the rouble to fluctuate freely and reflect the relative scarcity/abundance of US dollars given a drop/rise in oil prices and oil revenues. As a consequence, the economy adjusts to an oil price shock through import prices, domestic inflation and monetary policy, rather than nominal incomes. I use quarterly data on GDP and the Urals oil price from 1996 to 2015 to show that the link between the two became weaker towards the end of the sample, especially if the oil price is expressed in terms of roubles. However, the economy is likely to remain exposed to a drop in oil prices under a certain ’threshold’ level given a high dependence of oil companies on imported technology. I estimate the threshold effect to be somewhere around $45 and $55 per barrel.

8

References Andreas Benedictow, Daniel Fjaertoft and Ole Lofsnaes (2010) “Oil dependency of the Russian economy: an econometric analysis,” Discussion Papers of Statistics Norway, No. 617. Greene, William H. (2003) “Econometric Analysis,” Prentice Hall, No. 5th Edition. Hansen, Bruce E. (1996) “Inference When a Nuisance Parameter Is Not Identified Under the Null Hypothesis,” Econometrica, Vol. 64, No. 2, pp. pp. 413–430. (2000) “Sample Splitting and Threshold Estimation,” Econometrica, Vol. 68, No. 3, pp. 575–603. Johansen, Soren (1991) “Estimation and Hypothesis Testing of Cointegration Vectors in Gaussian Vector Autoregressive Models,” Econometrica, Vol. 59, No. 6, pp. pp. 1551–1580. Rautava, Jouko (2013) “Oil Prices, Excess Uncertainty and Trend Growth,” Focus on European Economic Integration, No. 4. Shafi, Khuram and Liu Hua (2014) “OIL PRICES FLUCTUATIONS & ITS IMPACT ON RUSSIAN’S ECONOMY; AN EXCHANGE RATE EXPOSURE,” Asian Journal of Economic Modelling, No. 2(4). Tokarev, A. H. (2015) “Can we substitute imports in the oil industry? (in Russian),” All-Russian Economic Journal (in Russian), No. 4.

9

Figure 1: Co-integration Rank Test. The figure plots the Johansen test statistic of the co-integration between the Russian GDP and the Urals oil price in US dollars (blue line) and roubles (green line) together with the critical values (green area). Panel A: Expanding Window

Panel B: Rolling Window

10

Figure 2: Slope Coefficient. The figure plots the slope coefficient γ from equation (1) with Urals oil price expressed in US dollars (blue solid lines) and roubles (red solid lines). I also report 99 percent confidence bounds around the letter estimates (red dashed lines). Panel A: Expanding Window

Panel B: Rolling Window

11

Figure 3: Hansen’s likelihood ratios and p-values. The figure plots the likelihood ratio statistics from equation (5) (panel A) and p-values of likelihood ratio statistics from equation (6) (panel B). Panel A: Homoscedastic Errors

Panel B: Heteroscedastic Errors

12