Helen Liefting 10417478 25-06-2018 Mr R.C. Sperna Weiland MSc Bachelor thesis: Finance and Organization
Old versus new: The relationship between oil prices and solar energy stock prices.
Abstract
In this study the relationship between oil prices and solar energy stock prices in the time frame 2013-2017 is analysed by means of a vector autoregressive model to capture the expected dynamic relationship between these variables. Due to developments in the solar industry, I hypothesise that a positive shock to oil prices causes a substitution effect towards solar energy and encourages investment in solar energy firms. The model includes solar stock prices, oil prices, S&P 500 index and the Nasdaq 100 technology index. Outcomes are analysed by Granger causality and impulse response functions. I found significant results for a positive impact on solar stock prices related to oil prices, whereas a significant but negative impact is found for the effects of oil prices on the S&P 500 and the technology index at first, which translates to a positive effect within a month. The results indicate that oil price changes provoke a substitution effect from oil towards solar energy.
2 Statement of Originality
This document is written by Helen Liefting, who declares to take full responsibility for the contents of this document.
I declare that the text and the work presented in this document are original and that no sources other than those mentioned in the text and its references have been used in creating it.
The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.
3 1. Introduction
The threats posed by climate change and related environmental issues are a widespread concern. Knowledge about the effects of fossil fuels and emissions to the environment have led to a surge in discussions about possible solutions like renewable energy, and these are furthered by energy security concerns. Partly, due to collective action problems it has been a relatively slow process towards clean energy (Smith and Mayer, 2018: 140). However, the Paris Climate Agreement in 2015 has accomplished to appeal to an unprecedented number of countries with concern about global warming. The agreement has focussed on reducing emissions, funding adaptation and using finance flows to keep global warming under two degrees Celsius (UN, 2015: 3). Besides the fact that fossil fuels have negative effects on the climate and environment, they are non-renewable. Alternative renewable energy sources have been developed and are still advancing. Solar energy is one of the renewable energy sources that has had an accelerated growth due to technological improvements and cost reductions (Spikes in investment 2014, 2015: 38). In 2014 investments in solar power accounted for nearly half of all investment in green energy. While wind energy takes a second place with an increase of 11% in investments (Idem: 39). This trend continues and in 2017 energy production from solar reached record highs according to the US energy information administration (EIA, 2018).
Non-renewable energy sources such as crude oil and natural gas are primary sources that make up more than 65% percent of energy consumption in the US. Especially crude oil is one of the main commodities that has a high global trading volume. Natural gas is not used as widely as oil, but still accounted for 29% of United States (US) energy consumption in 2017 (EIA, 2018). Solar energy as a complement or substitute provides an alternative for non-renewable energy sources that is available to industries as well as for private consumption, which is not the case for other renewables. Previous research has found that oil prices influence economic activity and stock markets in general (Jones & Kaul 1996; Driesprong et al. 2008; Kilian 2009; Kilian & Park 2009; Wang et al. 2013). Moreover, there has been research which is directed at the influence of oil prices on green energy indices (Kumar et al. 2012, Henriques & Sadorksky 2008). The results have been ambiguous, and circumstances have changed over the last decade with advancements in solar technology and a higher moral pressure towards using clean energy. The solar energy industry is maturing and available for use in industrial, commercial and private spheres (Malinowsky et al., 2017, p. 2144). Therefore, the focus in this research is on the question whether a positive relationship can be identified between prices and returns of crude oil and solar energy stocks.
Within the field of finance, the behaviour of stock markets has played a prominent role. Questions about which factors influence stock price behaviour remain relevant as changing circumstances have an impact on these factors. New industries that emerge through technological advancements offer new possibilities for creating optimal portfolio’s. The renewable energy industry and the solar energy industry specifically are interesting industries to study, because of recent increases in popularity and technological advancements. Information on the relationship between non-renewables and the specific industries in the renewable energy sector are very useful for governments advocating societal and environmental gains in proceeding towards more green energy in accordance with the Paris agreement. In addition, considering the dynamics of the market, they might be a relevant addition to diversified portfolios and provide opportunities for a hedging strategy.
A vector autoregressive model is used to capture the dynamic relationship between oil and solar stock prices over time. The results from granger causality tests and impulse response functions show a significant and positive effect of oil prices on solar stock prices. This knowledge is useful for investors
4 and governments to stimulate constructive investment behaviour and reach energy transformation goals. First, previous literature on this topic is examined and the hypothesis will be formed within the theoretical framework. The second section focusses on the methodology, after which the results and an analysis of the outcomes are given as well as remarks on its implications for investors and society.
2. Theoretical framework and literature review
The relationship between economic activity, economic growth and oil has been widely covered in previous literature and it was found that differences in oil prices and quantities have a significant impact on the state of the world economy (Jones & Kaul, 1996, p. 463: Sadorsky, 1999, p. 449). This has led to research on the relationship between oil, as a widely used and traded commodity, and the behaviour of stock markets and stock returns. Jones and Kaul (1996) find that oil shocks impact future cash flow expectations which influence stock price changes that can be completely accounted for in the US based on the theory of discounted cash flows. Sadorsky (1999) and Park and Ratti (2008) among others find a negative relationship between oil prices and stock market returns. Oil is used as an input in production processes. The price increase leads to higher costs and lower profitability for those companies. This combined with the uncertainty about the volatile oil prices causes the negative effect on the general stock market. Furthermore, Kilian and Park (2009) find that 22% of the variation in stock returns in the US can be explained by changes in demand or supply of crude oil. The relationship has been tested and confirmed in different time periods and throughout many countries.
Besides the general relationship between oil prices and stock markets the link between oil and stock returns in specific industries has been explored. The green energy sector has been subject to research since the oil shocks in the 1970’s (Kumar et al., 2012, p. 215). Renewable energy companies are considered likely to be related to oil price returns and Sadorsky studies the link and the effects in several articles. He finds that oil price returns have a positive effect on the beta in the capital asset pricing model (CAPM) for pricing renewable energy companies (Sadorsky, 2012a, p. 40, 42-43). Moreover, using a multivariate GARCH model he constructs optimal portfolio weights hedging technology and renewable energy stocks with oil futures (Sadorsky, 2012b, p. 253). An earlier study of Sadorsky and Henriques (2008, p. 1006) finds through a vector autoregression that returns of renewable energy companies can be explained by oil prices, technology stock returns and interest rates. However, he argues that the technology sector is more important for the renewable energy sector than oil prices, which show little significance, but sees this as a positive indication for the long-term future in the renewable energy market (Idem, p. 1009). The positive influence of developments in high technology for the use of renewable energy is a plausible and compelling argument for assuming a link between technology and the solar industry. The success of the solar industry has been through technological advancements and energy storage problems require high technology solutions. This relationship will be investigated further to see whether the industries behave similarly. When most investors indeed see the solar industry and the technology industry as interchangeable, this has consequences for the construction of optimal portfolio’s and policies. Kumar et al. (2012, p. 224) using a vector autoregressive model as well also finds this positive effect of the technology industry and has more importantly found a statistically significant and positive relationship between oil prices and clean energy stock returns. He ascribes this to a substitution effect of alternative energy sources for oil in a time of rising oil prices (Idem, p. 225).
The substitution effect relates to the supply and demand theory in economics and to the functioning of (cross-) price elasticities that follow from this theory. It shows the responsiveness of
5 demand (in Y) towards changing prices (of X). The theory states that when prices for X rise, demand for X falls. If Y is a substitute for X consumers will substitute Y for X, so that a rising price for X leads to an increased demand in Y. The goods do not need to be perfect substitutes in order for a substitution effect to arise.
Haug (2011) examines the possibilities and likelihood of substitution of oil and the progress within renewable energy sectors. She argues that the transition towards clean energy has become evident (Idem, p. 92, 113). In addition, she describes what necessary conditions we can derive for the renewable energy industry to become a substitute for oil (Idem, p. 96). Most importantly, technologies need to innovate and commercialize. Second, oil prices cannot drop too low or governments tend to abandon policies towards transition and investors incentives are reduced. In the last five years investments in renewable energy production have surged with a large proportion as a consequence of investment in the photovoltaics and solar industry (Spikes in investment 2014, 2015: 38: McCrone et al., 2017, p. 11). The solar industry has had the biggest share of investments compared to other renewables every year from 2014 to 2017 due to major efficiency increases and cost reductions. Along with expectations of more innovation within the industry in the years to come (Ibid). These prognoses increase investors trust in the feasibility of clean energy and in particular the solar industry. The sector has matured and commercialized as shown by the expansion of competitive pricing due to the use of auctions (EIA, 2017).
The focus in the literature has been on oil prices and clean energy indices in general. For example, Kumar (2012), Sadorsky (2012) and Sadorksy and Henriques (2008) have all used the Wilderhill Clean Energy Index or ETF. This index consists of companies operating in the hydrogen, biomass, geothermal, wind and solar industries. It is interesting to see if the same results are found for separate industries to see if the risks and opportunities associated with renewable energy are the same for an individual industry. In this research the focus will be on the solar industry because of its enhanced prospects in recent years as a substitute for oil and as a feasible option for transition towards green energy and reaching international environmental targets. Solar energy is a renewable energy source that individuals can easily integrate in their homes as an additional part of energy supply. This sets the solar industry apart from for example wind energy, geothermal and hydropower solutions. As Kumar (2012, p. 215) indicates one of the factors stimulating the investment in renewable energy is oil price behaviour. Therefore, the relationship between oil prices and solar industry stock returns will be examined in this research to see if this is the same for the solar industry. Based on the literature and theory the hypothesis is that solar energy stock prices will have a positive relationship with oil prices, because of the described substitution effect that might have even increased in recent years due to the thriving sector, energy security concerns and environmental issues. Furthermore, rising oil prices generally indicate a higher aggregate demand, which has a more pronounced effect than supply shocks (Kilian, 2009, p. 1054). The companies on the stock market face higher production costs, which partly causes the depression of stock markets (Jones & Kaul, 1996, p. 483). For solar companies the higher possibility of a substitution effect creates positive sentiments and if higher prices indicate higher energy demand and a booming economy this indicates additional opportunities for the solar sector.
This research will contribute to the existing work by focussing on one particular sector to see whether its behaviour can be compared to the whole renewable market. Recent data is necessary considering the rapidly changing circumstances within the sector. New companies in the solar industry have been listed on the US stock exchange since previous research has been conducted, whereas other companies have become private or acquired by companies that are not in the solar industry. Oil prices have been volatile throughout history, however they have been trending upward or downward more in
6 specific time periods. In addition to the innovations and changing moral sentiments, this might have led to changes in market dynamics.
An important part of research in finance has been to increase explanatory power of traditional capital asset pricing models and creating optimal diversified portfolio’s (Graham & Harvey, 2001). The CAPM model has been and remains influential for the valuation of stocks and explaining stock returns (Idem, p. 201). In addition to the CAPM which uses a market beta to look for individual stock sensitivity to market returns, other research focuses on creating models that include different factors to increase the explanatory power (Fama & French, 1993: Carhart 1997: Fama & French 2015). However, to establish dynamic relationships in financial time series, besides the simpler multifactor regression models, vector error correction and vector autoregression models have been used to provide deeper analysis of the relationship between variables and their lagged values. The next section will introduce these methods as well as the data that will be used in this research.
3. Methodology
First, to create a preliminary idea of the sensitivity of solar stock returns to the behaviour of oil prices, ordinary least squares (OLS) regressions will be run using a multi-factor capital asset pricing model. This method has been used by El-Sharif et al. (2005, p. 821) to determine the relationship between oil and equity values in the United Kingdom and similarly by Soyemi, Akingula and Ogebe (2017, p. 4) to investigate this same relationship in Nigeria. However more importantly, to be able to provide a more in-depth analysis of the possible dynamics between the movements of oil prices and solar stock prices, a vector autoregression will be run to capture the relationship over time and to account for some problems in the linear regression models. One of the problems that mostly arises when estimating a simpler model is the existence of autocorrelation, because in time series (and especially financial time series) it is likely that the current value of a variable depends on its previous values (Stock & Watson, 2015, p. 574). In a vector autoregression the variables are regressed on the lagged values of itself and on the lagged values of the other variables (Stock & Watson, 2015, p. 685). The variables are all assumed to be endogenous and the coefficients are estimated by using OLS. There are four variables included in the model; solar prices, oil prices, the broad market index and the technology index. Following Henriques and Sadorksy (2008, p. 1004) before running the vector autoregression the data will be transformed to logarithms to reduce heteroskedasticity.
In the first two models, weekly data is used for estimating the multifactor capital asset pricing model, because these have shown to mitigate problems of within week seasonality and non-trading problems in multivariate models (MacKinley, 1987, p. 1987). The dependent variable is the return on solar company stocks minus the risk-free rate. 𝛽1 measures the sensitivity of the solar portfolio to a
broad market index, whereas the added variable is the main variable of interest and consists of oil price returns to estimate 𝛽2. This coefficient will be used to test the extend of the relationship between solar
company returns and oil price returns.
𝑅𝑠𝑜𝑙𝑎𝑟− 𝑅𝑓 = 𝛽0+ 𝛽1(𝑅𝑚𝑘𝑡− 𝑅𝑓) + 𝛽2(𝑅𝑜𝑖𝑙 − 𝑅𝑓) + 𝜀 (1)
Additionally, the risk return effect of a technology index will be considered as multiple articles indicate that renewable energy companies behave very similar to technology companies, because of their dependence on innovation in technology for generation and storage of energy (Henriques &
7 Sadorsky, 2008, p. 1004: Kumar et al., 2012, p. 216: Sadorksy, 2012, p. 40). The excess returns on technology stocks will be calculated from the Nasdaq 100 Technology Index (NDXT).
𝑅𝑡𝑒𝑐ℎ− 𝑅𝑓 = 𝛽0+ 𝛽1(𝑅𝑚𝑘𝑡− 𝑅𝑓) + 𝛽2(𝑅𝑜𝑖𝑙− 𝑅𝑓) + 𝜀 (2)
The results of this exploratory method will provide an indication of how solar stock returns relate to returns on oil prices and how it compares to the technology sector. Moreover, for the sake of robustness and comparison to datasets in the existing literature, these regressions will be run after substituting returns for the logarithm of prices. The logarithm is used for interpretation purposes, since a log-log model can be interpreted as following; when X increases with 1%, Y increases with 𝛽%, keeping the other variables constant. However, dynamic relations between solar stock prices, oil prices, the S&P 500 and technology sector performance are not completely revealed by this multifactor model. The assumption that the S&P 500 is constant when analysing an oil price shock to solar prices is unrealistic. Therefore, the third model represented below in the form of a p-lag factor autoregressive model will be considered (Escanciano, Lobato & Zhu, 2013, p. 247).
𝒀𝒕= 𝐶 + 𝜷1𝒀𝑡−1+ 𝜷2𝒀𝑡−2+ ⋯ + 𝜷𝑝𝒀𝑡−𝑝+ 𝜺𝑡, 𝑡 = 1, … , 𝑇 (3)
Here 𝒀𝒕 is a (4 x 1) vector of time series variables in the form of 𝒀 = (𝑌1𝑡 , 𝑌2𝑡 , 𝑌3𝑡, 𝑌4𝑡 ) and 𝜷𝒊 are
coefficient matrixes (n x n).
The expectation is that weekly data contains too much noise for estimating long term trends using a vector autoregression model. Therefore, the model will be estimated using monthly data and weekly data will be used for the purpose of robustness checks. The data is collected for a time frame of five years stretching from 2013-2017. Data availability considerations are an important factor for the choice of time frame. Requiring companies with available data exceeding the past five years drastically decreases the sample size. Extending the time frame would not significantly increase the sample mainly due to mergers and acquisitions. These are often made by companies in the traditional energy sector, which could not be included in the portfolio of companies in the solar energy industry.
Solar companies that are listed on several clean energy indices and companies that are listed on US stock exchanges like the NASDAQ and NYSE have been used to establish the portfolio of solar companies. The indices include the Wilderhill Clean Energy Index, Mac Solar Index, Ardour Solar Index, the S&P Global Clean Energy Index and the Nasdaq Clean Edge Green Energy Index.
Since we are looking at US stock exchanges the S&P 500 is chosen as a suitable benchmark for the market proxy (Pilbeam, 2005, p. 210). For the risk-free rate the three-month US treasury bill will be used in accordance with Kumar et al. (2012, p. 218). As a benchmark for oil prices, the returns will be calculated for WTI crude oil (West Texas Intermediate). WTI and Brent are both used as benchmarks for oil prices, but since WTI is produced and consumed in the US it is a better benchmark for the US stock market than the more global brand Brent.
The descriptive statistics of the variables (Appendix, 1) show that oil prices were very volatile and have reached a low of $28.14 and a high of $108.77. The S&P index and the technology index had approximately the same minimum, that is 2036 and 2410 respectively. Though, the technology sector has increased at a faster pace over the five years that are accounted for. The technology sector reached
8 a maximum of 4094 compared to a maximum of 2683 for the S&P index. The portfolio of companies in the solar industry have seen a minimum price of $20.12 and a maximum of $39.57.
A correlation matrix of the returns is presented in table 1.1 and the same matrix for prices is presented in table 1.2. Correlations using prices in financial time series are often overvalued. This is because they are non-stationary, and the earlier prices are weighted more heavily than later prices. Using prices in models one and two will presumably lead to the additional problem of multicollinearity and transformation of the variable is necessary to prevent distorted outcomes. The correlation of prices is not the appropriate instrument to use for interpretation. They are only included for comparative purposes. The correlation between solar and technology returns is so far the most striking one and corroborates the expectation in previous literature that investors consider the industries as similar.
Table 1.1. provides a correlation matrix where solar indicates the return on the solar portfolio returns, oil are oil price returns, tech stands for returns on the technology industry and S&P stands for the returns on the S&P 500. Table 1.2 provides the correlation for solar stock prices, oil prices, the technology and the S&P 500 index levels.
4. Results two-factor model
An impression of the relation between solar and oil prices can be given by a graphical analysis. Below is the scatterplot with a fitted line of solar price returns minus the risk-free rate compared to oil price returns. Here we can carefully conclude that there seems to be a slightly positive relationship between oil price returns and solar stock returns.
Table 1.1. Correlation matrix (Returns)
Solar Oil Tech S&P
Solar 1
Oil 0.3451 1
Tech 0.6966 0.3187 1
S&P 0.2204 0.1726 0.2803 1
Table 1.2. Correlation matrix (Prices)
Solar Oil Tech S&P
Solar 1
Oil 0.1114 1
Tech 0.4832 -0.6177 1
9 Figure 1. Scatterplot with fitted line: solar returns to oil returns in US $
The most important outcomes of the regression of models 1 and 2 are displayed in the table below. However, they are fully taken up in the Appendix (figures 2 and 3). The models are both estimated using ordinary least squares (OLS) and robust standard errors.
The results show that the technology index and the solar portfolio do have similar characteristics especially in the risk-return relationship compared to the broad market index, which corroborates the findings in previous research concerning renewable energy in general. On the other hand, the results are surprising in that the betas for the market index are relatively low. The estimated coefficients indicate that the solar and technology industry returns are both less volatile than the broad market index of the S&P 500. This is striking because previous research indicates that the technology sector and solar sector are more risky and volatile than a broad market portfolio. However, previous research like Kumar et al. (2012) and Henriques and Sadorsky (2008), who determine risk and return relationships have varying results as well. Kumar et al. (2012, p. 218) finds market betas for the renewable and technology sector above 2 where the market betas for several renewable indices found by Henriques and Sadorsky (2008, p. 1004) revolve around 1.4. This difference can be explained by the way in which variables are presented. Whereas models 1 and 2 are estimated using the returns of stock prices for
Table 2. Regression results model 1 and 2
Coefficient S&P P-value Coefficient Oil P-value
Solar 0.32*** 0.005 0.25*** 0.000
Technology 0.35*** 0.000 0.17*** 0.000
0.15 0.15
*** Significant at the 1% level
-0 .1 0 0 0 -0 .0 5 0 0 0 .0 0 0 0 0 .0 5 0 0 0 .1 0 0 0 -0.1000 -0.0500 0.0000 0.0500 0.1000 0.1500 Returns Oil
Returns Solar Fitted values Weekly solar returns to oil returns ( in US $)
Table 2.1. Regression results model 1 and 2 (Returns)
Coefficient - S&P P-value Coefficient - Oil P-value
Solar 0.32*** 0.005 0.25*** 0.000
Technology 0.35*** 0.000 0.17*** 0.000
0.15 0.15
10 correlations and regressions, these previous studies on renewable energy have used absolute prices. Absolute prices can be used for prediction, but they are not stationary, which distorts the results and adds the problem of multicollinearity. When looking at the influence of the variance of a variable on another than the returns are stationary and give a better indication, as is the case when using capital asset pricing models and risk return relationships (Perold, 2004, p. 6-7).
When using prices and index levels for the sake of comparison an oil beta of 0.30 and a market beta of 1.18 are found for the solar sector (Appendix, 4 and 5). The difference compared to previous literature is small but might be explained by the use of different time periods and changed dynamics within the industry. As Henriques and Sadorsky (2008) use data before the financial crisis (up to 2007) and Kumar et al. (2012) use data from 2005-2008. This is a relatively short but turbulent time including part of the financial crisis and this might cause different relationships between assets. Furthermore, since the rapid increase in the production and consumption of solar technology the risk of the industry might have decreased relative to renewable energy indices in general. Still, the oil beta’s like the market betas are all significant at the 1% level which indicates that they are relevant for the risk return dynamics of solar and technology stocks and as hypothesised they show a positive relationship.
5. Vector Autoregression
The properties of the time series data need to be known before proceeding towards the vector autoregression to capture the more detailed and dynamic relationship under scrutiny. For a proper analysis stationarity, optimal lag length selection, cointegration and normality need to be accounted for. The conventional theory used to be that if data is cointegrated Vector Error Correction Models should be used (Todo & Yamamoto, 1995, p. 226). However, Todo and Yamamoto (Idem, p. 245) found it is best to use a lag augmented vector autoregression even with cointegrated data. That is if the lag length chosen consists of the optimal lag length as revealed by the selection criteria and is added to 𝑑𝑚𝑎𝑥, the
maximum order of integration. This is the approach used in most time series research considering renewable energy stock prices and oil relationship and will therefore be the method conducted in this research. The results are shown for monthly logged prices. Weekly data and vector autoregression using returns are covered in the section on robustness. The necessary pre-estimation tests and procedures will be explained and performed in the next paragraphs.
When using a vector autoregression it needs to be confirmed that the data is stationary. A first indication can be given by a graphical analysis. The graphs underneath show the oil price, solar stock and technology stock returns in the given time frame. For stationarity they should revolve around their mean as the mean and variance are constant over time in stationary series. This does not seem to be the case and the graphs give rise to specific concerns towards non-stationarity.
Table 2.2. Regression results model 1 and 2 (Prices)
Coefficient - S&P P-value Coefficient - Oil P-value
Solar 1.18*** 0.005 0.30*** 0.000
Technology 1.92*** 0.000 0.03*** 0.005
0.72 0.97
11 Figure 2. Time series
A rule of thumb to determine whether the data is stationary is that when the 𝑅2 of the regression is lower than the Durbin-Watson test statistics for autocorrelation this indicates stationary data. In this case the Durbin-Watson test statistic is 0.63 compared to an 𝑅2 of 0.76 signalling non-stationary data series (Appendix, 6). Additionally, to be sure a statistically valid model is created a unit root test will be carried out. The unit root test will be conducted using the Augmented Dicker Fuller test. The null hypothesis for this test is non-stationarity, so the null hypothesis should be rejected to be able to use the data for a vector autoregression. The test results are registered in the table below and are all insignificant for the normal test statistic, which indicates that the variables are indeed non-stationary. When using the first difference of all variables, the results of the Augmented Dicker Fuller test are significant at the 1% level. This means they are integrated of order one and the first-differenced data is stationary (Stock & Watson, 2015, p. 696). Using the first difference causes no problems for interpretation of the betas, since the betas will indicate the effect of growth or decline from the independent variables to the change of the dependent variable and this provides the knowledge needed
Table 3. Augmented Dickey Fuller tests for unit root show the test-statistic for the original logged data and for the first difference of all the logged data.
Table 3. Augmented Dickey Fuller Tests for Unit Root
Test Statistic P-value First Difference P-value
Solar -2.531 0.1081 -8.306*** 0.000
Oil -1.486 0.5405 -4.581*** 0.001
S&P -1.000 0.7242 -6.137*** 0.000
Tech 0.063 0.9634 -4.908*** 0.000
12 for determining the relationship. It is necessary to use the first difference in variables to be able to use OLS for estimation. Graphs showing the first differenced time series are shown in figure 3.
Figure 3. First differenced time series
Furthermore, the optimal lag lengths used in the vector autoregression need to be determined. This can be done by using the pre-estimation selection order criteria generated by Stata through the command ‘varsoc’. In this case different selection criteria lead us to differing optimal lag lengths. The criteria HQIC and SBIC indicates an optimal lag length of one. However, the lag length used will be 2 considering the LR, FPE and AIC all find an optimal lag length of two.
Table 4. Selection criteria for lag length selection
Lags LR df p FPE AIC HQIC SBIC
0 16 6.3E-09 -7.53 -7.47 -7.38 1 387 16 0.000 8.8E-12 -14.11 -13.82* -13.37* 2 35.999* 16 0.003 8.2e-12* -14.18* -13.67 -12.86 3 21.760 16 0.151 1.0E-11 -13.99 -13.25 -12.08 4 16.004 16 0.453 1.4E-11 -13.70 -12.73 -11.19 5 25.764 16 0.057 1.8E-11 -13.58 -12.39 -10.49 6 25.758 16 0.058 2.3E-11 -13.47 -12.05 -9.78
13 Moreover, as mentioned before cointegration of data series should be ruled out in order to use the basic vector autoregression. When cointegration exists, this might lead to biased estimators. In that case, the lag augmented vector autoregression model or the vector error correction model are more suitable to the data. The Johansen cointegration test is conducted to determine whether there are cointegrating equations in the model. The null hypothesis is that there is no cointegration.
The maximum rank zero shows a trace statistic below the critical value which leads to acceptance of the null hypothesis. Because no cointegration is found the basic vector autoregression model is the most appropriate compared to a lag augmented vector autoregression or a vector error correction model.
The vector autoregression can now be run and will be interpreted by means of Granger causality tests and impulse response functions. The Granger causality test will show whether the current and lagged values of one time series jointly affect the values of a different time series (Stock & Watson, 2015, p. 820). By using the logged values interpretations are in percentages. A 1% increase in X yields a β% increase in Y. This is also how the Y axis results from the impulse response functions can be interpreted. However, the shape of the impulse response function is most important for showing the relationship.
6. Results and discussion
The model fit values in table five suggest that the model fits well, since the R squared is high and the Chi squared statistics are significant at the 1% level.
Whether the chosen lag length is indeed appropriate can be confirmed post-estimation by testing for autocorrelation with the Lagrange-Multiplier. The lag length two is appropriate considering no residual autocorrelation is found. The stability condition is also satisfied since the modulus found are
The equations refer to the logged data of the solar portfolio, oil prices, S&P 500 index and the Nasdaq 100 technology index
Table 5. Johansen tests for cointegration
Maximum rank Trace statistic 5% critical value
0 46.5985* 47.21
1 17.5154 29.68
2 4.2843 15.41
3 0.8939 3.76
* Number of ranks with cointegration
Table 6. Goodness of Fit - vector autoregressive model Equation R squared Chi2 p>Chi2
Solar 0.881 429*** 0.000
Oil 0.9555 1244*** 0.000
S&P 0.9628 1502*** 0.000
Tech 0.9778 2554*** 0.000
14 all below one and all roots are inside the unit circle, which means there will be no distorting effects from lagged variables (Appendix, 8).
The data needs to follow a normal distribution for the results to make sense. Besides reliance on the central limit theorem, the Jarque-Bera test is run. The null hypothesis is that the distribution resembles a normal distribution. In this case none of the outcomes for test-statistics are significant. So, normality is assumed.
The coefficients of the vector autoregression are taken up in the appendix (table 11). Evaluation of the coefficients and relationships are performed by the granger causality test, which uses F-tests to show which variables, separately and jointly, are significant determinants for the other variables. The null hypothesis of the granger causality test is that all the lagged values of the variable listed under excluded do not explain the variable under equation (table 9). Table 9 shows that the oil prices significantly influence the solar stock prices in the portfolio. On the contrary, the granger causality test suggests the S&P 500 and the technology sector do not help to predict solar stock prices. Furthermore, the test reveals a significant impact of oil on the S&P 500 and on the technology index. The variables taken up in the vector autoregression do not show any significant causal impact on oil prices. The relationship between oil and stock markets was expected based on the theoretical framework. However, the results show no causal relationship between the solar and technology industry which goes against previous research. After the impulse response functions are defined a joint analysis will be given on their outcomes and that of the Granger causality tests.
Table 7. Lagrange-Multiplier test
Lags Chi2 df P-value
1 17.9771 16 0.32524
2 13.3273 16 0.64869
H0: no autocorrelation at lag order
Table 8. Jarque-Bera test for normality
Equation Chi2 P-value
Solar 0.338 0.84431 Oil 0.602 0.74013 S&P 3.502 0.17361 Tech 0.036 0.98204 All 4.479 0.81158 * significant outcome
15 The granger causality tests have shown which variables and which combinations have significant effects. However, when the effects start, how long they last, or which sign the impact of a variable has cannot be determined from this test. This is where impulse response functions are used. The impulse response functions graphically present the responsiveness of the variables used in the vector autoregression to one unit shock or one standard deviation shock in its own and the other variables. The graphs show whether there is positive or negative effect to a positive shock and how long the effects are visible.
Three kind of impulse response functions are shown for the effects of variables on the solar portfolio, since this is the main variable of interest. These are orthogonalized impulse response functions, generalized impulse response functions and the Cholesky forecast-error variance decomposition. The orthogonalized impulse response shows a graph for a one standard deviation shock, whereas the generalized impulse response function shows the response to a one unit shock in the impulse variable. They are both provided for robustness in line with Kumar et al. (2012, p. 221). The variance decomposition is of additional value when analysing the importance of the structural shocks (Idem: p. 217). All graphs show the effects up to 8 months in the future. The exact interpretation of the outcomes answers the question: what percentage of solar price change is associated with a 1% change of an oil price change, since the first difference of logged variables is taken. However, it is the shape of the impulse response that conveys the most interesting results in a clear and straightforward way.
Table 9. Granger Causality Wald tests Equation Excluded Chi2 P-value
Solar Oil 7.21** 0.027 Solar S&P 2.61 0.272 Solar Tech 0.86 0.649 Solar All 12.16 0.058 Oil Solar 1.81 0.404 Oil S&P 0.31 0.858 Oil Tech 0.86 0.651 Oil All 3.21 0.782 S&P Solar 2.49 0.287 S&P Oil 11.07*** 0.004 S&P Tech 2.75 0.253 S&P All 15.71** 0.015 Tech Solar 1.76 0.414 Tech Oil 14.57*** 0.001 Tech S&P 2.36 0.308 Tech All 16.25** 0.012 ** significant at 5%, *** significant at 1%
16 The shocks provided by Stata are positive and the impulse response functions show that a shock to oil gives a significant upward response to solar prices. The effect peaks after one month and decreases slowly after until the shock is extinguished around four months later. The hypothesis that oil prices significantly impact stock prices in the solar industry due to a substitution effect can therefore be accepted. Solar energy has become more widely available and accessible and has developed itself as a substitute for oil-based energy. The substitution effect might even be enhanced by the present trend in the use and quality of electric cars. The proportion of energy generated by oil has decreased in the last years, whereas the proportion of oil in transportation has increased (EIA, 2018). In the same years electric cars have been brought to the market and as more facilities are created the popularity rises. Individuals buying electric cars are encouraged to think about what electricity to use for ‘fuelling’ their car. Research has shown that electric car owners are more likely to install solar panels than conventional car drivers (Shahan, 2017). These tendencies might contribute to the demonstrated substitution effect even though the use of oil for electricity generation declines.
Figure 4 Impulse response functions (solar)
Figure 4 shows the impulse response functions (a. orthogonalized, b. generalized and c. the Cholesky forecast-error variance decomposition). Where the log of solar prices is the response variables and oil, S&P 500 index and the Nasdaq 100 technology index are the impulse variables. The shaded area provides a 95% confidence interval of the impulse responses showed by the blue line. The x-axis shows the steps in months, whereas the y-axis shows the change in percentages.
17 Furthermore, an impulse shock to the S&P 500 has a slight negative impact at first after which a positive impact can be seen. A shock to the technology index shows an exactly reversed tendency. However, neither are significant. This is surprising, because it goes against the results in previous research as stated in the theoretical framework. A possible cause for the insignificance of the technology index might be the use of a differently composed index. It could be the case that differences in weighting of countries and industries could generate different effects. Moreover, the effect was found for the renewable energy industry more than a decade ago. Possibly the solar industry is distinguished from other renewable energy resources in that the technology used is very specific and produced by companies in the industry itself rather than by technology companies taken up in the indices. Solar companies might show similar trends to the technology sector, but they do not impact each other or show causal relationships.
In the impulse response functions there is also a shock in solar to show the importance of lagged values of itself. This shows that solar prices react significantly positive to a previous stock price increase. This is in accordance with the conventional theory of momentum effects. The theory states that stocks with rising prices have a tendency to keep rising and a drop in the price is followed by a further drop. This is an anomaly since these effects last longer than should be expected based on the situation and information. Nonetheless, they are encountered throughout industries and years and the momentum effect has become somewhat expected (Moskowitz & Grimblatt, 1999).
From graph c in figure 4 the variance decomposition can be deducted. The variance can be explained by the lagged values of solar and the lagged values of oil. Which are the most important indicators for solar price behaviour within the vector autoregression.
On the next page, the impulses for every variable used in the vector autoregressive model are shown. The momentum effect can be identified in all four variables, they all react positively to a positive shock in their own values. Furthermore, the Granger causality results from table 9 showed a significant response of the S&P 500 and the technology index to oil prices. This makes sense based on previous research from Jones and Kaul (1996), Driesprong et al. (2008) and Kilian and Park (2009). They show that positive shocks to oil prices have a negative as well as a positive impact on stock markets (Jones & Kaul, 1996, p. 483 p. 483). The negative impact can be explained by a decrease in expected cash flows due to the higher production costs for most firms depending on oil. Moreover, it is explained that thereafter a negative as well as positive impact can be caused by demand shocks in oil. This is dependent on whether the cause is precautionary due to concerns about future oil supply in which case there is an additional negative effect or positive due to an expansionary economic environment (Kilian and Park 2009, p. 1268). In this case the second explanation seems to fit as in the period analysed the world economy has recovered from the financial crisis and the gross domestic product and demand are expected to keep growing. The general sentiment on the stock market is positive enough to withstand oil price increases in this period. Additionally, the impact is bigger for technology than for the S&P which is in line with the results from the coefficients of the multifactor regression in section 4. The technology sector reactions are more pronounced than the S&P 500 and behaves more like the solar sector, even though there is no significant causal relationship.
18 Figure 5. Impulse response function (total)
The vector autoregression using the variables solar stock prices, oil prices, a broad market index and a technology index has been used to provide information on the dynamic relationships between these variables. Granger causality tests and impulse response functions have been used to determine which variables cause an impact to other variables and to investigate what kind of, positive or negative, response they give. The most important finding is that the statistical tests show a significant positive result for the effects of oil prices on solar stock prices, indicating a substitution effect. On the contrary, a positive shock to oil prices slightly depresses the broad stock market index, but after 1 month the effect reverses. This knowledge can be used by investors and policymakers. This is useful information for governments and international institutions who can try to use taxes on fossil fuels and financial flows to deter investment from non-renewable energy and encourage investment in solar energy. Another approach would be to create positive incentive programs for investors to include solar stocks in their portfolio. The disclosed results provide information for investors forming investment portfolio’s as well. When solar stock prices initially move in the opposite direction of the general stock markets when oil shocks materialize, they can be added to the portfolio for hedging purposes as proposed by Sadorsky (2012, p. 253). As mentioned before the relationship is a dynamic one over time and might change if circumstances in either oil price behaviour, general markets or the renewables industry change. Therefore, further research should monitor the relationship between oil prices and solar stock returns, to see if the substitution effects persists in the rapidly developing solar industry. Eventually, the
Figure 5 shows the impulse response functions (orthogonalized). Where the first variables is the impulse variables and the second the response variables. Solar stands for the log of solar prices, oil for the log of oil prices, S&P for the log of the S&P 500 index and Tech for the log of the Nasdaq 100 technology index. The shaded area provides a 95% confidence interval of the impulse responses showed by the blue line. The x-axis shows the steps in months, whereas the y-axis in a. shows the change in percentages.
19 substitution effect might vanish when the solar industry has significantly matured to a point where oil prices and solar energy are no longer substitutes. As a consequence of cost efficiency advancements, energy consumers might at a point have transferred away from non-renewables. Or due to increasing concerns about global warming and the effects that are created using non-renewable energy and fossil fuels, individual consumers as well as companies may favour solar power based on principles despite oil price behaviour. In the meantime, other factors influencing investment in solar and green energy in general should be investigated to update knowledge on the behaviour of assets and to aid the progress towards renewables. Furthermore, it would be valuable if the effects of other non-renewables, like natural gas or nuclear energy, would be taken into account as well. This research is based on a limited number of variables and many interesting relationships may exist that provide us with a deeper understanding of the dynamics in the industries. This is where future research can provide an extension to current insights.
7. Robustness check
So far, I have used monthly data on prices to run the vector autoregression model. To check if the results are robust, the same tests are run using monthly returns and weekly data. There are no differences in significant results when monthly returns instead of the monthly prices are used. In this case it is unnecessary to use the first differences, because the returns are stationary and do not contain a unit root (Appendix, 10). On the contrary, there are several striking differences using weekly logged prices. Most importantly, the Granger causality statistics show no causal effect of oil prices on solar stock prices (Appendix, 7). There are two explanations for the different outcomes. First, weekly data might still contain too much noise. There are ongoing discussions about which data series are best to use in general, but it is often argued that there is a deterministic trend of holidays and events in weekly data. The time series retrieved from DataStream and the Federal Reserve are not seasonally adjusted, which might lead to problems in weekly data. Second, though the pre-estimation tests show mostly similar outcomes, the post-estimation test reveals a problem with the data. Looking at the vector autoregression post-estimation Jarque-Bera test for normality, the null hypothesis of normality is rejected at the 1% level. These issues might have caused the different outcomes from monthly returns. Furthermore, different lag lengths have been tested, but all optimal lags as chosen by one of the information criteria lead to the same Granger causality significance of test statistics.
Additionally, two vector autoregressions have been run including wind energy as the second biggest renewable energy resource. A five-variable regression where wind is added and a four-variable regression where wind is substituted for solar to see if similar results are retrieved. Indeed, Granger causality tests show that oil has a significant impact on wind energy as well as on solar energy (Appendix, 9). The impulse response functions show that they are also in the same direction, indicating that the substitution effect does work for wind energy as well as solar energy. Solar and wind energy do not have
Table 10. Jarque-Bera test for normality (weekly)
Equation Chi2 P-value
Solar 1.617 0.445 Oil 13.056*** 0.001 S&P 18.228*** 0.000 Tech 91.875*** 0.000 All 124.777*** 0.000 *** significant at 1%
20 any causal effect on each other. Which indicates they are used as complements to each other to provide green energy rather than that they are seen as substitutes of one another.
8. Conclusion
Energy security issues and environmental concerns have translated to national and international challenges and antipollution measures. In the Paris Agreement many countries agreed on goals including the use of financial flows and techniques to stimulate a transition towards renewable energy. Especially solar energy has advanced in recent years and is able to generate electricity more cost efficiently. This has lead to a wider adoption of solar energy. Understanding the dynamics in the solar industry is important for investors as well as governments who want to produce sound incentive policies for the transition away from fossil fuels. To capture the dynamic relationship between solar stock prices and oil prices a four-variable vector autoregression is run, including the effects of the broad market and a technology index. This research contributes to the existing literature by focussing on solar energy individually instead of on the whole array of renewable energy. Furthermore, the data is updated to recent years which have shown different tendencies in oil price behaviour. A significant and positive effect of oil price shocks to solar stock prices has been found. This underlines the hypothesis of the substitution effect arising between these assets. The same oil price shocks have a significantly negative effect on the S&P 500 and the technology stock index in the first month and reverses afterwards. This is in line with previous research and expectations. Similar results have been found when running the vector autoregressive model on wind energy. This indicates that a transition towards renewable energy can be aided by creating oil price shocks. Governments and international organisations can use this information to give a nudge in the desired direction. Investors can use the knowledge about the assets to create strategic portfolio’s and hedge risk related to oil price shocks. It would be interesting if further research conducted in the years to come will focus on different non-renewable sources, like natural gas, and at the effects of negative oil price shocks. A substitution from solar energy back to fossil fuels when prices are dropping seems unlikely given the large share of upfront costs of solar installations. After the initial investments have been made, the generation of solar energy does not require additional costs except for maintenance. Furthermore, studies on this relationship on different countries can be conducted to see if high oil consuming or producing countries show different relationships between oil and solar. For example, influential countries like China are important oil consumers that face severe pollution problems and started investment programs in the solar industry. Besides doing research on those countries, research on the effects between different industries within renewables, like geothermal, hydropower and biomass, can be conducted. Even though nuclear power is officially not a renewable energy source, it will be worthwhile studying it’s effects and influences since China, an influential player in world markets, has plans to start large scale nuclear power plant operations. When looking at different industries there might be competition and substitution between renewables and other energy sources. This is all important knowledge when trying to advocate and increase investment in renewables.
21
Literature
Broadstock, D. C., & Filis, G. (2014). Oil price shocks and stock market returns: New evidence from the United States and China. Journal of International Financial Markets, Institutions and Money, 33, 417-433.
Driesprong, G., Jacobsen, B., and Maat, B. 2008. “Striking Oil: Another puzzle.” Journal of Financial Economics, 89, 307-32
El-Sharif, I., Brown, D., Burton, B., Nixon, B., & Russell, A. (2005). Evidence on the nature and extent of the relationship between oil prices and equity values in the UK. Energy Economics, 27(6), 819-830.
Escanciano, J. C., Lobato, I. N., & Zhu, L. (2013). Automatic specification testing for vector autoregressions and multivariate nonlinear time series models. Journal of Business & Economic Statistics, 31(4), 426-437.
Graham, J. R., & Harvey, C. R. (2001). The theory and practice of corporate finance: Evidence from the field. Journal of financial economics, 60(2-3), 187-243.
Haug, M. (2011). Clean energy and international oil. Oxford Review of Economic Policy, 27(1), 92-116. Henriques, I., & Sadorsky, P. (2008). Oil prices and the stock prices of alternative energy
companies. Energy Economics, 30(3), 998-1010.
International Energy Agency (IEA) (2017, 17 October). Renewables 2017. Retrieved on 20-5-2018. https://www.iea.org/publications/renewables2017/#section-4-2
Investment spikes in 2014. (2015). Renewable Energy Focus, 16(1), 38-39.
Jones, C. M., & Kaul, G. (1996). Oil and the stock markets. The Journal of Finance, 51(2), 463-491. Kilian, L. (2009). Not all oil price shocks are alike: Disentangling demand and supply shocks in the crude
oil market. American Economic Review, 99(3), 1053-69.
Kilian, L., & Park, C. (2009). The impact of oil price shocks on the US stock market. International Economic Review, 50(4), 1267-1287.
Kumar, S., Managi, S., & Matsuda, A. (2012). Stock prices of clean energy firms, oil and carbon markets: A vector autoregressive analysis. Energy Economics, 34(1), 215-226.
MacKinlay, A. C. (1987). On multivariate tests of the CAPM. Journal of Financial Economics, 18(2), 341-371.
Malinowski, M., Leon, J. I., & Abu-Rub, H. (2017). Solar photovoltaic and thermal energy systems: current technology and future trends. Proceedings of the IEEE, 105(11), 2132-2146.
McCrone, A., Usher, E., Sonntag-O’Brien, V., Moslener, U., & Grüning, C. (2015). Global trends in renewable energy investment 2015. Frankfurt School UNEP Collaborating Centre for Climate
and Sustainable Energy Finance.
McCrone, A., V., Moslener, U., D’Estais, F.& Grüning, C. (2017). Global trends in renewable energy investment 2015. Frankfurt School UNEP Collaborating Centre for Climate and Sustainable
22 Moskowitz, T. J., & Grinblatt, M. (1999). Do industries explain momentum? The Journal of
Finance, 54(4), 1249-1290.
Park, J., & Ratti, R. A. (2008). Oil price shocks and stock markets in the US and 13 European countries. Energy economics, 30(5), 2587-2608.
Perold, A. F. (2004). The capital asset pricing model. Journal of economic perspectives, 18(3), 3-24. Pilbeam, K. (2005). Finance and financial markets. Macmillan International Higher Education. Sadorsky, P. (2012a). Modeling renewable energy company risk. Energy Policy, 40, 39-48.
Sadorsky, P. (2012b). Correlations and volatility spillovers between oil prices and the stock prices of clean energy and technology companies. Energy Economics, 34(1), 248-255.
Shahan, Z. (2017, June 25). 28–40% of EV Drivers Have Solar Panels. Retrieved from: https://cleantechnica.com/2017/06/25/28-40-ev-drivers-solar-panels-cleantechnica-ev-report/
Smith, E. K., & Mayer, A. (2018). A social trap for the climate? Collective action, trust and climate change risk perception in 35 countries. Global Environmental Change, 49, 140-153.
Stock, J. H., & Watson, M. W. (2015). Introduction to econometrics: Global edition (Third updated edition). Essex, Harlow: Pearson Education.
Soyemi, K. A., Akingunola, R. O. O., & Ogebe, J. (2017). Effects of oil price shock on stock returns of energy firms in Nigeria. Kasetsart Journal of Social Sciences.
Toda, H. Y., & Yamamoto, T. (1995). Statistical inference in vector autoregressions with possibly integrated processes. Journal of econometrics, 66(1-2), 225-250.
United Nations (2015). Paris agreement. Retrieved on 25-4-2018.
https://unfccc.int/files/essential_background/convention/application/pdf/english_paris_agree ment.pdf
US energy information administration (EIA) (2017, 26 October). Use of natural gas. Retrieved from: https://www.eia.gov/energyexplained/index.php?page=natural_gas_use
US energy information administration (EIA) (2018, 16 May). US energy facts. Retrieved from: https://www.eia.gov/energyexplained/?page=us_energy_home
Wang, Y., Wu, C. & Yang, L. (2013.) Oil price shocks and stock market activities: Evidence from oil-importing and oil-exporting countries. Journal of Comparative Economics, 41, 1220-1239.
23
Appendix
Summary of variable abbreviations
WROilRf = Weekly returns for oil prices – risk-free rate WRSolRf = Weekly returns for solar stocks – risk-free rate WRSPRf = Weekly return on the S&P 500 – risk-free rate
WRTechRf = Weekly return on the Nasdaq 100 technology index – risk-free rate
L is used to indicate that the logarithm of the variable is used, whereas an R stands for Returns. 1. Descriptive Statistics
The first table shows the descriptive statistics for the weekly data and the second for the monthly data used.
2. Regression multifactor model 1 (Return Solar, Return S&P, Return Oil)
The regression results are based on weekly data.
_cons .0012819 .0017055 0.75 0.453 -.0020767 .0046406 WSPRf .3200339 .1132225 2.83 0.005 .0970677 .5430001 WRoilRf .2476719 .045888 5.40 0.000 .1573059 .3380379 WRsolRf Coef. Std. Err. t P>|t| [95% Conf. Interval] Total .223812142 258 .000867489 Root MSE = .02733 Adj R-squared = 0.1391 Residual .191185289 256 .000746818 R-squared = 0.1458 Model .032626853 2 .016313427 Prob > F = 0.0000 F(2, 256) = 21.84 Source SS df MS Number of obs = 259
Wind 60 11.38716 1.647217 6.94 13.53 Sol 60 30.95108 4.367321 20.35815 38.3175 Tech 60 2386.974 643.067 1394.75 3940.48 SP500 60 2025.822 281.0412 1426.19 2642.22 Oil 60 66.77617 25.13375 30.32 106.57 Variable Obs Mean Std. Dev. Min Max WRf 261 .0024808 .0033151 -.0002 .0128 WTech 261 2410.192 652.5455 1437.01 4094.04 WSP 261 2036.836 276.3942 1466.47 2683.34 WOilPrice 261 66.79824 25.03746 28.14 108.77 WSolarPrice 261 31.69171 4.469967 20.11909 39.56973 Variable Obs Mean Std. Dev. Min Max
24 3. Regression multifactor model 2 (Return Tech, Return S&P, Return Oil)
The regression results are based on weekly data.
4. Regression multifactor model 1 absolute values (Log Solar Price, Log S&P Price, Log Oil Price)
The regression results are based on weekly data.
5. Regression multifactor model 2 absolute values (Log Technology Index, Log S&P Price, Log Oil Price)
The regression results are based on weekly data.
_cons -6.760341 .3991151 -16.94 0.000 -7.546279 -5.974403 SP500 1.177056 .0461097 25.53 0.000 1.086257 1.267855 Oil .3021716 .0169752 17.80 0.000 .268744 .3355992 Solar Coef. Std. Err. t P>|t| [95% Conf. Interval] Total 5.68454126 260 .02186362 Root MSE = .07886 Adj R-squared = 0.7155 Residual 1.60466138 258 .006219618 R-squared = 0.7177 Model 4.07987988 2 2.03993994 Prob > F = 0.0000 F(2, 258) = 327.98 Source SS df MS Number of obs = 261
_cons -6.944486 .2248127 -30.89 0.000 -7.387187 -6.501784 SP500 1.916833 .0259725 73.80 0.000 1.865688 1.967978 Oil .0267722 .0095618 2.80 0.005 .0079432 .0456013 Tech Coef. Std. Err. t P>|t| [95% Conf. Interval] Total 17.6551415 260 .067904391 Root MSE = .04442 Adj R-squared = 0.9709 Residual .509129928 258 .001973372 R-squared = 0.9712 Model 17.1460116 2 8.57300581 Prob > F = 0.0000 F(2, 258) = 4344.34 Source SS df MS Number of obs = 261 . _cons .0023062 .0013112 1.76 0.080 -.0002758 .0048882 WSPRf .3462841 .0870413 3.98 0.000 .1748759 .5176924 WRoilRf .1684235 .035277 4.77 0.000 .0989535 .2378936 WRTechRf Coef. Std. Err. t P>|t| [95% Conf. Interval] Total .133542138 258 .000517605 Root MSE = .02101 Adj R-squared = 0.1473 Residual .112989955 256 .000441367 R-squared = 0.1539 Model .020552183 2 .010276091 Prob > F = 0.0000 F(2, 256) = 23.28 Source SS df MS Number of obs = 259
25 6. Regression and Durbin Watson test statistic (Return solar, Return Oil, Return S&P Return Technology)
The regression results are based on monthly data.
7. Weekly data: Granger Causality Walt tests after VAR prices (First Difference)
8. Stability condition vector autoregression
The results are based on monthly data.
D_Tech ALL 1.0077 3 0.799 D_Tech D.Oil .11524 1 0.734 D_Tech D.SP500 .89429 1 0.344 D_Tech D.Solar .00986 1 0.921 D_Oil ALL 13.181 3 0.004 D_Oil D.Tech 1.9188 1 0.166 D_Oil D.SP500 .24759 1 0.619 D_Oil D.Solar 2.2683 1 0.132 D_SP500 ALL .12097 3 0.989 D_SP500 D.Tech .04291 1 0.836 D_SP500 D.Oil .02103 1 0.885 D_SP500 D.Solar .03307 1 0.856 D_Solar ALL 93.846 3 0.000 D_Solar D.Tech 3.7853 1 0.052 D_Solar D.Oil 1.307 1 0.253 D_Solar D.SP500 10.123 1 0.001 Equation Excluded chi2 df Prob > chi2 Granger causality Wald tests
VAR satisfies stability condition.
All the eigenvalues lie inside the unit circle. -.03537337 - .1574916i .161415 -.03537337 + .1574916i .161415 -.2178903 .21789 .2493333 .249333 Eigenvalue Modulus Eigenvalue stability condition
26 9. Five/Four variable Granger Causality Wald test including wind energy
The results are based on monthly data.
10. Monthly data: Granger Causality Walt tests after VAR using returns Using the returns on all variables (no unit root).
*** significant at 1% RTechRf ALL 12.275 3 0.006 RTechRf RSPRf .08335 1 0.773 RTechRf ROilRf 12.256 1 0.000 RTechRf RSolRF .77921 1 0.377 RSPRf ALL 8.4642 3 0.037 RSPRf RTechRf .37781 1 0.539 RSPRf ROilRf 7.8833 1 0.005 RSPRf RSolRF .20848 1 0.648 ROilRf ALL 1.0647 3 0.786 ROilRf RTechRf .73267 1 0.392 ROilRf RSPRf .09884 1 0.753 ROilRf RSolRF .0133 1 0.908 RSolRF ALL 11.015 3 0.012 RSolRF RTechRf .47748 1 0.490 RSolRF RSPRf 2.3576 1 0.125 RSolRF ROilRf 6.5524 1 0.010 Equation Excluded chi2 df Prob > chi2 Granger causality Wald tests
Granger Causality Wald tests (four-variable wind) Equation Excluded Chi2 P-value
Wind Oil 12.37*** 0.000
Wind S&P 0.04 0.837
Wind Tech 0.52 0.471
Wind All 15.36 0.002
Granger Causality Wald tests (five-variable wind) Equation Excluded Chi2 P-value
Wind Solar 0.43 0.511
Wind Oil 14.72*** 0.000
Wind S&P 0.09 0.768
Wind Tech 0.04 0.838
27 11. Vector autoregression (First difference monthly prices)
_cons .0229602 .0062605 3.67 0.000 .0106899 .0352305 L2D. .2111414 .2115506 1.00 0.318 -.2034902 .6257729 LD. -.1458515 .2113824 -0.69 0.490 -.5601534 .2684505 LTech L2D. -.4626106 .3061645 -1.51 0.131 -1.062682 .1374608 LD. .058433 .2954691 0.20 0.843 -.5206758 .6375418 LSP L2D. -.009224 .0726455 -0.13 0.899 -.1516065 .1331585 LD. .2351107 .0649421 3.62 0.000 .1078266 .3623949 LOil L2D. .047886 .1246029 0.38 0.701 -.1963311 .2921032 LD. -.1645845 .1286677 -1.28 0.201 -.4167685 .0875995 LMS D_LTech _cons .0142439 .0043054 3.31 0.001 .0058055 .0226823 L2D. .2324888 .1454856 1.60 0.110 -.0526578 .5176354 LD. -.0232167 .14537 -0.16 0.873 -.3081366 .2617032 LTech L2D. -.5240391 .2105526 -2.49 0.013 -.9367147 -.1113635 LD. -.1475445 .2031973 -0.73 0.468 -.5458039 .2507148 LSP L2D. .0235153 .0499591 0.47 0.638 -.0744027 .1214332 LD. .1305145 .0446614 2.92 0.003 .0429798 .2180492 LOil L2D. .0892434 .0856907 1.04 0.298 -.0787073 .2571942 LD. -.1069009 .0884861 -1.21 0.227 -.2803306 .0665287 LMS D_LSP _cons .002349 .0134402 0.17 0.861 -.0239932 .0286913 L2D. -.2504398 .4541634 -0.55 0.581 -1.140584 .639704 LD. -.3760351 .4538023 -0.83 0.407 -1.265471 .5134011 LTech L2D. -.3067517 .6572834 -0.47 0.641 -1.595004 .9815002 LD. .1753282 .6343223 0.28 0.782 -1.067921 1.418577 LSP L2D. .0009878 .1559575 0.01 0.995 -.3046833 .3066589 LD. .2900392 .1394197 2.08 0.037 .0167816 .5632968 LOil L2D. .3567256 .2675013 1.33 0.182 -.1675673 .8810185 LD. .0437192 .2762277 0.16 0.874 -.4976772 .5851156 LMS D_LOil _cons .0119209 .0081935 1.45 0.146 -.004138 .0279798 L2D. -.1141498 .2768696 -0.41 0.680 -.6568042 .4285046 LD. .2074339 .2766495 0.75 0.453 -.3347891 .7496569 LTech L2D. .2000784 .4006968 0.50 0.618 -.5852728 .9854297 LD. -.5836663 .3866991 -1.51 0.131 -1.341582 .17425 LSP L2D. .0295167 .0950757 0.31 0.756 -.1568282 .2158616 LD. .2033448 .0849938 2.39 0.017 .03676 .3699296 LOil L2D. .158358 .1630756 0.97 0.332 -.1612644 .4779803 LD. -.0269688 .1683955 -0.16 0.873 -.3570179 .3030803 LMS D_LMS Coef. Std. Err. z P>|z| [95% Conf. Interval]
