Carbon politics: The effect of regulatory events on carbon emission prices in the European Emissions Trading System

12-6-2019

Abstract

This paper investigates the impact of regulatory events and Brexit proceedings in the European Union on the abnormal returns of EU carbon allowances in the period between December 2011 and April 2019. In order to estimate expected returns, we use a relatively new econometric technique called Dynamic Model Selection, which allows the predictive model to evolve over time. We found significant abnormal returns as the carbon allowance market reacted to Brexit, and measures taken by the EU to postpone scheduled allowance auctions (“backloading”). Furthermore, the market’s reaction to

‘negative’ events was markedly stronger than to ‘positive’

events, showing that the market considers regulatory events to bring high downside risks. These findings show that political processes and decisions have the potential to be a significant price driver in the carbon market, reinforcing the political responsibility towards investors and the success of the system in their actions.

Keywords: Dynamic Model Selection, Event study, EU ETS, Political uncertainty, Allowance price, Backloading, Brexit

Carbon politics: The effect of regulatory events on carbon emission prices in the European Emissions Trading System

Tim Hendriks, BSc.

S2390612 T: +31 6 83 20 74 52 E: T.Hendriks.3@student.rug.nl MSc Finance Faculty of Economics and Business University of Groningen Thesis supervisor: Dr. A. Dalò Word count excluding appendices: 14879

2

Introduction

In an effort to curb climate change by limiting the emission of greenhouse gases in the European Union, the EU implemented the first phase of the European Emissions Trading System (ETS) in 2005. The EU ETS is a cap-and- trade measure limiting carbon emissions EU wide. It is the first and largest emissions trading system globally, currently covering 11.000 power- and manufacturing plants, as well as aviation activities, in 31 European countries. In 2018, 1.754 million tons of CO2 (and equivalent greenhouse gases) have been emitted by participants in the EU ETS. Together they are responsible for 45% of European greenhouse gas emissions, and almost 5% of global emissions, which reached 37,1 billion tons in 2018 (European Commission, 2019).

The EU ETS auctions emission allowances up to a previously determined cap. Companies then have to cover their yearly carbon emissions with allowances. These can be bought and sold on the free market, which establishes a market price based on supply and demand. In this system, investment in carbon reduction becomes efficient if a company’s abatement cost¹ for one ton of CO2 is lower than the cost of a carbon allowance. Companies that can abate CO2 for less than the cost of an allowance will do so, while companies for which CO2 abatement is costly are not forced to take carbon reducing measures (Boyce, 2018; Fuss et al, 2018; Knopf et al., 2018). In a completely efficient market, the market price of the allowance is therefore equal to the marginal cost of abating one ton of CO2 in the whole market (Fuss et al., 2018), leading to greenhouse gas reduction at the lowest possible price. By reducing the total amount of available allowances over time, the EU can reach its primary goal: decreasing carbon emissions throughout the EU. A related secondary goal of the EU ETS is incentivizing the investment in carbon decreasing investments, as a higher price for carbon emissions make alternative sustainable investments more worthwhile (Boyce, 2018). Changing public opinion regarding the necessity of taking measures against climate change further increases the need for a credible and robust system to tackle carbon emission reduction.

The opinions regarding the performance of the EU ETS however, have been mixed. A period of low carbon prices between 2012 and 2017, followed by a rapid jump in 2018 (which can be observed in figure 1) has called the

1 Abatement cost refers to the cost of reducing one ton of CO² emissions

Figure 1: This graph shows the price of EU carbon allowances during the investigated period.

3 effectiveness of the carbon trading system into question. The low carbon prices caused the system to be less effective at reducing emissions than expected (Salant, 2016), while the price volatility in 2018 caused unrest in the industry (Point Carbon, 2018).

Research in environmental economics has traditionally explained allowance prices by developing a model in order to estimate the marginal abatement cost. However, the marginal abatement cost could only explain a small amount of variance in allowance prices, and did little to explain large price breaks occurring in ETS prices (Koch, Grosjean, Fuss & Edenhofer, 2016). What is missing from models based on marginal abatement cost is the effect that regulatory events can have on price formation. Salant and Henderson (1978) used the Hotelling model in order to predict the effects of regulatory events in a market where these events can have a large effect on price.

The influence of regulators is magnified by the fact that the carbon market is not just regulated by the (European) government, but completely created by it, as the price on carbon emissions is manufactured through regulation (Newell, Pizer & Raimi, 2014). This causes the high impact that government regulation has in the carbon market – policy revisions can strongly alter the value of financial assets (such as the carbon allowances) and financial liabilities of the affected parties (Newell et al., 2014).

The ability of policymakers to alter the carbon cap reduces the ability of allowance holders to accurately predict the cap in the future, introducing a significant risk for the participants, which reduces the price they are willing to pay (Koch et al., 2016). Newell et al. (2014) predict that a large amount of price formation in the emissions market is driven by regulatory events and policy making, due to the associated risk. Announcements that change the market’s perception regarding changes in the cap can therefore trigger large price jumps or falls. Decreasing regulatory uncertainty is therefore an important step towards more effective measures against climate change, and improves the chances of meeting the climate goals of the Paris agreement². This leads to the research question of this proposed research:

What is the influence of regulatory and political events on carbon emissions prices in the European Emissions Trading System?

This paper aims to answer this question using an event study methodology to examine the abnormal returns generated due to announcements of 55 regulatory events in the period between 1-11-2011 and 31-4-2019.

Answering the research question helps to understand and improve the European ETS, as well as cap-and-trade carbon reduction initiatives around the world. The chosen time period allows this paper to replicate and validate the results of Koch et al. (2016), while adding knowledge to the literature by testing the hypothesis using data in a timeframe that has not been investigated before.

The choice for an event study methodology is made as this method is ideal to measure the impact of an economic event on the price of an asset (MacKinlay, 1997). Event studies utilize the semi-strong form efficient market

2 The Paris Agreement was signed in 2015 and ratified by 195 countries, binding these countries to their commitment to

‘holding the increase in the global average temperature to well below 2 °C above pre-industrial levels and pursuing efforts to limit the temperature increase to 1.5 °C’ (Boyce, 2018; Schleussner et al., 2016)

4 hypothesis, which stipulates that any public information is immediately reflected in the price of an asset. This means that the effect of this additional information can be measured in a short timeframe (MacKinlay, 1997). The studied events are divided in four categories: Events related to implementing the backloading³ of allowances, the implementation of a Market Stability Reserve⁴, medium and long term policy, and Brexit.

Brexit is a distinctive category from the other three, as it mainly contains events that are not necessarily directly linked to the ETS, and furthermore include events that were induced unilaterally by the UK, instead of by the legislative bodies of the EU. The reason for including these events are the high impact of the UK leaving the union (and in which manner) on the ETS regarding the available allowances, whether UK firms will still be bound by previous regulation, and whether the UK would implement a linked, separate, or no ETS at all after Brexit.

This paper will first present an overview of the literature on price formation in the EU ETS, after which the research methods and data collection will be discussed. In the next section, we provide details about the results found, after which the implications of the results are reviewed in the discussion section, followed by the conclusion of this research.

Literature review

Ideally, a cap-and-trade scheme, such as the EU ETS, is an efficient solution for curbing emissions to reach established goals in an ideal world (Salant, 2016). The policy has also brought desirable effects, such as an increase in patenting of low carbon technologies (Calel & Dechezlepretre, 2016), and a decrease in emissions to put it on track to meet its goal to reduce carbon emissions by 20% in 2020, relative to 1990 (European Commission, 2016). However, the EU ETS has not been able to bring about all of the desired effects. As the EU ETS entered its third stage in 2013, the market was heavily oversupplied. This was partially caused by the results of the economic crisis driving down demand, and partially due to the lenient initial stages of the ETS in which many allowances were distributed for free, instead of auctioned, and yearly EU emissions were markedly lower than the cap.

The oversupply lead to low allowance prices in the market, which was the main cause of ineffectiveness of the system. Emitting one metric ton of CO2 has cost between €5 and €8 between 2013 and the end of 2017, instead of the expected €25 to €39 (European Commission, 2009; Salant, 2016).

Climate policy analysts recommend significantly higher prices than found in the ETS market in the past years, as low prices do not provide enough incentives to move away from carbon emitting technologies (Boyce, 2018).

Low carbon prices also lead to a lock-in problem for short-term oriented companies, as they invest in long-term carbon-intensive infrastructure with low carbon prices in mind (Knopf et al., 2018). Lastly, the low carbon prices

3 “Backloading” of emission allowances was proposed by the European Parliament committee on the environment in 2011, in response to low allowance prices threatening the effectiveness of the EU ETS. The implemented backloading proposal pushed back the auction date of 900 million allowances from 2014-2016 to 2019-2020 in order to curb short term supply and increase allowance prices.

4 The Market Stability Reserve (MSR) was proposed in 2015 and introduced a mechanism in the EU ETS that would store allowances in times of low demand, which can be auctioned later in times of high demand. The MSR entered into force in 2019 and is meant to increase the system’s resilience to shocks and stabilize the allowance prices.

5 have been damaging to the reputation of the project, increasing political uncertainty (Boyce, 2018; Knopf et al., 2018; Salant, 2016)

In 2018, the price of carbon allowances rose from €8 to €25 during the year. While this puts the carbon price close to effective levels, this introduces a second issue. If the price of carbon allowances rises quickly, especially after a long period of low prices, the rise will lead to a sudden increase in costs for business and consumers. This leads to a difficult political situation for governments, who are very likely to receive political backlash (Knopf et al., 2018). Relaxation of the emissions cap to bring carbon prices down then becomes an attractive opportunity for politicians. An example in practice was the 2008 U.S. presidential campaign, when candidates Hillary Clinton and John McCain were campaigning during times of high fuel costs. Both candidates were public proponents of cap-and-trade systems, but both called for a temporary relief of gasoline tax under political pressure, in order to bring relief to the public suffering the high prices (Bosman, 2008; Boyce, 2018).

Another recent example is Ontario’s 180 degree turn on carbon trading. After the previous premier Kathleen Wynne entered the Canadian province into the Western Climate Initiative (WCI) and instituted a provincial cap- and-trade carbon emissions system, newly elected premier Doug Ford immediately cancelled the system and withdrew from the WCI. This abruptly impacted the overall carbon supply and demand in the WCI, depressing carbon prices and leaving Ontario emitters with worthless allowances (Point Carbon, 2018).

Newell et al. (2014) showed that the market is unable to have full confidence in political commitments made in the introduction of carbon tax or cap-and-trade schemes. Incentives to renege on prior commitments remains as policy makers also require a degree of flexibility in order to counter unexpected circumstances that render the system less effective. The “backloading” decision made by the EU, also investigated in this paper, is an example of policy makers coming back on previously agreed upon auction schemes in an effort to boost the allowance price, thereby impacting the system’s participants who are affected by such price changes. Furthermore, a vulnerability of the EU ETS is the fact that allowance holders are in no way protected from changes to the system that reduce the value of their allowances, requiring participants to trust the legislators to not introduce changes to the system that destroy value. This harms the effectiveness of the EU ETS by both driving down the allowance price and introducing additional volatility (Libecap, 2013).

Market participants will take the expected probability of an intervention in to account, as well as the expected price effect of an intervention when valuing carbon allowances. This behaviour is described by the Hotelling model (Salant & Henderson, 1978). Investors in this model expect a regulatory intervention will occur in the future, which can have positive or negative price effects. Investors base their prediction of the future allowance price on the probability of good news and the associated probability of bad news, and the allowance prices resulting from good or bad news (Koch et al., 2016). This results in the expected price changing if investors’

expectations about the probability of good news versus bad news changes, even if no regulation has changed. In a market where risks of a price collapse (downside risks) are dominant, the mean price change caused by news should be negative, and negative price events will lead to stronger price reactions than positive events (Salant &

Henderson, 1978).

6 Koch et al. (2016) draw a parallel between the gold and carbon market⁵, as the value of a carbon allowance is also driven for a large part by the possibility of government intervention, causing price jumps at news announcements regarding ETS policies. In these cases, that means that even speculation regarding a change in political will to change the ETS system will cause price falls (Salant, 2016).

Government intervention, even uncertainty about intervention, in cap-and-trade systems thus is damaging to the system, because they reduce the trust that participants have in the system (Koch et al., 2016). Uncertainty about whether the government might intervene in the system discourages investment in costly green technologies. If the government intervenes during periods of high prices in order to bring these prices down, the costly investment suddenly loses all its value, because the opportunity cost of using the carbon-intensive technology decreases together with the carbon price. This causes companies to wait-and-see in cases of regulatory uncertainty, instead of investing in green technology, driving carbon prices down due to uncertainty alone (Fuss et al., 2018; Grosjean, Acworth, Flachsland & Marschinski, 2016). Even regulatory events aimed at increasing the carbon price, such as the EU deciding to backload allowances eventually caused price decreases due to the signal that the EU ETS was susceptible to policy changes (Koch et al., 2016).

In this thesis, we will investigate the effect of political events by predicting the expected returns around announcements of significant policy alterations in the EU ETS, and comparing these with the actual returns to find whether the news caused abnormal returns in a seven day window.

H1: Regulatory announcements about medium- to long term policy packages regarding the EU ETS will cause abnormal returns of carbon allowances in the ETS.

H2: Announcements regarding the exit of the United Kingdom from the European Union and its ETS will cause abnormal returns of carbon allowances in the ETS

Research method and data

Event study methodology

This paper will utilize event studies in order to measure the impact of events on the value of stocks, securities or tradeable assets such as (carbon) allowances. With this approach, we adopt the methodology used by Koch et al.

(2016) used to determine the effect of regulatory events on the carbon allowance price from 2008 to 2014. Event studies are currently the standard in the fields of economics and finance as the methodology to investigate price changes caused by events (MacKinlay, 1997). Because the generation of expected returns is critical to a dependable outcome of the event study, we will discuss the metrics and the return generating process below.

5 As countries moved away from the gold standard, governments could start auctioning their vast gold reserves. The result of this was the gold price becoming very susceptible to any news implying an increased likelihood of a government gold auction, regardless of the auction actually happening (Salant & Henderson, 1978).

7 (1)

(2) Metrics

An event study compares the abnormal returns of a security because of an event, with the expected returns predicted during the same time frame without the conditioning of the event (MacKinlay, 1997). This allows isolating the effect of the event from the expected value deviations during the event window. In algebraic terms:

𝐴𝐴𝐴𝐴_𝑡𝑡 = 𝑟𝑟_𝑡𝑡− 𝐸𝐸(𝑟𝑟_𝑡𝑡|Ω_𝑡𝑡)

In this formula, 𝐴𝐴𝐴𝐴𝑡𝑡 represents the abnormal returns in time period t, 𝑟𝑟𝑡𝑡 represents the realized returns, and 𝐸𝐸(𝑟𝑟_𝑡𝑡|Ω_𝑡𝑡) stand for the predicted expected return in the time period (Koch et al., 2016). The expected return is unobservable and therefore will have to be estimated using predictors. Ω𝑡𝑡 represents the conditioning information set, the value of the predictors during time period t, (Koch et al., 2016). In order to draw a conclusion about the impact of the event throughout the complete event window, the daily returns are summed to create the cumulative abnormal return (CAR) of the event.

As mentioned before, the impact of new information will have an immediate impact on the value of the security as the event happens, following the weak-form efficient market hypothesis (MacKinlay, 1997). However, a wider event window is used to capture investor anticipation of a certain event outcome and/or leaked information before the event. Examining the full event window is especially important in regulatory events, as information leakage from committees or other legislative bodies prior to an announcements is very much in the realm of possibility (Koch et al., 2016). The delayed transmission of the news and initial investor over- or under-reaction become visible in the days after the announcement day, especially as news regarding EU regulations is often the subject of political debate in the following days (Koch et al., 2016). The length of the event window should be long enough to capture most event related price distortions, while not being susceptible to other unrelated events generating abnormal returns during the event window (McWilliams & Siegel, 1997). Conforming to MacKinlay (1997) and Koch et al. (2016) we will not only look at the day on which the event occurred, but use a wider event window of 7 days, from day -3 to day 3, as is customary in the literature.

In the event window, the cumulative abnormal returns of the EU allowance are calculated by summing the abnormal returns over the event window.

𝐶𝐶𝐴𝐴𝐴𝐴_{𝜏𝜏1,𝜏𝜏2}= � 𝐴𝐴𝐴𝐴_𝑡𝑡

𝜏𝜏2

𝑡𝑡 = 𝜏𝜏1

In accordance with a 7 day event window, −3 < 𝜏𝜏₁< 3 and −3 < 𝜏𝜏₂< 3, while 𝜏𝜏₁≤ 𝜏𝜏₂. This study looks at four intervals specifically. The interval 𝐶𝐶𝐴𝐴𝐴𝐴_−3,−1 is used to assess market pre-announcement anticipation and its reaction to potential information leaks. 𝐶𝐶𝐴𝐴𝐴𝐴0 studies the market reaction on the day of the announcement and is expected to show the largest deviation from zero in the event window. The interval 𝐶𝐶𝐴𝐴𝐴𝐴1,3 is used to examine post-announcement performance, related to price corrections after the announcement. The full event window 𝐶𝐶𝐴𝐴𝐴𝐴_−3,3 is used to assess the overall impact of the event.

8 (4)

(5)

(6) Following McKinlay(1997) and Koch et al.(2016), the expected returns and the variance of the residuals of the expected returns are calculated from the beginning of the dataset, up until the beginning of the event window.

This period is called the estimation window. The variance of the residuals of the expected returns estimates the variance of the abnormal returns. This means that the expected returns and the variance of abnormal returns are calculated using all observations up until three days before the event window of each event. The variance of the residuals of the expected returns is calculated by taking the large sample variance of the residuals of the expected return model during the estimation window.

𝜎𝜎_{𝐴𝐴𝐴𝐴}²_𝑡𝑡 = 1

𝑇𝑇₁− 𝑇𝑇₀− 2 � (𝑟𝑟^𝑡𝑡− 𝐸𝐸(𝑟𝑟_𝑡𝑡|Ω𝑡𝑡))²

𝑇𝑇₁

𝑡𝑡=𝑇𝑇0

The variance across (intervals in) the event window is calculated by multiplying the variance by the amount of days studied.

𝜎𝜎_{𝐶𝐶𝐴𝐴𝐴𝐴}² _{𝜏𝜏1,𝜏𝜏2} = (𝜏𝜏₂− 𝜏𝜏₁+ 1)𝜎𝜎_{𝐴𝐴𝐴𝐴}² _𝑡𝑡

The root of this variance is the standard error of the prediction and is used to test the abnormal returns for significance, following a two-sided t-test (MacKinlay, 1997).

𝜃𝜃𝑖𝑖= 𝐶𝐶𝐴𝐴𝐴𝐴𝜏𝜏₁,𝜏𝜏₂

�𝜎𝜎𝐶𝐶𝐴𝐴𝐴𝐴2 _{𝜏𝜏1,𝜏𝜏2}

When investigating cumulative abnormal returns over multiple events i, the variance of the cumulative abnormal returns is calculated.

𝜎𝜎_{𝐶𝐶𝐴𝐴𝐴𝐴}������2 _{𝜏𝜏1,𝜏𝜏2} = 1

𝑁𝑁²� 𝜎𝜎_{𝐶𝐶𝐴𝐴𝐴𝐴}2 𝑖𝑖,(𝜏𝜏1,𝜏𝜏2)

𝑁𝑁

𝑖𝑖=1

What differentiates this event study from many others is the fact that only one asset is examined, over multiple events, where most event studies investigate multiple assets over one event. This is because it is not possible to use other relevant assets in the emissions trading market except for the EUA price and its correlated derivatives and futures. Therefore, it is not possible to average cumulative returns from multiple assets over one event, increasing standard errors.

Estimating expected returns

Theoretically, expected returns on carbon allowances should be equal to the marginal abatement cost of CO2 in the market, as firms that find it cheaper to reduce CO2 emissions rather than buy allowances would do so.

Measuring the abatement costs are challenging, however. Many different fuel and commodity prices, as well as macroeconomic indicators can be used to predict abatement costs, but there is no consensus about which combination of variables most accurately represents abatement costs (Koch et al., 2016).

In order to estimate the returns that would have materialized had the studied event not occurred, measured by predicting the marginal abatement cost, we follow Koch et al. (2016), who employ a Dynamic Model Selection

(3)

9 (8)

(9) (7) (DMS) approach. DMS uses a relatively small amount of explanatory variables in order to predict the returns of the allowance. DMS allows a large number of potential models, using different subsets of predictors and their weights, which are allowed to change over time. At the beginning of the event window of the studied event, the model which predicted the returns up to that point with the least amount of error, is used to predict the returns in the event window. (Koch et al, 2016). DMS outperforms static models, as models using a limited amount of static variables run the risk of not predicting abatement costs accurately enough, while large models often run into statistical problems such as over-parameterization due to the high number of variables (Koch et al., 2016).

Secondly, as literature pointed out that the abatement cost curve can be discontinuous, volatile and variable during the year (Fezzi & Bunn, 2009), it is unlikely that a static model can always accurately capture abatement costs (Koch et al., 2016). A third reason to choose Dynamic Model Selection is the ability to drop (or add) explanatory variables, or alter the weights of the variables, if these explanatory have decreased (or increased) in importance (Raftery, Kárný & Ettler, 2010; Koop & Korobilis, 2012) .

A DMS approach to predicting expected returns in the event window is preferred by Koch et al. (2016) because of a current lack of a reduced-form, proven model to price abatement costs, and due to the large impact of a wrongly specified model on the results in an event study. Dynamic Model Selection is defined by the use of a set of models to predict expected returns. These models consist of an intercept, a set of explanatory variables over the time period (𝑥𝑥_𝑡𝑡) and a lag of the previous predicted allowance price (𝑦𝑦_𝑡𝑡−1) in order to predict allowance prices, which follow a random walk. This creates the vector of predictors:

𝑧𝑧_𝑡𝑡 = [1, 𝑥𝑥_𝑡𝑡, 𝑦𝑦_𝑡𝑡−1]

Koch et al. (2016) suppose a set of K models can then utilize different subsets of 𝑧𝑧_𝑡𝑡 as predictors. The main ‘model universe’ in this DMS approach consists of two main formulas, the observation equation and the state equation (Raftery et al., 2010). The observation equation (8) predicts the expected returns using the following formula:

𝑦𝑦_𝑡𝑡 = 𝑧𝑧_𝑡𝑡^(𝑘𝑘)𝜃𝜃_𝑡𝑡^(𝑘𝑘)+ 𝜀𝜀_𝑡𝑡^(𝑘𝑘)

In this formula, 𝑦𝑦𝑡𝑡 represents the expected returns of the allowance, 𝑧𝑧_𝑡𝑡^(𝑘𝑘) represents the chosen subset of predictors, 𝜃𝜃_𝑡𝑡^(𝑘𝑘)represents the regression parameters of the chosen predictors, and 𝜀𝜀_𝑡𝑡^(𝑘𝑘) is an error term depicting model innovations, normally distributed with a mean of 0 and a variance of 𝐻𝐻_𝑡𝑡^(𝑘𝑘) (Koch et al., 2016; Rafferty et al., 2010). We will cover the error variance later. The superscript (k) indicates that the quantities indicated are for the specific model 𝑀𝑀𝑘𝑘. The regression parameters of the chosen predictors are determined by the second formula, the state equation (9):

𝜃𝜃_𝑡𝑡^(𝑘𝑘) = 𝜃𝜃_𝑡𝑡−1^(𝑘𝑘) + 𝜂𝜂_𝑡𝑡^(𝑘𝑘)

The regression parameters at time t for the subset of predictors k (𝜃𝜃_𝑡𝑡^(𝑘𝑘)) are given by the state at t-1 (𝜃𝜃_𝑡𝑡−1^(𝑘𝑘)), altered by state innovations represented by 𝜂𝜂_𝑡𝑡^(𝑘𝑘), which is normally distributed with a mean of 0 and a variance of 𝑊𝑊_𝑡𝑡^(𝑘𝑘). These state innovations allow the regression parameters to change over time (Rafferty et al., 2010). For ease of

10 (13) (14) (11)

(12) (10) understanding, this paper will first explain the technique used to predict returns for each subset of predictors (k), and explain the process of selecting the right subset later.

In order to allow the model to converge as closely as possible towards reality, Kalman filtering is used. The Kalman filter uses the prediction equation (10) in order to adjust the model (Rafferty et al., 2010). Assuming the state at t - 1 can be expressed as follows;

𝜃𝜃_𝑡𝑡−1|𝑌𝑌^𝑡𝑡−1~𝑁𝑁(𝜃𝜃�_𝑡𝑡−1, 𝐴𝐴_𝑡𝑡)

in the sample, where 𝑌𝑌^𝑡𝑡−1 represents all time intervals up to t – 1. In order to avoid having to use a transition matrix with the size of 𝑁𝑁², the formula to predict the variance of the coefficients uses a ‘forgetting factor’ 𝜆𝜆, which makes the calculation a lot faster without sacrificing much accuracy. The forgetting factor dictates the weight given to the variance in the state equation (9) at t - 1. This forgetting factor is a number between 0 and 1, but is typically chosen to be slightly under 1 (Raftery et al., 2010).

𝐴𝐴_𝑡𝑡 = 𝜆𝜆⁻¹𝛴𝛴_𝑡𝑡−1

In general terms, the forgetting factor results in time-weighted estimation, where data at time point t has a weight of 𝜆𝜆^𝑡𝑡 (Rafferty et al., 2010). The last equation required to complete the inference is the updating equation:

𝜃𝜃𝑡𝑡|𝑌𝑌^𝑡𝑡~𝑁𝑁(𝜃𝜃�𝑡𝑡, 𝛴𝛴𝑡𝑡)

This equation looks like the prediction equation (10), but uses the following formulas to determine the mean and variance of the normal distribution:

𝜃𝜃�_𝑡𝑡 = 𝜃𝜃�_𝑡𝑡−1+ 𝐴𝐴_𝑡𝑡𝑥𝑥_𝑡𝑡^𝑇𝑇(𝑉𝑉 + 𝑥𝑥_𝑡𝑡^𝑇𝑇𝐴𝐴_𝑡𝑡𝑥𝑥_𝑡𝑡)^𝑇𝑇𝑒𝑒_𝑡𝑡 𝛴𝛴𝑡𝑡 = 𝐴𝐴𝑡𝑡− 𝐴𝐴𝑡𝑡𝑥𝑥𝑡𝑡(𝑉𝑉 + 𝑥𝑥𝑡𝑡𝑇𝑇𝐴𝐴𝑡𝑡𝑥𝑥𝑡𝑡)⁻¹𝑥𝑥𝑡𝑡𝑇𝑇𝐴𝐴𝑡𝑡

In these equations, 𝑒𝑒_𝑡𝑡 represents the one-step-ahead prediction error, given by 𝑒𝑒_𝑡𝑡= 𝑦𝑦_𝑡𝑡− 𝑥𝑥_𝑡𝑡^𝑇𝑇𝜃𝜃�_𝑡𝑡−1. After initializing the formula by specifying 𝜃𝜃�₀ and 𝛴𝛴₀, these values get updated to predict the actual returns with the least amount of error. Following Rafferty et al. (2010), we specify 𝜃𝜃�0 as a vector of zeroes for every model k, while 𝛴𝛴₀ is specified as a vector of the covariances of the predictors. The observations innovations variance V has to be specified by the model user. Rafferty et al. (2010) recommend using a recursive method to determine V. In the results section, we describe how we have initialized our models in terms of forgetting factors α (in equation 18), λ, and initial variance V, and what the influence on the accuracy of the predictions is.

When using multiple models, such as in the DMS approach, there should also be a method to select which subset of predictors, together with their respective states, should be used at what point in time, dictating the final outcome (Rafferty et al., 2010). All possible models in the DMS approach can be expressed using the aforementioned observation equation (8) and state equation (9). The model, with its specific subset of predictors, is ‘governing’ (e.g. driving the results) at a in time is depicted as 𝐿𝐿_𝑡𝑡 = 𝑘𝑘

It is necessary to assume that the governing model changes infrequently, and that its evolution follows a transition matrix of the size 𝐾𝐾 ∗ 𝐾𝐾, with 𝐾𝐾 representing all possible subsets of predictors. This transition matrix

11 (16)

(17)

(18) (15) would have to be specified by the user, which would both be so large as to be infeasible to compute, with there often not being enough information to construct the matrix in the first place (Rafferty et al., 2010). Again, a forgetting factor (α) is used to approximate the transition matrix to get around this problem.

In general, estimating the coefficients from the state equation (9) 𝜃𝜃_𝑡𝑡^(𝑘𝑘) becomes a problem with these amounts of models, as they involve a number of terms with the size of 𝐾𝐾^𝑡𝑡, for each possible model 𝐿𝐿_𝑡𝑡. This explodes the number of terms, making it impossible to directly evaluate the outcome, therefore requiring the use of an approximation (Raftery et al., 2010). The outcomes that are relevant when using the DMS approach is the outcome of the whole system 𝑦𝑦_𝑡𝑡, where 𝑌𝑌^𝑡𝑡−1 is given. 𝑦𝑦_𝑡𝑡 depends on 𝜃𝜃_𝑡𝑡^(𝑘𝑘), but only when 𝐿𝐿_𝑡𝑡= 𝑘𝑘. In other words, the state vector is only relevant for the outcome of the system for the model that is actually governing at that time. Raftery et al. (2010) therefore advise to use a simple approximation, where 𝜃𝜃_𝑡𝑡^(𝑘𝑘) is updated conditionally on 𝐿𝐿_𝑡𝑡 = 𝑘𝑘 for each sample. The underlying state of the system consists then of the pair (𝜃𝜃_𝑡𝑡, 𝐿𝐿_𝑡𝑡). The quantity of 𝜃𝜃_𝑡𝑡^(𝑘𝑘) is therefore only defined when the model is governing (𝐿𝐿_𝑡𝑡 = 𝑘𝑘), which means the probability distribution of (𝜃𝜃_𝑡𝑡, 𝐿𝐿_𝑡𝑡) can be expressed with the following formula:

𝑝𝑝(𝜃𝜃_𝑡𝑡, 𝐿𝐿_𝑡𝑡) = � 𝑝𝑝(

𝐾𝐾 𝑘𝑘=1

𝜃𝜃_𝑡𝑡^(𝑘𝑘)|𝐿𝐿𝑡𝑡 = 𝑘𝑘)𝑝𝑝(𝐿𝐿_𝑡𝑡 = 𝑘𝑘)

This formula takes the place of the prediction equation (10) in a multi-model setup, and is the formula that gets updated, using a prediction and an updating step, as time goes by and more information becomes available. To explain these steps, we suppose that the conditional distribution of the state at time (t-1), represented as 𝜃𝜃_𝑡𝑡−1,is given, as it would be in the use of this system.

𝑝𝑝(𝜃𝜃𝑡𝑡−1, 𝐿𝐿𝑡𝑡−1|𝑌𝑌^𝑡𝑡−1) = � 𝑝𝑝(

𝐾𝐾 𝑘𝑘=1

𝜃𝜃_𝑡𝑡−1^(𝑘𝑘)|𝐿𝐿𝑡𝑡−1= 𝑘𝑘, 𝑌𝑌^𝑡𝑡−1)𝑝𝑝(𝐿𝐿𝑡𝑡−1= 𝑘𝑘, 𝑌𝑌^𝑡𝑡−1)

In this expression, the conditional distribution of 𝜃𝜃_𝑡𝑡−1^(𝑘𝑘)is approximately normally distributed, expressed as 𝜃𝜃_𝑡𝑡−1^(𝑘𝑘)|𝐿𝐿_𝑡𝑡−1= 𝑘𝑘, 𝑌𝑌^𝑡𝑡−1~𝑁𝑁(𝜃𝜃�_𝑡𝑡−1^(𝑘𝑘), 𝛴𝛴_𝑡𝑡−1^(𝑘𝑘))

The prediction step now also has to predict the model that will be used in the next period, 𝐿𝐿𝑡𝑡, in addition to predicting the value of 𝜃𝜃_𝑡𝑡^(𝑘𝑘) like in the single-model explanation. To that end, the model prediction equation (18) and the parameter prediction equation (19) have to be used (Raftery et al., 2010). Because the parameter prediction equation is conditional on the chosen model 𝐿𝐿𝑡𝑡, the model is predicted first.

This model prediction equation (18) again requires a transition matrix of size 𝐾𝐾 ∗ 𝐾𝐾. In this case, we can avoid using a transition matrix again, by using forgetting factor 𝛼𝛼 to imply the transition matrix (Raftery et al., 2010).

Let 𝜋𝜋𝑡𝑡−1|𝑡𝑡−1,ℓ= 𝑃𝑃[𝐿𝐿_𝑡𝑡−1 = ℓ|𝑌𝑌^𝑡𝑡−1 , then

𝜋𝜋𝑡𝑡|𝑡𝑡−1,𝑘𝑘= 𝜋𝜋𝑡𝑡−1|𝑡𝑡−1,𝑘𝑘𝛼𝛼 + 𝑐𝑐

∑^𝐾𝐾_ℓ=1𝜋𝜋𝑡𝑡−1|𝑡𝑡−1,ℓ𝛼𝛼 + 𝑐𝑐

12 (19)

(20)

(21)

(13) (14) The forgetting factor α therefore governs the weight attached to all model probabilities at time 𝑡𝑡 − 1. Like the other forgetting factor λ, the literature advises to weight this variable slightly lower than 1, typically between 0,95 and 0,99 (Raftery et al., 2010). 𝑐𝑐 is a small positive number which is necessary to avoid mechanical computing problems, and is set to 𝑐𝑐 =^0,001_𝐾𝐾 . This model prediction equation (20) is used to decide which model will be governing in the DMS model.

The parameter prediction equation in the multi model system becomes:

𝜃𝜃_𝑡𝑡^(𝑘𝑘)|𝐿𝐿_𝑡𝑡 = 𝑘𝑘, 𝑌𝑌^𝑡𝑡−1~𝑁𝑁(𝜃𝜃�_𝑡𝑡−1^(𝑘𝑘), 𝐴𝐴_𝑡𝑡^(𝑘𝑘))

where 𝐴𝐴_𝑡𝑡^(𝑘𝑘)= 𝜆𝜆⁻¹𝛴𝛴_𝑡𝑡−1^(𝑘𝑘).

After the prediction steps, like in the single model system, the updating steps are again used to align the model with additional information. The updating equation also has two parts, the model updating equation (20) and the parameter updating equation (21) (Raftery et al., 2010). The model updating equation is expressed as follows:

𝜋𝜋_{𝑡𝑡|𝑡𝑡,𝑘𝑘}= 𝜔𝜔_{𝑡𝑡𝑘𝑘}

∑^𝐾𝐾_ℓ=1𝜔𝜔_𝑡𝑡ℓ

where 𝜔𝜔_𝑡𝑡ℓ=𝜋𝜋_{𝑡𝑡|𝑡𝑡−1,ℓ}𝑓𝑓_ℓ(𝑦𝑦_𝑡𝑡|𝑌𝑌^𝑡𝑡−1), in which 𝑓𝑓_ℓ(𝑦𝑦𝑡𝑡|𝑌𝑌^𝑡𝑡−1) is the density of a 𝑁𝑁(𝑥𝑥_𝑡𝑡^(ℓ)T𝜃𝜃�_𝑡𝑡−1^(ℓ), 𝑉𝑉^(ℓ)+ 𝑥𝑥_𝑡𝑡^{(ℓ)𝑇𝑇}𝐴𝐴_𝑡𝑡^(ℓ)𝑥𝑥_𝑡𝑡^(ℓ) distribution, evaluated at 𝑦𝑦𝑡𝑡.

The parameter updating equation is the last equation we need in order to obtain a working DMS model, and is expressed here:

𝜃𝜃^(𝑘𝑘)|𝐿𝐿_𝑡𝑡 = 𝑘𝑘, 𝑌𝑌^𝑡𝑡~𝑁𝑁(𝜃𝜃_𝑡𝑡^(𝑘𝑘), 𝛴𝛴_𝑡𝑡^(𝑘𝑘))

In this equation,

𝜃𝜃�𝑡𝑡 = 𝜃𝜃�𝑡𝑡−1+ 𝐴𝐴𝑡𝑡𝑥𝑥_𝑡𝑡^𝑇𝑇(𝑉𝑉 + 𝑥𝑥_𝑡𝑡^𝑇𝑇𝐴𝐴𝑡𝑡𝑥𝑥𝑡𝑡)^𝑇𝑇𝑒𝑒𝑡𝑡

𝛴𝛴_𝑡𝑡 = 𝐴𝐴_𝑡𝑡− 𝐴𝐴_𝑡𝑡𝑥𝑥_𝑡𝑡(𝑉𝑉 + 𝑥𝑥_𝑡𝑡^𝑇𝑇𝐴𝐴_𝑡𝑡𝑥𝑥_𝑡𝑡)⁻¹𝑥𝑥_𝑡𝑡^𝑇𝑇𝐴𝐴_𝑡𝑡

which are the same formulas as used to predict the mean and standard deviation of the normal distribution of the parameter updating equation (21) as in the parameter updating equation for the single model system (12).

With these systems of equations, it becomes possible to select the most accurate model with which to generate the returns for each event.

Data Events

The data required for this paper consist of the dates of EU regulatory events of long term policy changes or introductions, as well as Brexit events. This data is sourced using the European Commission’s DG CLIMA official climate news website. This website was created for the purpose of releasing price-sensitive information and

13 create a complete overview of relevant developments in the EU ETS (Koch et al., 2016). Secondly, Thomson Reuters Point Carbon is used to verify the news dates of the events as reported on the DG Clima website, and provide more detailed information about the political process during the policy creation. Point Carbon was also used to assess whether the impact of the events were high enough to include as an event in the study. Consistent with Koch et al. (2016), this paper only regards primary information as published by the regulator. In this case, legislative proposals, votes, communications and votes from the European Commission, Council or Parliament (and its relevant committees) are considered as events in this study. Information stemming from journalists or statements from individual member states are therefore not used. An exemption is made for events concerning the impact of Brexit on the EU ETS, in which case we have also considered statements and publications from the UK Government. The reason for this exemption is because in the Brexit process, announcements from the UK itself regarding their handling of the ETS in Brexit will have a profound impact on the situation and the behaviour of ETS participants.

The events are classified ex-ante and ex-post, based on the expected impact and the actual impact on allowance prices, respectively. In order to classify the results, we utilize the fact that certain events are evidently opposition to either proposed legislation or to (strengthening) the ETS in general. If the event was deemed likely to have a negative impact on the allowance returns, for example by (a proposal for) increasing the cap, the event was classified as negative ex-ante. If the event was likely to increase allowance prices or if its impact was not straightforward to predict, a positive ex-ante classification was given. This ex-ante distinction is made because we do not want policy events which intend to boost the price of the allowance to be conflated with events attempting to prevent price increases, as this would lead to inconclusive results. In this approach, we follow the methodology of Koch et al. (2016). The ex-post classification is based on the actual returns in the event window.

Dependent variable

The data regarding the price of carbon in the market has been gathered using Thomson-Reuters EIKON, which has a comprehensive database of historic daily time-series data for carbon prices in the EU market. In this study, we use data on settlement prices for December 2019 EUA future contracts traded on the ICE ECX. The December futures are the most actively traded contracts in the futures market, as the allowances for the emissions of the previous year have to be surrendered on the 20^th of December. Futures prices are chosen over spot prices

Table 1: Measures of the dependent variable and independent variables used as predictors EU ETS allowance price (EUA) ICE ECX settlement prices December 2019 futures contracts Oil price Brent crude oil price – rolling futures

Coal price API2 index price for coal with an ARA (Rotterdam area) delivery Gas price Natural gas monthly future price at the Title Transfer Facility Electricity price EEX baseload electricity – Phelix baseload month-ahead price CER offset price Official carbon offset credit price

GSCI Non-energy commodity index Goldman Sachs Non-energy Commodity Index STOXX 600 index STOXX 600 European stock index

Interest rate 3-month ECB interest rate

Credit risk spread Default spread between Aaa and Baa rated corp. bonds Volatility index Chicago Board Options Exchange volatility index (fear index) This table shows an overview of the selected explanatory variables and where the data has been gathered.

14 because the majority of transactions in EUAs take place in futures, instead of on the spot market. Secondly, most of the price discovery for EUAs happens in the futures market (Mizrach and Otsubo, 2014).

By choosing December 2019 futures, we deviate from Koch et al. (2016), who used year-ahead EUA December contracts. The reason for this is the availability of data, as the year-ahead prices are no longer available for the complete dataset. Year-ahead contracts are traded in higher volumes than contracts further in the future, which means that the price of the future is less susceptible to individual buyers and sellers. However, a small test sample done showed that the prices of December 2019 and December 2022 futures had a correlation coefficient of 99,90%, while the logarithmic returns of both futures showed a 99,36% correlation. We therefore expect that using December 2019 futures will not alter our findings strongly.

Independent variables

This paper elects to use the ten abatement price predictors used by Koch et al. (2016), which they have shown to perform adequately in the DMS framework. The daily time series for these predictors are retrieved from Thomson Reuters’ EIKON.

The first four independent variables used are the oil price, coal price, gas price and electricity price. The mentioned fossil fuels play a large role in power production in the EU, and are established to be linked to the carbon price (Alberola, Chevallier & Chèze, 2008; Mansanet-Bataller, Pardo & Valor, 2007; Chevallier, 2013).

Fezzi and Bunn (2009) have also proven a passthrough of carbon prices into electricity prices, making these four predictors adequate to use in predicting the carbon price. The connection between the carbon price and these predictors can be found in the fact that energy producers are often able to switch between different fuel sources in electricity production. As these fuel sources have different levels of carbon emissions and energy intensities, differences in price ratio between these three fuel sources will change the demand for carbon allowances, resulting in a change in the carbon price (Lutz, Pigorsch & Rotfuss, 2013).

Secondly, the Certified Emission Reduction certificates (CER) price is used as a predictor, as CER allowances can be used in lieu of EU allowances in EU ETS phase 2, and could be exchanged for EUAs in phase 3. This allows an increase in supply of EU allowances by the exchange of CERs, and gives participants the option to shop between EUAs and CERs in order to find the cheapest way to cover their carbon emissions (Koop & Tole, 2013). While the role of CERs is getting smaller, due to a large decrease in market size and the ineligibility of using CERs of projects after 2012, they could still be exchanged for EUAs throughout the investigated period in this study, and are therefore taken into account.

15 The STOXX 600 index, the GSCI non-energy commodity index, the ECB interest rate, the credit risk spread and the volatility index are used as predictors for macroeconomic risk. Economic slowdown will have a downward effect on allowance prices due to lower aggregate demand, which reduce demand for carbon allowances (Lutz et al., 2013; Chevallier, 2009). The stock index reflects collective expectations of dividends and serve as an expectation of the overall future economic environment (Lutz et al., 2013). The credit risk spread reflects the default risk in credit markets, and serves a good indicator of economic uncertainty. High economic uncertainty is expected to have a downward effect on prices (Lutz et al., 2013). Lastly, the commodity index captures the risk found in fluctuations of global commodity markets. Lower commodity prices foreshadow a downturn in economic activity due to lower aggregate demand, which puts negative pressure on allowance prices (Lutz et al., 2013). The same applies to the ECB interest rate, which reflects European monetary policy, and tends to decrease with economic downturn (Chevalier, 2009). The volatility index reflects risk perception in the financial markets, and an increased risk perception seems to increase the link between stock prices and EU allowances (Koch, 2014).

Table 2: Stationarity of variables

Original data Logarithmic difference ADF statistic KPSS statistic ADF statistic KPSS

statistic

EUA Price -0,096 4,082 *** -12,90 *** 0,601

Oil price -1,611 15,79 *** -12,16 *** 0,149

Coal price -1,732 7,602 *** -11,28 *** 0,245

Gas price -2,278 8,973 *** -11,40 *** 0,071

Electricity price -3,025 4,915 *** -12,71 *** 0,049

ECB Bonds -1,949 18,18 *** -13,87 *** 0,059

CER offset price -2,206 10,74 *** -14,32 *** 0,149

GSCI non-energy index -2,401 18,33 *** -12,47 *** 0,080

Stoxx index -2,459 17,70 *** -13,90 *** 0,043

Volatility index -4,555 *** 4,039 *** -15,71 *** 0,009

Credit spread -1,979 2,440 *** -11,12 *** 0,110

In this table the results of the Augmented Dickey Fuller test and the Kwiatkowski–Phillips–Schmidt–Shin test for stationarity of the used variables are reported. Because these tests show that the original price and index data is not stationary, the variables are converted by taking the log difference, transforming them into log return data. ***, ** and * indicate significance at the 1%, 5% and 10% level, respectively.

Table 3: Descriptive statistics

Mean^a Std. Deviation^a Skewness Kurtosis

Allowance returns 0,04 0,59 -1,22 15,3

Oil returns -0,04 0,37 -0,09 3,26

Coal returns -0,13 0,23 -0,61 55,8

Gas returns -0,08 0,36 0,40 6,55

Electricity returns -0,05 0,57 2,92 42,0

ECB Bond returns -0,22 0,78 0,57 17,9

CER offset returns -0,51 1,44 18,7 634

GSCI non-energy index returns -0,07 0,14 -0,18 2,07

Stoxx index returns 0,06 0,20 -0,51 3,86

Volatility index returns -0,05 1,51 1,09 7,64

Credit spread returns -0,03 0,36 -0,30 5,26

This table shows the descriptive statistics of the stationarized dependent and independent variables

a Means and standard deviations are annualized

16 While DMS and DMA do not necessarily require stationary time series in order to work correctly, having stationary variables can help the variance updating equation (Koop and Koribilis, 2012). However, the alternative forecasting methods, such as recursive OLS, require stationary time series in order to make reliable predictions (Stock & Watson, 2015). All variables are therefore tested for stationarity using the Augmented Dickey Fuller test, and the Kowalski-Phillips-Schmidt-Shin test. The results of these tests can be found in Table 2. The significant results of the KPSS test and the absence of significance in the ADF test show that the original data was not stationary. After transforming the data into compounded returns by taking the natural logarithm of the first difference of the variable values, the ADF tests become significant, while the SPSS tests are no longer significant, showing that the variables are stationary after the transformation.

The descriptive statistics of the transformed data can be found in Table 3. Notable in the descriptive statistics are the mean and standard deviation of the Allowance price. While the standard deviation of the data is almost equal to the findings of Koch et al. (2016), the annualized mean return has become positive, mainly due to the resurgence of allowance prices between 2014 and 2019. The CER offset returns show a very low mean, as the price of the certified offsets has declined from over €12 in 2011 to just €0,22 in 2019, as these offsets have lost a lot of their value due to an oversupply of allowances in the ETS, and cannot be used in phase 4 of the ETS starting in 2021.

Table 4: Studied events

Date Studied events Ex ante

class. Event code

Backloading events

20-12-2011 Committee passes amendment of EE Directive including withdrawal of 1.4 bil. EUAs + A01 28-2-2012 Committee adopts amendment of EE Directive and calls for report on set-aside plan + A02

14-6-2012 EE Directive is agreed but inclusion of set-aside does not pass - A03

25-7-2012 EC plan to backloading Three variants for consultation: 400, 900 or 1200 mil. EUA; change of ETS

Directive + A04

13-11-2012 Official proposal to postpone 900 mil. EUAs; further reform measures announced + A05

24-1-2013 EP ITRE Committee votes against the proposal - A06

19-2-2013 EP ENVI for proposal Committee votes in favour of proposal but delays the decision to start fast-track

legislation + A07

25-2-2013 EP ENVI no fasttrack, committee fails to agree to start fast-track legislation talks - A08 16-4-2013 EP negative vote, rejection of backloading proposal in plenary vote - A09 19-6-2013 EP ENVI amended backloading proposal with stricter conditions (one-time-measure) + A10 3-7-2013 EP positive vote Second plenary vote in favour of amended backloading proposal (311 against vs 344

in favour) + A11

17-12-2013 EP + Council compromise Final compromise between Parliament and Council is reached + A12 30-1-2014 EP ENVI fast tracking Committee agrees to fast track backloading legislation + A13 24-2-2014 Council agrees adoption Council adopts fast-track backloading officially + A14

Market stability reserve events

22-1-2014 Commission launches second step to reform the EU ETS setting 40% target by 2030, LRF to 2.2% + B01

24-10-2014 Council secures deal on 2030 climate and energy framework + B02

22-1-2015 European Parliament ITRE committee fails to agree on MSR and unallocated allowances - B03

24-2-2015 EP's ENVI committee votes to start MSR by the end of 2018 + B04

5-5-2015 Parliament and council confirm Market Stability Reserve agreement + B05

13-5-2015 Council votes positively on MSR trilogue result + B06

18-9-2015 Council officially rubberstamps market stability reserve plan + B07

15-2-2017 EU Parliament votes on increasing LRF and strengthening MSR in ETS phase 4 + B08