![Sample our Engineering & Technology journals, sign in here to start your access, latest two full volumes FREE to you for 14 days]({"type":"image" , "src" : "/sda/5933/TOC-Subject-Sample-2021-Engineering-Tech.jpg"})
Abstract
We study the informational efficiency of the European Emissions Trading Scheme, EU ETS market, by simulating the trading in this emerging market. If the market is efficient, profitable trading should only exist locally in time. We adopt the Timmermann and Granger (2004) definition of efficiency and run a large set of econometric, technical analysis and combined models to forecast the emissions allowance price changes. These forecasts are then used as trading signals in the trading simulation. We find that the combined models outperform the other models in forecasting ability. Trading simulation based on models combining time series and technical analysis trading rules shows that there have been possibilities for profitable trading in the EU ETS market during the study period of 2008–2010. This suggests that the EU ETS market shows periods with no informational efficiency.
Acknowledgements
We acknowledge funding support from TEKES (Climus/POMAR) and Academy of Finland (BEET, project no. 124480). The earlier versions of the paper have been helpfully commented by the participants of the EAERE2009 (Amsterdam) conference and HECER seminars (Helsinki) and participants in the HEC Energy & Finance Chair workshop ‘The Behavior of Carbon Prices’ (Paris, January 2012). The authors thank also Andreas Löschel, Adriaan Perrels, Janne Tukiainen and two anonymous referees for useful comments.
Notes
1. Extension of the information set, Xt, to cover also other aspects than the asset's own price history is always conditional on the researcher's choices and thus there is a possibility for bias: What information should be included? Which variables should be chosen for the fundamental analysis? In the asset price modelling, the information set is often extended by macro- or micro-level variables like dividend yield, T-bills and T-bonds as well as growth and inflation rates (Pesaran and Timmermann Citation1995).
2. The first phase of the EU ETS was 2005–2007, the second 2008–2012 and the third 2013–2020. The first phase was considered as a ‘learning-by-doing’ period and the initial allocation of the EUAs was generous. The first publication of the emissions data in May 2006, an important event for the EU ETS market, showed how quickly information is absorbed by the market. As the information on the great surplus of allowances was revealed, the market reacted in a couple of days. In a week the price crashed from almost 30 euros close to 5 euros and after a month the EUA2007 had almost no value at all.
3. EU ETS market includes trading for both compliance and speculation and the trading behaviour of the two types of agents may differ, as Maeda Citation(2004) discusses. In both cases we think, the advanced firms base their decisions at least on some systematic procedures focusing on the market price movements. For the analysis in the paper we assume, therefore, that all the traders, regardless of their background, act similarly in the market based on the price signal that the models give. During our study period, until year 2010, most of trade was conducted in forward markets, so in that sense both compliant and non-compliant traders acted consistently.
4. The API 2 price is the primary price reference for physical and OTC coal contracts in northwest Europe. Some 90% of the world's derivatives are priced against the Argus/McCloskey API2 and API 4 indexes. CIF ARA means coal delivered to Amsterdam, Rotterdam and Antwerp inclusive of costs, freight and insurance.
5. Particularly, we use Wednesday observations to avoid the possible weekday anomalies. There is evidence of returns being abnormal on Mondays and Fridays in the stock markets (e.g. Gibbonsand Hess Citation1981; Cross Citation1973).
6. The level series of the Clean spark spread includes some negative values and we have taken only the first difference. Furthermore, in the trading simulation, volume data are treated in a way that only observations of the growing volumes were taken into the estimation. In addition, the positive log returns of the volume series are multiplied by minus 1 if the EUA return is negative during the same week. By this we want to control the increased volatility of the EUA forward price approaching its maturity date.
7. This Timmermann and Granger approach to study market efficiency has been applied in many settings. See e.g. Sanders, D R and Manfredo M R 2005, Hsu, Po-H and Chung-Ming, K. 2005.
8. Note that we do not have to assume anything about firms’ risk preferences. However, we assume that new information accumulates identically to all products of the EUA, i.e. spot and forward contracts. When forward markets are present, Holthausen Citation(1979) shows that firm makes its production decisions based on the forward price, irrespective of the degree of its risk aversion. The risk preferences affect only the hedging decision.
9. See, for example, Box and Jenkins Citation(1976), Hamilton Citation(1994) and Lütkepohl Citation(2007) for the state of art forecasting models.
10. With all regressors included in the model the regression equation includes, with the constant term, 28 regressors and the whole model set includes 262,144 different model combinations.
11. Hence, using a vector characterisation the used mean equations of the GARCH models can be defined by vectors AR(1) = (1,0,…,0) and FU(1) = (1,…,1). To save space we abbreviate the models as AR- and FU-.
12. See, e.g., Lanne and Luoto Citation(2008) and Arshanapalli, Fabozzi, and Nelson Citation(2011).
13. See Nelson Citation(1991) and Pierre Citation(1998) for detailed discussion. Further, we run models that combine the GARCH-M and EGARCH model specifications. These models are noted with EGARCH-M.
14. Daskalakis and Markellos Citation(2008) found evidence of profitable first and second phase futures trading. They also checked the profitability of another trading rule, namely trading range break-out (TRB). With the rule of TRB(1,30), a buy (sell) signal is generated if the current price is lower (greater) than the minimum price during the 30-week window. In our simulation, TRB rules turned out to be really unprofitable and the results are not reported.
15. To ease the computational burden, we, in fact, use a two-stage process in the model selection. In every eight weeks, we choose best 500 models with respect to the BIC value and fix the model set for these models for the following eight-week period.
16. Pseudo out-of-sample forecasting is used to evaluate the forecasting power of the model. We use recursive and rolling methods. Recursive RMSFE is calculated by first reducing the original sample of observations by 15% and forecasting the omitted values with this shortened model. Rolling RMSFE is calculated in the same manner as in the recursive method except that the number of observations included in the regressions is kept constant. In our models, the forecasting window is 30 weeks.
17. We use the Marquardt maximising algorithm and the normal distribution. The models assume normal distribution of error terms and the back casting parameter is set at 0.7.
18. This is a relatively low-risk premium. Nevertheless, due to the economic downturn the reference for the risk-free rate, the yield of 3-month Germany Treasury bill for example, has been on average during the estimation period 1.5% but with a huge variance from 3.2% to 0.25%. This is also a relatively low rate. Besides, we use the 5% proxy as an upper bound for the risk-free rate when calculating the Sharpe ratio. Thus, the values of the Sharpe ratio are biased downwards.
19. Some of the worse GARCH models are not in line with this result. We also have a significantly smaller sample than Fang and Xu Citation(2003).
20. Recursive RMSFE is only marginally profitable with the 0.6% filter.
21. In the trading simulation, volume data are treated in a way that only observations of the growing volumes were taken into the estimation. By this we want to control the increased volatility of the EUA forward price approaching the maturity date. In addition, the positive differences of log volume series are multiplied with minus 1 if the EUA return is negative during the same week.