Sentiment Analysis as a Source of Gaining Competitive Advantage on the Electricity Markets

Wioletta Soko?owska, Tymoteusz Hossa, Karol Fabisz, Witold Abramowicz, and Mateusz Kubaczyk

Sentiment Analysis as a Source of Gaining Competitive Advantage on the Electricity Markets

Wioletta Soko?owska, Tymoteusz Hossa, Karol Fabisz, Witold Abramowicz, and Mateusz Kubaczyk

—The electricity retail markets are evolving toward more competitive and customer-oriented. The deployment of smart meters and a wealth of new technologies create customers’ eagerness for taking control of their electricity consumption. By being better-informed about the energy usage, people are encouraged to switch deals among existing suppliers or move to a new energy provider. Moreover, as customers are more socially interconnected, the Internet portals and social media become a place for discussion, comparison, and evaluation of the available offers. Unfortunately, in case of the energy sector there is a lack of understanding that such information, when taken into account and properly analyzed, can be a completely new and a powerful source of competitive advantage.

In the paper, we introduce a solution that the use of quasi real-time automated sentiment analysis on the energy suppliers and the relevant aspects of their offers may enable energy companies to adapt quickly to changing circumstances, prevent potential customer churn, and harness new business opportunities.

Index Terms—Customer attrition, electricity retail market, sentiment analysis.

1. Introduction and Motivation

The currently observed deployment of smart grid infrastructures accompanied by a wealth of new technologies, together with an increasing utilization of renewable energy sources, an ongoing liberalization and emergence of new participants (e.g., prosumers and new competitors) brings major changes to the electricity retail markets[1],[2]. Finally, customers are not only the subjects of energy provider’s activities but they also start to be recognized as the first-class participants influencing various processes. Moreover, an access to different tools and information sources enables them to reduce their energy consumption and also take control of their own energy supply needs[3]. As people become better-informed about the energy usage, they start to exercise their right to switch deals among existing suppliers or move to a new energy provider[1],[4].

According to recommendations proposed by the Council of European Energy Regulators (CEER) followed by the European Union (EU) directives, the possibility of switching should be easy and free, and should be carried out within three weeks of informing the current supplier[5],[6]. Moreover, CEER recognizes supplier switching as one of the most relevant processes on the retail markets, mainly due to the fact that if it functions well it might result in building customer trust and engagement[4]. However, an effective energy retail markets[7]can only be developed if customers’ satisfaction regarding energy supplier’s offer, service quality, and price, etc. is really being taken into account. For this end, energy providers must realize that making rational decisions and reacting to the changes occurring in their business environment are only possible when various information sources (e.g., new regulations, user feedback, and competitors moves, etc.) are being monitored and analyzed[8]. At this point however, the customers’ opinions expressed on the Internet portals and social media seem to be out of the scope of the suppliers’interest. Unfortunately, even though the volume of available data is rapidly increasing, in case of the energy sector we think that there is a lack of understanding that such information can be a completely new and a powerful source of competitive advantage.

In the paper we present the results of our early stage research aiming at developing a solution that the use of quasi real-time automated sentiment analysis on the energy suppliers and the relevant aspects of their offers may enable energy companies to adapt quickly to changing circumstances, prevent potential customer churn, and harness new business opportunities. The paper is structuredas follows: Section 2 presents the statistics of supplier switching in the retail electricity markets in EU. Moreover, the determinants for brand switching are identified and thorough analysis of the market situation is conducted. In Section 3 the related work in sentiment analysis is shortly discussed in order to position the paper towards the work of others in this field. Section 4 presents the proposed scenario, providing the description of the developed method and the resources used. Moreover, the evaluation of the obtained experimental results is also presented. Section 5 concludes the paper with final remarks and outlook on the future work.

2. Supplier Switching in the Retail Electricity Markets

The liberalization and integration of the EU energy retail markets are still far from completed. There is a number of indicators responsible for jeopardizing market openness and competition, such as attachment to regulated prices, low consumers’ awareness, high level of customer protection, lack of Internet access, not yet developed or poor business, and switching models[9]. The latter is distinguished between a bilateral and a centralized model. In the first one, the winning retailer interacts with the payer’s Distribution System Operator (DSO)—one or many. In the centralized switch, the winning party is a central entity that processes the switch and interacts with all other market agents, omitting the customers. The logic of those models is not a main case of this paper, however according to [10], 11 (44%) out of the surveyed jurisdictions (not only from Europe) are using the centralized model, as it is perceived as a faster for the customer and has a general positive impact on switching rate (bilateral model with single DSO—4% of the measured impact on switching, bilateral model with multiple DSO—7%, centralized model—11%)[11].

In 2014, the average switching rate in the EU countries amounted to 6%, both for electricity and gas. Therefore, it might seem that customers were not yet really interested in changing their energy suppliers. In fact, according to the research of the European Consumer Organisation (BEUC), 41% of household energy consumers did not even know if there was any opportunity to find a cheaper deal. The study showed also that less than 50% of the consumers were even aware how much electricity they consumed. On the other hand, over 72% of individual clients were dissatisfied with the level of complaints and incidents handling by the energy suppliers but only less than one third of payers had ever compared deals from different providers[12]. If we consider the energy consumer being in fact an organization, we learn that for the same year (2014) average 46% of companies looked into their suppliers’ offer and started comparing various tariff options. However, a positive incentive on taking action was noted by companies with higher levels of energy spend (￡10000 plus). 57% of them took action, and 42% looked for other suppliers and tariffs[13].

To achieve the integration of the European markets leading to EU, more benefits to energy consumers as well as stronger regulatory incentives for retail market design, information, switching, and billing must be delivered and adopted, such as in [4] and [14]:

· Promotion of easy to use, understandable comparison, and switching tools that will allow for reducing the conviction about that switching.

· Products for optimal customer service and engagement.

· Deletion of regulated prices policies.

· Harmonization of regulatory regimes and licenses.

· Harmonization of billing process for the needs of better payment’s handling.

· Enforcement of services unbundling.

Following the outcomes of the Agency for the Cooperation of Energy Regulators’ (ACER) 2014 market monitoring report, switching rates and the number of active suppliers might be considered as factors that differentiate European countries. There have been already few countries like Finland, Sweden, Norway, Netherlands, the Great Britain, and Ireland that have adopted efficient and effective switching procedures. Unfortunately, most of the EU countries still need to develop and deploy appropriate regulations in this manner. In general, countries with regulated energy retail price have both a lower number of active suppliers and a lower switching rate[15]. In 2013, the active switching levels of the electricity retail markets in Finland, Sweden, Norway, Netherlands, the Great Britain, and Ireland amounted from 5% to 18%. At the same time, Austria, France, Denmark, Greece, Hungary, Northern Ireland, Poland, Bulgaria, Cyprus, Estonia, Latvia, Lithuania, Luxembourg, and Romania recorded low or very low switching rates[16].

Another issue regarding the ongoing changes of the electricity retail markets is an observation presented in the Consumer Focus’ survey on the consumer experiences of energy switching. The research showed that the popular assumption stating that once consumers change their energy supplier and make savings they become repeat switchers not always holds true. Thus, the EU policies should put a lot of emphasis on a long-term consumers’ engagement as well as on optimization of surveying processes[17]. In this paper we focus on UK countries, especially the Great Britain, as they are considered the best examples of how to adjust electricity retail markets to nowadays expectations and the EU regulations. Additionally, it is possible to collect and analyze data (comments on energy suppliers) from the Internet fora and surveying services regarding UK energy providers and their offers that work as a basis for our experiment (see Section 3).

The year of 2013 was a year of significant changes and reforms in the energy UK retail market (e.g., the set of reforms called Retail Market Review was completed and two schemas for retail: Green Deal and Energy Companies Obligation were defined). The retail energy market was opened for customers[15]allowing for the switching process to start.

Currently, the overall energy consumption level in the UK amounts to around 294 TWh per year. There are over 27.3 m households, and over 2.8333 m business entities that are clients of the UK energy suppliers[18]. On the energy market, there are 27 licensed companies, 5 of which sell either electric energy and 22 of which sell both electric energy and gas[18]. The examples of major suppliers in the households market are British Gas, E.ON UK, EDF Energy, RWE npower, Scottish Power, and SSE. Companies like Ecotricity, Good Energy, Co-op Energy, Economy Energy, and Spark Energy represent medium and small energy providers. According to [18], nowadays an erosion of the major energy suppliers’ shares at the households market could be observed. It is worth underlying that incumbent companies (with shares over 10% each) have recently experienced a fall in the aggregate market share. Thus, having a big market share does not always imply that clients are satisfied with the current level of service[18]. In 2015, over 9400 customers took part in the satisfaction survey where they were asked to evaluate their experience (using stars from 1 to 5) with their gas and electricity suppliers[19]. The survey contained questions on: Overall satisfaction of customer service, value for money, bills accuracy and clarity, complaints handling, supplier’s efforts to encourage consumer to be energy efficient, and likelihood to recommend the energy supplier to a friend.

The results showed that minor companies were actually winning the competition. The best one turned to be a small energy supplier—Ecotricity with the 84% of the overall satisfaction score. Its followers, also small companies, namely Good Energy and Ebico received respectively 82% and 81%. The first from the big six—Eon, ranked at the 12th place with only 50% of the overall satisfaction score. Npower, another big company, was scored as the worst energy supplier of all the twenty-two that were considered in the survey, achieving only 35%[19].

Another research in [20] confirmed the increasing interest in smaller suppliers. The monthly statistics published by the Energy UK showed that in April 2015 the number of switching customers reached a total of 335826. In comparison to March 2015, there was a 24% decline observed and over 15.93% (46000) increase in relation to the previous year. The 25.15% (84471) of switching consumers turned to small suppliers[20]. The biggest number of changing of supply event (COS Events—an industry term) was observed in November 2013, when 615363 clients moved from one supplier to another due to the fact that there was an announcement of a series of electricity price rises[20],[21].

What is more, according to the outcomes of the Central Research Institute of Electric Power Industry (CRIEPI) project, residential switching rates are also boosting because of the growth in price differences between energy suppliers, sudden price increases, seasonality, negative publicity (e.g., on the Web) and the emergence and expansion of services or third parties promoting switching deals, comparing tariffs and energy suppliers[22]. Moreover, one can see the robust of commercial and non-commercial weekly and monthly surveys and industry reports on payers’ opinions of their energy suppliers. These reports investigate such metrics as ease of switching, customer effort and trust score, or net promoter score (the likelihood that customers will recommend their energy supplier)[23].

In the UK, the switching process is being supported by many governmental and non-ministerial agencies. One example can be Ofgem, an authority (respected by the EU) that developed the Confidence Code which is being considered as a code of practice that governs independent energy price and suppliers comparison services. Up to now, there are twelve websites accredited by Ofgem available in the UK that help individual households and companies compare energy suppliers’ offers[24]. Moreover, there are already several regulations and information campaigns established speeding the process of adjusting the energy market to a new reality[15],[25], e.g.,

· “Faster switching” to speed up the time of change.

· “Tariff choices” where one energy supplier can offer maximum four tariffs.

· “Lowest price” where suppliers are obliged to inform the customer which tariff is cheapest for him and how much he could save.

· “Going when you owe” where person who uses prepayment meter can switch deal even if he owes money to the existing supplier.

· “Exit fees” where fees are mandatory only in case when somebody cancels a fixed contract.

· “Switching numbers” which is a platform where monthly statistics on switching deals are published.

· “Help with switching” where the step by step guide and animation are published on the website.

It must be stated that only aware and more educated consumers are actually willing to change their sticky habits and move to other supplier. The elders least educated and persons without Internet access often choose monopolists’offer which is considered as a safer one[9].

The existing traditional switching sources, like 51% of door-step, 11% of telesales, 4% of collective switching, and 3% of mail, might be quite expensive and time consuming for the energy suppliers. On the other hand, for the customers, they might be considered as a financiallyeffective way to save clients’ money. For example, in 2013 in Belgium, 32995 consumers that switched using collective switching saved in total €6.8 m, whereas in Austria approximately 70000 consumers saved over €12.6 m from switching. At the same time in France over 71000 customers saved together over €13.7 m. In the Netherlands, 60547 payers used cumulative switching to gain over €16 m of total savings[26]. It must be emphasized that traditional switching sources result in 69% of total switches. The remaining number is generated by other means. In fact, it is estimated that from 1% to even 75% of nowadays electricity switches take place via Internet. Countries like Belgium and the Great Britain (approximately 45% of cases) are the leaders, whereas Italy oscillates around 5%[11]. Thus, the energy companies should start harnessing business opportunities connected with the Internet use. Especially, there have been already many web services that not only report statistics of traditionally conducted surveys of customer satisfaction, but also become a place where people voice their opinions on various aspects of energy supplier offers. In general, more consumers turn to Internet to seek quality information from online reviews, while many firms use them as an important resource in their product development, marketing, and consumer engagement[27]. However, taking into account the sheer volume and the noisiness of the free-style content available on the Web in order to automate the content-assessment process and to obtain the relevant information an application of proper sentiment analysis techniques is recommended[28].

3. Sentiment Analysis-Related Work

The sentiment analysis allows for examining an opinion, sentiment and subjectivity in a text[29]. It employs computational techniques, such as natural language processing, data mining, and text mining in order to review and decide whether each text’s emotional attitude is positive, negative, or neutral. Usually the sentiment analysis experiment is split into three parts: Part-of-speech tagging, subjectivity tagging, and polarity detection[28],[30]. In general, there are two main types of methods for sentiment analysis: Lexical-based and machine-learningbased. The later approach relies on supervised classification allowing for binary sentiment detection (i.e., positive or negative). This approach requires labeled data to train classifiers[31]. The methods that fall into this group have an ability to adapt and create trained models for specific purposes and contexts. Even though they are considered as fairy effective, their main disadvantage is that there is a low applicability on new data. On the other hand, the lexical-based methods make use of a predefined list of words, where each word is associated with a specific sentiment. Thus, the lexical methods vary according to the context in which they are created[28],[30],[32],[33].

Up to now, a lot of research has been done on the sentiment analysis applied in different domains. However, many experiments and real-life application examples focused their attention on two aspects. The first one is connected with applying sentiment analysis methods to analyze data coming from popular social media portals like Twitter or Facebook. The second one concerns all research activities that were undertaken to develop better technical features like performance, accuracy, precision, and recall of the sentiment analysis itself[28]. According to [34], different machine learning classifiers like Na?ve Bayes, support vector machine, maximum entropy, winnow classifier, decision trees, k-nearest neighbor, and Adaboost have been used to examine data sources like movie and book reviews, product reviews, customers’ feedback, and restaurant reviews (in general consisting of more characters than a traditional tweet).

For many marketing or public relations companies, mastering sentiment discovery techniques can be very useful in the current situation of data and information robust. Popular social media websites offer dedicated application programming interfaces (APIs) for those interested in data extraction. In the recent years, there was a robust of applications and websites handling the automated summaries using Twitter data. Some successful examples of automatic sentiment analysis portals or applications include tweetfell.com, engagor.com, umigon.com, thestocksonar.com, and sentiment140.com.

It has become almost a standard to use the sentiment analysis of social media data for election campaigns, financial markets, and big events like Super Bowl. In fact, nowadays every action, speech, acquisition, and trade can be analyzed in terms of its popularity and sentiment[35]-[37]. However, even though the sentiment analysis methods found their way into a wide range of applications for, inter alia, topics related to sociology, politics, tracking companies reputations and competitors, discovering customers opinions about movies or products or services, and monitoring positive or negative trends in the social media[29],[31],[38], and so on. It is impossible to find a single research that would address the topic of the sentiment analysis for the needs of benchmarking the energy market in the context of switching energy deal.

4. Application of Sentiment Analysis on the Electricity Market-Experiment

4.1 Data and Technology Used

To conduct our experiment, three different sets of comments on the energy operators were used. Thecomments were extracted directly from three UK Internet fora where people discuss and rate energy suppliers and their offers. In total, we were able to automatically retrieve over 10900 comments from February 2012 to September 2014 concerning 17 British energy suppliers. The average comment consisted of 539 characters, being in fact more challenging to analyze than an average Twitter comment. The sample comments are presented in Fig. 1.

The experiment was conducted using SAP HANA provided by the Hasso Plattner Institute Future SOC Lab in Potsdam. The instance was installed on the machine equipped with 24 processor cores and 64 GB RAM. The code for automatically extracting data from the identified websites and loading it into SAP HANA table was written in Python language and the sentiment analysis was performed with the use of the SAP HANA Text Analysis module. The method proposed in this paper was implemented in SAP HANA using SQL Script. In this way, the overall computing time (excluding the time needed for data retrieval) was shortened to 12 s.

Fig. 1 Opinions of energy providers’ clients—examples. Source: www.moneysupermarket.com and www.reviewcentre.com.

4.2 Proposed Method

The overall experiment was divided into two parts. In the first part, we were mainly testing the SAP HANA Text Analysis module to gain knowledge on how the default settings work and to develop a suitable sentiment analysis method. Moreover, we wanted to access how well a commercial software can perform if it would be applied for the same purpose. For this end, we used a smaller dataset (initially around 900 comments). The goal of the second part was to test our method on a bigger dataset (initially over 10000 comments) and compare the obtained results for the bigger and smaller dataset.

The research consisted of the following steps: Data extraction, data cleanup, sentiment identification, and evaluation.

As it was already mentioned, the data extraction was performed using Python script written from scratch for this purpose. This script was embedded in the SAP HANA environment to retrieve data and load it directly to the SAP HANA database. Thus it was possible to fasten the whole process.

Within data pre-processing (data cleanup), the comments with less than 10 characters were removed. As a result of few initial tests, we then decided to remove also the comments where our method was unable to find neither positive, negative, or neutral entity (this removal was executed during the next step). The aforementioned situation could occur in a comment written with words that were not included in the default dictionary provided by the SAP HANA Text Analysis tool (e.g., slang or misspelling). Finally, in the first part of the experiment, 656 comments out of 893 extracted became a subject to further analysis and were employed to develop the method. Respectively, in the second part 9583 comments out of 10 012 extracted were used to test and adjust the proposed solution. It is worth to mention that although the obtained reviews referred only to 17 UK utilities, there were many possible abbreviations and variants of the same company’s name in the corpus, i.e., for British Gas there were 5 different variants identified.

The identification of the sentiment value was performed with the usage of the SAP HANA Text Analysis component, in particular the EXTRACTION CORE VOICE OF CUSTOMER configuration with the default lexicon settings. Thus, the proposed method was built using the lexicon-based approach for the sentiment analysis. A result of this procedure was a new table with comments divided into text entities together with the assigned sentiment. Only entities with one of the following sentiments were taken into further consideration: Weak positive sentiment, weak negative sentiment, major problem, minor problem, strong positive sentiment, and strong negative sentiment. It should be stated that during this step, after the sentiment tokens were distinguished, we decided to remove also those comments that lacked of neither positive, negative, or neutral entity (no sentiment assigned) and the method was tested again. Consequently, a few values were then calculated for each comment: POS is a sum of a number of weak positive sentiment and strong positive sentiment entities, NEG is a sum of a number of weak negative sentiment, major problem, minor problem, and strong negative sentiment entities, POSNEG is a sum of POS and NEG. Next the Measure value was calculated as a ratio of POS to POSNEG, like (1). Within the following step, the Polarity value was assigned, depending on the level of Measure value, as in (2). Polarity describes the calculated sentiment value for each comment.

At this phase of research, some basic statistics on the sentiment for each tested company were calculated. For instance, the energy sector business analyst can learn that the company EDF Energy Electricity’s positive opinion share is about 58% percent. At the same time, this company has slightly less than 13% of negative reviews share. The conclusion for this stage was that larger companies usually have more negative comments than the smaller energy suppliers. Thus, our experiment confirmed the results of the surveys presented in Section 2.

The expert-driven feedback phase was used to evaluate the proposed method. For this end, the gold standard approach was used. The 3 experts willing to participate in this evaluation were provided with a representative sample of comments for each of the analyzed datasets. For the smaller one it was 242 comments and 370 randomly chosen comments for the bigger one. The experts’ task was to state (separately) whether each comment has a positive, a neutral, or a negative sentiment. When two or more persons agreed on a chosen sentiment (negative, neutral, or positive), this sentiment was assigned to the comment. If there was an issue, namely experts did not agree on the sentiment, a fourth person was asked to decide which sentiment should be assigned for a particular comment.

To compare the sentiment assigned by humans and the machine, three resulting values were calculated for the neutral, positive, and negative sentiment: Precision, Recall, and F-measure. Precision is a percent of the correctly assigned sentiment by the machine to sum of the correctly assigned sentiment and wrongly assigned sentiment separately for each sentiment polarity. Recall is a percent of the correctly assigned sentiment penalized by a number of missed items in each sentiment classification. F-measure is an effectiveness measurement. It is measured as a weighted average of the Precision and Recall. The best value for F-measure is 1 and the worst is 0. These values as an example for positive polarity were calculated as follows (3), (4), and (5).

4.3 Evaluation of the Experimental Results

The gold standard evaluation results, precisely values of the Precision, Recall, and F-measure for each sentiment polarity are discussed in this section. Table 1 presents the results that our method obtained for the smaller dataset. According to F-measure, the developed method is quite effective in assigning positive sentiment correctly (about 80%), but it has lower effectiveness in assigning negative sentiment (about 59%). More unfortunately, the method does not give the satisfying results for neutral polarity classification. Probably, this partially might be caused by the customer attitude to the company, i.e., when somebody writes an opinion about the energy supplier usually has a clearly defined positive or negative attitude. Very rarely someone writes a neutral comment about his energy supplier. The test performed on the smaller sample showed that a default configuration in the SAP HANA Text Analysis (the default SAP HANA set of lexicons) can provide quite satisfactory results.

Table 1: Gold standard values of Precision, Recall, and F-measure parameters for positive, neutral, and negative comments for a smaller comments set.

Table 2: Gold standard values of Precision, Recall, and F-measure parameters for positive, neutral, and negative comments for bigger comments set.

In the Table 2, the results of gold standard values of the Precision, Recall, and F-measure for the bigger dataset are presented. It is easy to notice that for the second set of the Precision of positive polarity classification is at very good level of 91%. But effectiveness of positive polarity classification stays on a very similar level like in the previous test (82%). It is worth to notice that in this dataset the negative polarity has a very good classification effectiveness (83%), with lower Precision but higher Recall. The classification of neutral polarity stays on a very low level, amounting for 19%.

5. Final Remarks and Future Work

The energy sector is a very sophisticated domain with many of terms that are not supported by the default dictionaries in the SAP HANA Text Analysis or probably in any other commercial tool. Moreover, the clients while expressing their opinions about the quality of the servicesand their suppliers in general often came up with a domain specific slang. Nevertheless, our research showed that it is possible to conduct the sentiment analysis research in this area and obtain a quite satisfactory result even with the default lexicon configuration.

It is worth to notice that for our best knowledge this sentiment analysis research based on customer opinions about energy operators is the only one that was conducted within this specific domain and its results are available to the public. Even when we expand this domain not only to the energy operators but also to the related topics like oil and gas providers, we were able to find only one research conducted by the Industry Sentiment Intelligence about impact of controversial movie on gas and oil drillers business[39].

In our future work, we would like to better the results by replacing the default configuration of the SAP HANA Text Analysis with the domain specific lexicon and by incorporating in our solution not only a lexicon-based approach but also support it with machine learning techniques.

[1] R. Hoffman, K. Fabisz, A. Filipowska, and T. Hossa,“Profiling of prosumers for the needs of energy demand estimation in microgrids,” in Proc. of the 5th Intl. Renew. Energy Congress, 2014.

[2] R. Hoffman, A. Filipowska, K. Fabisz, T. Hossa, and M. Mucha, “Towards forecasting demand and production of electric energy in smart grids,” in Proc. of the 12th Intl. Conf. on Perspective Business Informatics Research, 2013, pp. 298-314.

[3] Utility of the Future: A Customer-Led Shift in the Electricity Sector, an Australian Context, PricewaterhouseCoopers, Apr. 2014.

[4] Retail Market Design, with a Focus on Supplier Switching and Billing Guidelines of Good Practice, Council of European Energy Regulators, Jan. 2012.

[5] European Energy Consumers’ Rights, European Commission, 2014.

[6] Retail Market Design, with a Focus on Supplier Switching and Billing Draft Guidelines of Good Practice, Council of European Energy Regulators, Jul. 2011.

[7] The Role of Distribution System Operators (DSOs) as Information Hubs, Eurelectric, Jun. 2010.

[8] K. Haniewicz, M. Kaczmarek, M. Adamczyk, and W. Rutkowski, “A case study of sentiment orientation identification for polish texts,” in Proc. of 2014 European Network Intelligence Conf., 2014, pp. 46-51.

[9] S. Benedettini and C. Stagnaro. (Janaury 2015). Failure to liberalise energy retail markets jeopardizes Energy Union. Energy Post. [Online]. Available: http://www.energypost.eu /failure-liberalise-energy-retail-markets-jeopardizes-energyunion/

[10] Customer switching process: How to find the right model. International Metering. [Online]. Available: http://www. metering.com/customer-switching-process-how-to-find-the-r ight-model/

[11] P. E. Lewis and A. Bogacka. (March 2014). The impact of switching process centralisation on levels of competition and the role of switching web sites in liberalised electricity markets globally writers. Vaasa ETT. [Online]. Available: http://www.vaasaett.com/wp-content/uploads/2014/03/Switc hing-Centralisation-and-Websites-040314.pdf

[12] N. Goodwin, “Making the internal market work,” British Medical Journal, vol. 308, no. 6922, pp. 206, 1994.

[13] S. Lomax. (March 2015). Micro and small business engagement in energy markets. BMG Research. [Online]. Available: https://www.ofgem.gov.uk/publications-and-up dates/micro-and-small-business-engagement-energy-market-2015-quantitative-research-report

[14] A. Pototschnig, “Overcoming the obstacles to the completion of the internal energy market,” present at the EU Energy Security Conf., Brussels, May 2015.

[15] Annual Report on the Results of Monitoring the Internal Electricity and Natural Gas Markets in 2013, Agency for the Cooperation of Energy Regulaters/Council of European Energy Regulators, Oct. 2014.

[16] Annual Report on the Results of Monitoring the Internal Electricity and Natural Gas Markets in 2012, Agency for the Cooperation of Energy Regulaters/Council of European Energy Regulators, Nov. 2013.

[17] Switched on consumer experiences of energy switching. Consumer Focus. [Online]. Available: http://www.consumer focus.org.uk/files/2013/01/Switched-on.pdf

[18] R. Buckley and A. Moss. Competition in British household energy supply markets: An independent assessment. Cornwall Energy. [Online]. Available: https://www.energy -uk.org.uk/publication.html?task=file.download&id=4886

[19] 2015 Energy Companies Satisfaction Survey, Which?, 2015.

[20] Electricity Switching, Energy UK, Apr. 2014, pp. 20-22.

[21] Electricity supplier switching ‘a(chǎn)t record’ in November. BBC News. [Online]. Available: http://www.bbc.com/news/ business-25812352

[22] T. Airu, P. E. Lewis, H. Goto, C. Dromacque, and S. Brennan. (March 2012). Impacts and lessons from the fully liberalized European electricity market-residential customer price, switching and servicest. Central Reaserch Institute of the Electric Power Industry of Japan. [Online]. Available: http://www.mojastruja.net/CRIEPI_Report_EU_energy_mar ket.pdf

[23] Consumer experiences of the energy market-wave 2. Energy-UK. [Online]. Available: https://www.ipsos-mori. com/Assets/Docs/Publications/SRI_Environment_EnergyU K_CMS_FINAL_REPORT_PUBLIC.pdf

[24] The confidence code. Ofgem. [Online]. Available: https: //www.ofgem.gov.uk/information-consumers/domestic-cons umers/switching-your-energy-supplier/confidence-code.

[25] Switching Energy Nanomagnet. Oct. 2014.

[26] Colective Energy Switch Organised by Number of Customers that Swithced Total Savings, UK Goverment, 2014.

[27] J. Yu, Z. Zha, M. Wang, K. Wang, and T. Chua,“Domain-assisted product aspect hierarchy generation: Towards hierarchical organization of unstructured consumer reviews,” in Proc. of the Conf. on Empirical Methods in Natural Language Processing, 2011, pp. 140-150.

[28] K. Haniewicz, M. Kaczmarek, M. Adamczyk, and W. Rutkowski, “A case study of sentiment orientation identification for polish texts,” in Proc. of the 2014 European Network Intelligence Conf., 2014, pp. 46-51.

[29] L. Lee and B. Pang, “Opinion mining and sentiment analysis,” Foundations Trends in Information Retrieval, vol. 2, no. 1-2, pp. 1-135, 2008.

[30] M. Hu and B. Liu, “Mining and summarizing customer reviews,” in Proc. of the 10th ACM SIGKDD Intl. Conf. on Knowledge Discovery and Data Mining, 2004, pp. 168-177.

[31] P. Goncalves, B. Fabrício, A. Matheus, and C. Meeyoung,“Comparing and combining sentiment analysis methods categories and subject descriptors,” in Proc. of the 1st ACM Conf. on Online Social Networks, 2013, pp. 27-38.

[32] R. Feldman, “Techniques and applications for sentiment analysis,” Communications of the ACM, vol. 56, no. 4, pp. 82-89, 2013.

[33] A. Mudinas, D. Zhang, and M. Levene, “Combining lexicon and learning based approaches for concept-level sentiment analysis,” in Proc. of the 1st Intl. Workshop on Issues of Sentiment Discovery and Opinion Mining, 2012, pp. 1-8.

[34] A. Sharma and S. Dey, “A comparative study of feature selection and machine learning techniques for sentiment analysis,” in Proc. of the 2012 ACM Research in Applied Computation Symposium, 2012, pp. 1-7.

[35] The election will be tweeted (and retweeted). NY Times. [Online]. Available: http://www.nytimes.com/interactive/us/ politics/2010-twitter-candidates.html?_r=0

[36] Sentiment140-Discover the Twitter Sentiment for a Product or Brand, Sentiment140, 2015.

[37] Sentiment Tool Scans Twitter to Set Super Bowl Odds, Information Week, 2012.

[38] Y. He, “Incorporating sentiment prior knowledge for weakly-supervised sentiment analysis,” ACM Trans. on Asian Language Information Processing, vol. 11, no. 2, pp. 4:1-19, 2012.

[39] C. Hansen. (July 2013). Delivering next-generation sentiment intelligence: Tailoring actionable analytics to improve. IHS Industry Sentiment Intelligence. [Online]. Available: https://gnip.com/docs/IHS-Sentiment-Intelligence -White-Paper.pdf

Wioletta Soko?owskareceived her B.S. degree in international logistics in 2010, accounting and public finance in 2011, international logistics in 2013, and informatics and econometrics in 2014, all from the Poznan University of Economics (PUE). She is currently pursuing her Ph.D. degree with the Department of Information Systems, PUE. Her research interests include energy system, big data, energy forecasting, future energy markets, energy market modeling, sentiment analysis, business process modelling, and simulation.

Tymoteusz Hossareceived his B.S. and M.S. degrees from the PUE in 2011 and 2013 respectively, both in informatics and econometrics. He is currently pursuing the Ph.D. degree with the Department of Information Systems, PUE. His research interests include energy systems, energy forecasting and modeling, and big data analysis.

Karol Fabiszreceived his B.S. and M.S. degrees from the PUE in 2011 and 2013 respectively, both in informatics and econometrics. He is currently pursuing his Ph.D. degree with the Department of Information Systems, PUE. His research interests include application of process and data mining techniques in the energy domain, energy prosumers profiling, smart grids, and business process modeling and simulation.

Witold Abramowiczis a full professor and the Chair of the Department of Information Systems, PUE. His areas of interest cover information acquisition and filtering. He was involved in a number of EU and national funded projects. He is the author of a number of journal publications and books.

Mateusz Kubaczykreceived the B.S. degree in informatics and econometrics from the PUE in 2011. He is currently pursuing the M.S. degree with the Department of Information Systems, PUE. His research interest is sentiment analysis.

Manuscript received May 26, 2015; revised July 10, 2015.

This work was supported by the HPI Future SOC Lab and Tableau Software.

W. Soko?owska, T. Hossa, K. Fabisz, and M. Kubaczyk are with the Department of Information Systems, Faculty of Informatics and Electronic Economy, Poznan University of Economics, Poznań 61-895, Poland.

W. Abramowicz is with the Department of Information Systems, Faculty of Informatics and Electronic Economy, Poznan University of Economics, Poznań 61-895, Poland (Corresponding author e-mail: w.abramowicz@gmail.com).

Digital Object Identifier: 10.11989/JEST.1674-862X.505261