The predictive power of search queries and sentiment analysis in the sales of high involvement goods

The Predictive Power of Search Queries and

Sentiment Analysis in the Sales of High Involvement

Goods

Date of Submission: 22-06-2018 | Final Version

Qualification: MSc Business Administration - Digital Business Name: Edo Santema

Student Number: 11406518

Institution: University of Amsterdam

Supervisor: Prof. em. dr. ir. Hans J. Oppelland Word count: 11.384

Abstract

It was researched that 70% of the car buyers who used YouTube as part of their buying process were influenced by the video content they watched prior to the purchase (Jarboe, 2016). This research examines the use of Google Trends data and sentiment analysis on YouTube comments as a tool to explain and forecast automobile sales in the United States. A YouTube comment scraper used to mine data. In particular, 481 YouTube videos were manually selected for 15 car models, yielding 47.090 useful comments. Contemporary research has extensively examined the use of sentiment analysis as a tool for the explanation of sales of low-involvement goods. Many of those indicate that further research is needed on the application of sentiment analysis for the sales prediction of higher involvement goods, such as car sales.

The results of the research show that YouTube sentiment and Google Trends data have negligible explanatory power for automobile sales. The results were weak and mostly in the wrong direction of the intended research framework. The moderation of price was significant, showing stronger results for lower-priced automobiles. For a minority of the car models it was exemplified that the implementation of a time lag significantly enhanced the correlations. The results of this research impose new limits on the application of sentiment analysis in the area of high involvements goods.

Statement of originality

This document is written by Edo Santema who declares to take full responsibility for the contents of this document. I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it. The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

Acknowledgements

Foremost, I would like to thank my supervisor Hans Oppelland for his continued support and input during the writing process of this Master thesis. In addition I would like to thank Lexalytics for providing this research with the needed sentiment analysis tools, and Bart Demand for providing the required data on the car industry. Besides, I want to thank Sunny de Blok and Bram Walda for their fruitful discussions and input helping this research finalise.

1. Introduction 1

1.1 Problem Statement 1

1.2 Research Objective 2

1.3 Research Methods 3

1.4 Structure of Research 3

2. State of the Art 4

2.1 Sentiment Analysis 4

2.2 Sales Prediction with Social Platforms 6

2.3 YouTube and Sentiment Analysis 9

2.4 Customer Behaviour and Social Platforms 11

2.5 Time Lag 12

2.6 Automobiles and Sentiment Analysis - Causality 14

3. Research Design 15

3.1 Conceptual Model 15

3.2 Hypotheses 16

4. Method and Research Sample 18

4.1 Data Sources 18

4.1.1 Car Sales 18

4.1.2 Google Trends 19

4.1.3 YouTube Data 20

4.1.4 YouTube Sentiment 20

4.2 Data Collection and Extraction 22

4.2.1 Car Sales 22

4.2.2 Google Trends 23

4.2.3 YouTube 24

5. Analysis & Interpretation 26

5.1 Sample Characteristics 26

5.2 Development of Sales 26

5.3 Independent Samples T-test - Influence of Price 27

5.4 Regression Analysis 29

5.5 Multiple Regression 31

5.6 Time Series 34

6. Discussion 37

6.1 Key Findings 37

6.2 Theoretical and Practical Implications 40

6.3 Limitations and Future Research 41

7. Conclusion 43

References 44

1. Introduction


1.1 Problem Statement

The opinions of others have proven to be an essential factor influencing many decisions in our life. (Pang & Lee, 2008). Long before the internet existed people used the opinions and recommendations of others to choose whom to vote for, which restaurant to go to or which lawnmower to buy (Pang & Lee, 2008). Previously, when companies researched the opinions and sentiments of their customers they needed to conduct physical polls, interviews or surveys (Liu, 2010). The widespread evolution of the World Wide Web has made it possible to review the opinions and experiences of large amounts of consumers (Pang & Lee, 2008). For many consumers social media has become an extension of their lives and therefore has become an appendix of their body and mind (Lassen et al., 2014). The rise of social media platforms in the last decade has created enormous amounts of user-generated content (UGC), providing multiple platforms of online discourse for products and services (Asur & Huberman, 2014). The social media platforms could be beneficial for firms to deduct information from. Millions of consumers generate online word of mouth (eWOM) creating massive amounts of useful information for business owners (Asur & Huberman, 2014). It has become evident that during the consideration phase consumers are not solely dependent on their family and friends but are able to use the internet to explore the overall sentiment. Likewise, for firms it is not necessary to exclusively use offline forms of opinion mining such as surveys to research the sentiment of their customers. As companies always have the urge to control or monitor the reputation of their products, new methods to analyse large data flows have been developed. An increasingly popular approach to analyse these social media data flows is by using machine learning algorithms to classify the data flow in either positive, neutral or negative categories (Pang & Lee, 2008). Ostrowski (2010) defines consumer sentiment as the harmony of feelings that consumers have about a product or service. Understanding these sentiments

can be useful to explain and predict events, such as the valuation of stocks, election outcomes or sales volumes (Pang & Lee, 2008). Past research has exemplified how the sales of low-involvement goods, such as movie sales, books and iPhone sales can be explained and predicted using sentiment analysis (Asur & Huberman 2010; Dijkman, Ipeirotis, Aertsen, & Van Helden 2015; Lassen, Madsen, & Vatrapu, 2014). However, does this also hold for high involvement goods such as automobiles? Current literature indicates that further research is needed in this matter (Lassen et al., 2014; Wijnhoven & Plant, 2017). This thesis will use car sales in the US as an example of a high-involvement product to test several hypotheses. What could be the potential explanatory role of online search activity and online sentiment on car sales volumes? Is the repeated positive or negative mentioning of a car model in YouTube comments, or increased search activity on Google related to its sales volumes? The results of this research will contribute to the current academic literature by researching the effect of sentiment in the comments of YouTube car reviews and Google search activity on the sales of high-involvement goods. Through the addition of this type of products the explanatory value of sentiment analysis can be further analysed, and possible conclusions can be drawn to its explanatory value. The results of this research can be particularly crucial for car manufacturers. Automobiles are either custom-built or built based on a sales forecast. Even if cars are custom built to order, forecasting the demand for these types of cars can help car manufacturers to allocate their resources more efficiently. Understanding how sales relate to sentiment and search queries opens a new possible source to forecast these car sales.

1.2 Research Objective

This research will exploit Google Trends data and YouTube data to explain car sales volumes. In gathering all these forms of data this research attempts to create an explanatory model for car sales in the US market. It builds upon the research of Choi & Varian (2012) to explain the

present and build upon that model to create a prediction model. The central research question will be as follows: “What is the explanatory power of Google Trends and sentiment analysis mined from Youtube comments for car sales in the US ?”

1.3 Research Methods

This research will use multiple case studies exemplified by the sales volumes of several car models. This will be done by using datasets derived from the Carsalesbase database, Google and YouTube. The Carsalesbase database can provide accurate datasets for the sales volumes and pricing of cars. The second dataset will be deducted from Google Trends which gives us an insight in search queries of consumers for particular car brands in a selected timeframe and market. Lastly, the sentiment analyser ‘Semantria’ is used to analyse selected YouTube comments for sentiment. The Google Trends and sentiment dataset derived from YouTube will be used to find a correlation to the sales datasets of the Carsalesbase. To explain changes in sales it is essential to research the time lag for the product under scrutiny. In this research it is important to acknowledge how long the process from the initial search to the purchase of a car will take.

1.4 Structure of Research

After the introduction this research will start with exploring the contemporary literature. In the third chapter a research framework with hypotheses is presented, which will be the conceptual basis of this research. The fourth chapter will go into the method and research sample, mainly by discussing the various data sources and the means of data extraction. The fifth chapter will statistically analyse and interpret the collected data. The sixth chapter will discuss the main findings, the implications, limitations and will propose future research objectives. The seventh chapter will conclude the research.

2. State of the Art

This state of the art will present the recent findings in the area of sentiment analysis and opinion mining as well as the recent developments in the academic debate about prediction models based on social media feeds for product sales. In doing so, the literature review concludes with a knowledge gap that will be the basis of the research in this thesis.

2.1 Sentiment Analysis

Sentiment analysis, also known as opinion mining or subjectivity analysis has become a hot topic in many research fields, ranging from finance and sales to international relations (Pang & Lee, 2008). The area of sentiment analysis has enjoyed a great burst in activity since the early 2000s, mainly motivated by computational progress and the availability of data through the advancement of the World Wide Web (Serrano-Guerrero, A. Olivas, P. Romero, & Herrera-Viedma, 2015). The goal of sentiment analysis is to analyse the opinions, emotions, attitudes of authors towards products, services, organisations and properties (Serrano-Guerrero et al., 2015). Nasukawa and Yi (2003) describe sentiment analysis as a method to identify the expression of sentiments in texts and the technique to indicate whether these expressions are favourable or unfavourable opinions towards the subject of the text.

The subject of sentiment analysis can clearly be explained by investigating the inherent challenges while using sentiment analysis tools. Serrano-Guerrero et al. (2015) classify sentiment analysis tools into five tasks with accompanying challenges. The first task is the classification of sentiment, which is based on the idea that a specific text expresses an opinion on an entity from a holder and tries to measure that sentiment. This measurement can be categorised into 3 broad categories: negative, neutral and positive. This seems a simple task but it is highly sophisticated, especially when a single author holds multiple opinions or when the author attributes these opinions to different attributes of the target (Serrano-Guerrero et

al., 2015; Wijhoven & Plant, 2017). This is closely related to sentiment ranking which reflects the measurement of the intensity of the sentiment. Different scales are used to indicate the intensity of the sentiment ranging from stars (1 to 5) to points (-1 to +1). The second challenging task of sentiment analysis is the classification of subjectivity. Subjectivity classification has the primary purpose of detecting whether a sentence holds a certain degree of subjectivity and should be seen as a task preceding the classification of sentiment. Liu (2010) indicated that subjective information implies that a person holds a personal belief, feeling or attitude towards an object in question, while objective information consists of factual information about the world. Subjectivity is not a prerequisite for sentiment, but both go hand in hand in the majority of the cases (Liu, 2010). The process of subjectivity classification is a process considered more difficult than sentiment classification but when done right will increase the accuracy of sentiment classification to a great extent (Serrano-Guerrero et al., 2015). Opinion Summarisation is the process in which the sentiment analysis tool summarises the sentiment of the author throughout the document or multiple documents. The challenges with summarisation include the aggregation of votes, the selective highlighting of some opinions, the representation of points of disagreement and the accounting for different levels of authority of the opinion holders (Pang & Lee, 2008). The fourth challenging task for sentiment analysers is to extract and present the sentiment of the author, which is closely related to opinion summarisation. The fifth, and perhaps major challenge that remains present for sentiment analysis is the detection and correct classification of ironic information (Serrano-Guerrero et al., 2015). Irony is extremely hard to identify for non-humans, since on the surface the meaning of the information differs radically from what the actual case is (Serrano-Guerrero et al., 2015). Most of the aforementioned challenges can be summarised by an example by Liu (2010): 


“(1) I bought an iPhone a few days ago. (2) It was such a nice phone. (3) The touch screen was really cool. (4) The voice quality was clear too. (5) Although the battery life was not long, that is ok for me. (6) However, my mother was mad with me as I did not tell her before I bought it. (7) She also thought the phone was too expensive, and wanted me to return it to the shop.…” (Liu, 2010).

The text holds different opinions, both positive (2,3 and 4), negative (5,6 and 7) and neutral (1). Parts of the opinions are targeted towards the iPhone as a whole (2) or attributes of the phone (3,4,5). In addition the sentence holds multiple holders of opinions. In sentence (2), (3) (4) and (5) the reviewer (the “I”) holds an opinion, but in sentences (6) and (7) the mother of the reviewer holds a certain opinion about the iPhone. 


The challenges recognised by Serrano-Guerrero et al. (2015) and exemplified by Liu (2010) have been resolved to a great extent with the current accuracy of contemporary sentiment analysers such as Semantria, which will be used in this research. Semantria is a multilingual sentiment processing engine which allows the processing of information from many sources such as tweets, comments but also plain texts (Serrano-Guerrero et al., 2015). It has to be recognised that even the most sophisticated sentiment analysers cannot reach full accuracy. However, high accuracy scores can be reached (Serrano-Guerrero et al., 2015).

2.2 Sales Prediction with Social Platforms

Multiple studies have been performed in the area of sales prediction with the assistance of sentiment analysis or web search queries, most of them successful but many indicate that future research is needed towards different product categories. Pioneers using online content and sales prediction were Gruhl, Guha, Kumar, Novak, & Tomkins, (2005). In their research

they mined textual data from online books blogs to forecast spikes in book sales (Gruhl et al., 2005). They researched whether book blogs exhibited a certain pattern prior to sales spikes. The introduction of comprehensive online social platforms introduced new methods to mine content. The first researchers successfully using the combination of social media and sentiment analysis were Asur and Huberman (2010). In their research they pioneered by using social media content to predict future movies sales. In particular they used Twitter data to predict box-office revenues for movies (Asur & Huberman, 2010). Asur & Huberman (2010) categorised tweets as being either positive, neutral or negative and introduced the Positivity to Negativity Ratio (PNratio) “p/n” where the “p” is a positive tweet, and “n” is used to indicate a negative tweet. They have demonstrated that both exploiting the buzz (the total mentions of a subject) as well as the sentiment analysis of the buzz (categorising the buzz into a negative, positive or neutral category) has proven to be a more powerful prediction method than using market-based approaches, such as the Hollywood Stock Exchange (Asur & Huberman, 2010). They exemplified that there is a strong correlation between the amount of attention for a movie and its future ranking (R .90, R-square .80). Recently Twitter announced that historical access to tweets will be limited to 7 days, making it incompetent for any future application of academic and business research without running a real-time server that collects data over a longer time period. Goel, Hofman, Lahaie, Pennock, and Watts (2010) used search query data gathered from Yahoo to predict the future sales of movies, first month-sales of video games and the ranking of songs in the Billboard top 100 chart. Results showed that there was a strong correlation between the search queries and the real-life outcomes (Goel et al., 2010). Building upon this research Choi & Varian (2012) researched the ability to explain current and future sales with search query data for cars, unemployment applications, excursion planning and consumer confidence using Google Trends. Lassen et al. (2014) progressed using sentiment analysis by analysing iPhone sales. Their research

demonstrated that there is a strong correlation between iPhone related tweets and iPhone sales, and that the correlation becomes increasingly stronger after incorporating the sentiment of the tweets. Dijkman et al., (2015) built upon the research of Asur and Huberman (2010) by exemplifying how Twitter activity is able to moderately predict the sales of products that are less related to a digital environment than for example phones and movies. Voortman (2015) set the first steps analysing a higher involvement good with Google Trends. The level of involvement indicates how personally interested you are in consuming an entity and how much information you need to decide to acquire the entity (Mittal, 1989). The level of involvement can be mapped on a continuum which reaches from routine purchases to purchases that need extensive thought and a high level of involvement (Mittal, 1989). Voortman (2015) used Google Trends as an independent variable to predict the sales of automobiles in the Dutch market. By introducing a time lag for the purchase Voortman increased the explanatory power of the independent variable to a great extent. Wijnhoven & Plant (2017) built upon this research by the inclusion of social media content and sentiment analysis performed by the software tool Coosto relating to car sales in the Dutch market. Coosto is an automated software tool that gathers data from multiple social platforms relieving the researcher from using API technology and independent sentiment analysis. Coosto recently announced that it has ceased its functionality to look further than one month in the past, which is mainly due to the Twitter restriction which is their main source of data (Wijhoven & Plant, 2017) This restriction makes Coosto incompetent for future historical research looking further than one month in the past. This research will be closely related to Wijnhoven & Plant (2017) in that it is using the same industry. However this research will be inherently different by researching a larger geographical scope, by using an inherently different social media platform (YouTube) and an independent method to analyse the opinions of YouTube viewers (Semantria). YouTube has not been used intensively as a

platform for sentiment analysis although it has a significant effect on the decisions of consumers (Jarboe, 2016). It was it researched that 60% of the car shoppers enter the market unsure of which car to buy and that 70% of the people who used YouTube as part of their buying process were influenced by the video content they watched (Jarboe, 2016). Besides, the amount of views of test-drives, reviews and walkthroughs has significantly increased in recent years (Jarboe, 2016). Moreover, car reviews of independent reviewers are the number one source of information for car buyers in the United States (Kandaswami & Tiwar 2014). What is evident from the current academic literature is that the correlation between total buzz, sentiment and sales have increasingly been focused on lower involvement goods such as movies, books and games and that the data was mainly based on Twitter data (Asur & Huberman, 2010; Goel et al., 2010; Gruhl et al., 2005; Lassen et al., 2014). This research will add to the current academic literature by exploring the correlation between Google search query data and sentiment deducted from Youtube feeds to explain the sales of high involvement goods, exemplified by cars in the second largest automobile markets in the world, the United States.


2.3 YouTube and Sentiment Analysis

YouTube is one of the most popular websites worldwide, ranking second on the Alexa top 500 most popular websites (Alexa, n.d.). Moreover, YouTube has been ranked as the second largest search engine after Google, and we watch more than 1 billion hours of YouTube video content every day, more than Netflix and Facebook combined (Smith, 2018). The platform has started to exist in 2005 with the simple functionality to share videos. Since then the platform has evolved in the most popular video sharing website with functionalities to create personal profiles, friending, sharing, subscribing, liking and commenting. The increased importance of Youtube is reflected by its increased usage: 18-49 years olds spent 4% less time

watching tv while time on YouTube went up by 74% (Smith, 2018). On mobile devices alone, YouTube reaches more 18-49 years-old than any cable tv provider (Smith, 2018). YouTube can be classified as a typical online social network which is extensively used by multiple layers of the society to project identities and associate with specific social groups (Smith, Fischer & Yongjian, 2012).

Although YouTube’s increased importance and popularity it seems to have attracted little social science research on sentiment analysis compared to other social media platforms (Thellwall & Sud, in press). Several researchers have addressed issues with the help of sentiment analysis for YouTube for comments. In 2009 Bermingham, Conway, McInerny, O’Hare & Smeaton used sentiment analysis to identify the potential for jihadi radicalisation on YouTube channels. They found that radicalisation mostly takes place on designated forums and that YouTube plays no significant role in this (Bermingham et al., 2009). Sierdorfer, Chelaru, Nejdl, & Pedro (2010) researched the ability to use sentiment analysis to predict the number of likes a certain comment could get. In several experiments they could predict the community acceptance for a particular comment to a great extent (Sierdorfer et al., 2010). In 2011 Thelwall and Sud researched YouTube sentiment to examine who participates in discussions in the comment section. Their research revealed that a typical YouTube comment was posted by a 29-year old male and contained 58 characters (Thelwall & Sud, 2011). Further research on sentiment analysis in comments was performed by Oksanen et al. (2015). They researched the emotional content of comments in pro-anorexia and anti-pro-anorexia YouTube content. They found that anti-pro-anorexia content yielded more positive sentiments (Oksanen et al., 2015). Furthermore, in 2015 Severyn, Moschitti, Uryupina, Plank, and Fillipova researched the targeting of the sentiment in comments. The central question was whether a comment was targeted towards the video itself, or the product under scrutiny in the video. The results were mixed, showing that both the video as well as the

product were the target of the comment. This is important for this research and will be further discussed in the limitations. Contemporary literature on YouTube sentiment analysis shows that there is no research in the area of sales explanation or prediction, in which this research will be a pioneer.

2.4 Customer Behaviour and Social Platforms

The AIDA model has proven to be a helpful model while mapping the influence of social media and UGC along the road to purchase decisions (Lassen et al., 2014). The AIDA model stands for (1) attention, (2) interest (3) desire, and (4) action (Lassen et al., 2014). According to Lassen et al., (2014) social media and UGC can play in role in multiple stages of the AIDA model. The attention stage can result from reading social media posts, whereas the interest stage can be represented by Google search queries or directed search for videos by consumers (Lassen et al., 2014). Since Google Trends data only shows the relative amount of search queries and not the sentiment of search queries it might not be applicable to the desire stage which reflects a sentimental opinion about the product. A comment in a YouTube video can reflect an opinion and thus a desire. Following the hierarchy of the AIDA model we would expect that the desire stage would lead to a higher probability of purchase (action) than the interest and attention stage. This is the results of the so-called Hierarchy of Effects (HoE) formulated first by Lavidge and Steiner (1961). Therefore there is the expectation that data which reflects an opinion with a desire would lead to higher correlations with sales figures than data found in the interest stage. In this research we are looking at a high involvement good where we expect that consumers carefully scrutinise the product in question, opposite to routine purchases. Google might therefore still be a very helpful indicator as we expect that the consumer will routinely search for the product on the internet before purchasing the product.

A comparable theory which explains the customer purchasing journey is the Consumer Buying Process Theory created by Kotler (1994). The stages of Kotler’s theory are as follows: (1) problem recognition, (2) information search, (3) evaluation of alternatives, (4) purchase decision, (5) post-purchase behaviour. This theory is different from the AIDA model in that the problem recognition stage comes from the consumer self, opposite to creating attention from externalities (Lassen et al., 2014; Wijnhoven & Plant, 2017). The model might be more appropriate for the purchase of high involvement goods since consumers get less persuaded to purchase a car after gaining attention but is to no small extent self-reliant for the recognition of the problem (Wijhoven & Plant, 2017). The consumer might recognise the problem of purchasing a new product, and might be directly influenced in the information search stage by exploiting social media where we expect that the consumer is more approachable to new information. We expect that consumers look up other alternatives in the “evaluation of alternatives” stage after the influence of social media in the “information search” stage (Wijnhoven & Plant, 2017). In the “evaluation of alternatives” the applicable time lag is present, which will be discussed next (Voortman, 2015).


2.5 Time Lag

A time lag represents the moment between which the consumer is researching the product and the conclusion of the research by deciding to purchase the product (Voortman, 2015). This is reflected in the AIDA model between the t1(attention) and t4 (action) and in the Kotler model (figure 1) between the information search stage (2) and the purchase of the product (4). This research will use Kotler’s (1994) theory to map the buying process, because it is deemed more appropriate for high involvement purchases where consumers recognize the problem inherently (Wijhoven & Plant, 2017).

"

The time lag might differ between products. Lassen et al. (2014) exemplified that a time lag of 20 days between the start of the consideration to acquire an iPhone until the purchase of an iPhone shows the strongest correlation. Asur & Huberman (2010) researched the time lag related to the prediction of future movie sales and concluded that a time lag of 14 days to be most accurate. Somervuori & Ravaja (2013) have researched that higher priced products lead to an increase in the customer decision time, which is particularly applicable to the product under scrutiny in this research. Putsis & Srinivasan (1994) have studied the time it takes for consumers to purchase an automobile. The moment from recognition until the purchase is for 16% of the consumers less than 1 month. For 34% of the car buyers the purchase is completed in 3 months, 16% within 6 months, and 15% within 7 months. More recently Kandaswami & Tiwar (2014) researched that consumers spend more than 10 hours researching possible vehicles and that men tend to use more time to research car purchases than females. The applicable time lag is applied by Voortman (2015), who discovered that 15 out of 17 car models showed a higher correlation between sales and web search queries after imposing a time lag. This research will examine if a time lag within the time capabilities of our datasets leads to more insights.

2.6 Automobiles and Sentiment Analysis - Causality

Automobile sales are subject to many variables (Heinig, Nanda, & Tsolacos, 2016). The intention of this research is not to exclusively explain sales volume changes by YouTube data, sentiment analysis or Google activity, rather than to contribute to the explanation of differences in sales volumes between car models based on the data mentioned above. As indicated by Lassen, La Cour, and Vatrapu (2017) it might be that many variables explain a particular phenomenon, social media data may just be one of the variables that explain a phenomenon. In building a model to explain any phenomena related to social media data it is increasingly important to exploit the quantity and quality of the data (Lassen et al., 2014).

3. Research Design

3.1 Conceptual Model

The research model will start by exploiting the correlation between Google Trends search query data and car sales in the US. Subsequently, in a search for more explanation, YouTube data will be used by analysing the sentiment in the comments of selected YouTube videos. The dependent variable will be the volume of car sales. In doing so this research will exploit if web search queries data and sentiment analysis are useful sources to explain volume changes in the dependent variable. The research is limited to a timespan of 11 months due to data availability, which is inherent in the usage of YouTube comment data. This is due to time identification issues which will be covered in a later section.

H3 H5 H1 H2 H4 Ind. Variables

• Google Query Data • YouTube Comments • Sentiment of Comments Dep. Variable • Car Sales Social Data • Google • YouTube • Price • Time Fig.2: Conceptual Model

3.2 Hypotheses

Following from the literature review several hypotheses are formulated which are reflected in the conceptual model. Multiple researchers have shown that the number of search queries has a high probability to be correlated to the number of sales of a product (Asur & Huberman, 2010; Choi & Varian, 2010; Dietzel, 2016; Goel et al., 2010) The first hypothesis will therefore be:

H1: “The number of sales of a particular car model will increase with the increased search query volume of Google Trends of the particular car model.” 


Besides, past research has exemplified that a more positive sentiment will lead to a positive influence in sales (Asur & Huberman, 2010; Lassen et al., 2014). We expect that more positive UGC surrounding car models will increase the sales of the model. The second hypothesis will therefore be as follows:

H2: “A more positive sentiment (higher PNratio) surrounding a certain car model on YouTube will have a positive influence on the sales of that car model.”

Taken the assumption that people who own a car will create content about their car holds for a majority of the car brands. However, this research has to take into account that certain cars, such as exotic cars, will receive a lot of positive sentiment from a fan base who will not always purchase the car. Therefore the third hypothesis will be as follows:

H3: “Higher priced cars will have a weaker correlation than lower priced cars between social media data and car sales.”


The fourth hypothesis will deal with the more considerable explanatory value of sentiment analysed data than Google Trends data. Following the hierarchy of effects in the consumer buying process model we expect that data which reflects an opinion would lead to higher correlations with sales volumes than search query data which is unable to reflect an opinion (Asur & Huberman, 2010; Gruhl et al., 2005; Goel et al., 2010; Lassen et al., 2014). The fourth hypothesis will therefore be as follows:

H4: “The correlation between the sentiment of YouTube comments and car sales is stronger than the correlation between Google Trends and car sales.”

Voortman (2015), Wijnhoven & Plant (2017) and Lassen et al. (2014) have researched that the introduction of a time lag will increase the correlation between sales and social media content. This research will exploit if this is also applicable to the social media under scrutiny in this research, Therefore the fifth hypothesis is as follows:

H5: “When a time lag is introduced the correlation between sales and social media content will strengthen.”

Table 1: Hypotheses Summary

H1 Higher search query index Higher sales

H2 More positive sentiment (Higher PNratio) Higher sales

H3 Higher priced cars Weaker correlation between variables

H4 Sentiment Stronger correlation between variables

than search query

4. Method and Research Sample

Due to the increased possibilities of data collection through the use of online tools such as Google Trends and the Youtube API it has become possible to conduct quantitative research on a larger scale. The collected data will be used for deductive research (Saunders, Lewis and Thornhill, 2012). In the literature review we have created a platform of hypotheses which will be tested with the respective datasets.


4.1 Data Sources

To test the relationship between car sales and social media 3 data types are collected: car sales, Google Trends data and YouTube data and sentiment. There are challenges in collecting and combining these 3 data types which will be reflected in the next chapters.

4.1.1 Car Sales

This thesis is building upon earlier research in the automotive industry by extending the geographical scope to the second largest car market in the world, the United States. The car sales dataset that will be used is derived from Carsalesbase. Carsalesbase is a database of all monthly car sales in the US, Europe and China created by the Dutch automotive analyst Bart Demand (http://carsalesbase.com). The sales data can be specified at a detailed level including body style, fuel type, gearbox, colour or even specific options. This research has selected the US as a geographical demarcation for several reasons. The largest car market in the world is China ("Largest automobile markets worldwide", n.d.). However, mainland China does not have access to Google and thus has no resulting Google Trends data. In addition the Chinese population has no access to YouTube voiding any possibilities to mine Chinese Youtube search and comment data. The second and third car markets in the world

are the US and Europe, which both have access to YouTube and Google. However, Google Trends does not allow the regional specification of “Europe” and is only able to select separate countries inherently making the Google search data for Europe as a whole impossible. The US as a whole can be selected in Google Trends and can be used as a location setting in YouTube, which Europe cannot. This leads to the conclusion that the United States is the most relevant market which we can research with the available data.

4.1.2 Google Trends

The second sample that will be used consists of Google Trends data. As more and more people turn towards the internet as a source of information it is tempting to view the search activity of customers as a reflection of the current state of interest for the searched attribute. As stated by Goel et al. (2010) it is from this perspective a short step to conclude that what people are looking for today might have predictive value of what they will do tomorrow. The data provided by Google Trends can be specified according to subject, time, market and country. The data is shown as a relative monthly number to the total search query activity in the selected region: Google is dividing the query volume of the search term by the total number of searches in the specified region in the selected time frame (Choi & Varian, 2012). The data is shown as a time series index of the volume of queries consumers insert on the Google search frame (Choi & Varian, 2012). The maximum query share over the time is normalised at 100 points with the initial query share being at 0 points. When the score hits 100 points, the interest is at its peak within the selected timeframe. The normalisation of data is particularly important because of the increased popularity of Google searches over time. Raw numbers would not give any way to compare between contemporary searches and historical search queries (Rogers, 2016). Moreover, the exact number of searches are not published by Google. The data retrieved from Google Trends is relevant to the research

question. Following the consumer buyer decision theories Google Trends exemplifies the information stage in the Kotler (1994) model.

4.1.3 YouTube Data

The choice for the usage of YouTube data is based on the amount of content and activity for automobiles on the platform, its influence on consumers and the addition of this research to the current literature on the explanatory power of YouTube in the subject of sales explanation and prediction. As exemplified by multiple studies there are many other platforms which potentially could be used to research the correlation between sentiment, mentions and sales, such as Twitter, Facebook and Instagram (Asur & Huberman, 2010; Goel, Hofman, Lahaie, Pennock, and Watts 2010; Gruhl, Guha, Kumar, Novak, & Tomkins, 2005; Lassen, Madsen, & Vatrapu, 2014). However, all platforms mentioned above never had the ability to mine content, have ceased the ability to mine content over a more extended period or have ceased the option to mine content entirely. This is mainly because of privacy breaches or economic justifications. YouTube is still an open platform from which sentiment can be mined, although with a historical limitation of 11 months. This is a limitation to the research but inevitable when using YouTube as a data platform.

4.1.4 YouTube Sentiment

The sentiment analyser Semantria will process the sentiment analysis of YouTube comments. Semantria is one of the most used and most reliable sentiment analysers on the market (Lexalytics, 2015). The first developer package starts at $1500 per month (Lexalytics, 2015) but after contacting the Lexalytics sales department two packages of 50,000 queries were offered for this research. The Semantria algorithm is a black box, all we know is that it is optimised for 22 languages and uses intensive ways of deep learning. Semantria has been

optimising its engine for the last 10 years and is one of the (self-proclaimed) major players in the natural language processing (NLP) industry (Lexalytics, 2015). The unknown algorithm is a limitation to the research but also present in other related research and understandable from an economic standpoint (Lassen, et al., 2014; Wijnhoven & Plant, 2017). The actual sentiment analysis can be executed by using different programming languages such as Python, Java or by using an MS Excel plugin. The timeline of the research will be 11 months based on the availability of data and related research (Asur & Huberman, 2010; Goel, Hofman, Lahaie, Pennock, & Watts 2010; Gruhl, Guha, Kumar, Novak, & Tomkins, 2005; Lassen et al., 2014). Similar to Lassen et al. (2014) and Asur & Huberman (2010) this research will use the PNratio to classify the comments. The researchers mentioned above used the following definition to define the sentiment of comments.

p: comments with positive sentiment
 n: comments with negative sentiment o: comments with neutral sentiment With subjectivity being:

And the positive to negative ratio (PNratio):

In this research the PNratio will be calculated as the amount of positive to negative comments in a particular month for a car model. It is chosen to leave neutral comments outside the ratio because they do not show subjectivity that can be analysed. The total amount of comments

p + n Subjectivity = ——— o p PNratio = — n

will be included as a separate variable in the statistical analysis, which is also a reflection of the neutral comments.

4.2 Data Collection and Extraction

The data for this research has been extracted from 3 independent databases: Carsalesbase, Google Trends and YouTube with subsequent analysis through Semantria. The extraction of the data has different challenges which will be discussed in the next subchapters.

4.2.1 Car Sales

Carsalesbase provides the car sales data with sales numbers per month of different car models in separate regions. This dataset is directly derived from car manufacturers and national automobile registration offices and is extremely accurate. For several car models, such as Ferrari and Tesla it is impossible to gain monthly sales numbers because the manufacturers only publish yearly numbers. For this reason these car models are excluded from the research. Besides, cars with multiple branding names such as “BMW 5-Series”, “BMW 5-er”, “BMW-5 Serie” are excluded from the research. The selection of cars is based on sales volumes in the US (Thompson, 2018) in combination with price differences (low to high) to test the price moderator. Prices are retrieved from www.truecar.com. The selected car models have no new model introduction in the selected timeframe and the selected automobiles fall into multiple segments as indicated by the Euro Market Segment (European Commission, 1999). The Euro Market segmentation is the most detailed classification of cars available. Despite its name it is

not limited to car models in the European market (European Commission, 1999).

4.2.2 Google Trends

The data collection procedure for the Google Trends is fairly straightforward; the car brand can be inserted as a search query, together with the time frame, market and region. The data can then be downloaded as a .csv file. When selecting certain time frames Google Trends shows the data in weeks of which some naturally tend to be in separate months. This has to be corrected by manually inputting it from the graph on the Google Trends website into SPSS. The search terms that are being used in Google Trends are the same as the brand

Table 2: Selected Car Models

Model Base Price US $ Class (Euro

Segmentation) Search Term Google Trends Search Term Youtube Hyundai Elantra US$ 16.423,00 C Hyundai Elantra Hyundai Elantra

Review

Honda Civic US$ 18.840,00 C Honda Civic Honda Civic Review Ford Fusion US$ 22.120,00 D Ford Fusion Ford Fusion Review Toyota Camry US$ 23.495,00 C Toyota Camry Toyota Camry

Review Honda Accord US$ 23.600,00 D Honda Accord Honda Accord

Review

Ford Escape US$ 23.940,00 J Ford Escape Ford Escape Review Toyota RAV4 US$ 24.510,00 D Toyota RAV4 Toyota RAV4

Review Ford Explorer US$ 35.170,00 J Ford Explorer Ford Explorer

Review Jeep Grand

Cherokee US$ 39.740,00 J Jeep Grand Cherokee Jeep Grand Cherokee Review

BMW i3 US$ 41.350,00 B BMW i3 BMW i3 Review

Chevrolet

Silverado US$ 42.290,00 J Chevrolet Silverado Chevrolet Silverado Review Volvo XC90 US$ 46.900,00 J Volvo XC90 Volvo XC90 Review Chevrolet

Corvette US$ 55.495,00 S Chevrolet Corvette Chevrolet Corvette Review Porsche 911 US$ 91.100,00 S Porsche 911 Porsche 911 Review Mercedes AMG

name from Carsalesbase. In addition the brand names are inserted with quotations to ensure that the car brands are not taken out of context (for example: the “Fusion” in “Ford Fusion” could refer to other search terms, such as nuclear fusion). The time frame is selected in accordance with the data available from YouTube, and the geographical location is marked to the United States.

4.2.3 YouTube

YouTube videos were scraped using a YouTube scraper which is available on GitHub. The scraper is developed by ‘Philbot9’ and is able to mine comments from individual YouTube videos. The aforementioned scraper is used because can scrape the comments of individual videos whereas the YouTube API allows developers to scrape channels, in which this research is not particularly aimed at. The process of scraping a video could take up to 9 minutes, dependent on the amount of information that was being scraped. The comments can be downloaded as a .csv file together with a timestamp, username, comment text, and date. For every video this process has to be completed manually. The monthly timestamp of the video is limited to a maximum of 11 months. All videos between 12 and 23 months are labeled as one year old, from 24 months to 35 months as two years old and so forth. This is complicating a monthly sentiment ratio for different car models. Since this research is interested in a monthly timespan it is limited to mine comments of a maximum age of 11 months (until June 2017). For every car model as many videos as possible are collected in the selected timeframe plus 12 months (until 01-06-2016). Sampling has shown that videos up to a year old still yield a significant amount of new comments (Thelwall, 2017). The YouTube videos are selected based on multiple criteria: the video has to fall within the selected timeframe minus 12 months (01-06-2016 until 01-05-2018), comments have to be enabled, the main language has to be English and the video needs to have a view count of at least 50.000 views. The latter has

been done to confront the lengthy manual processing of videos without the valuable comments for sentiment analysis (Thellwall & Sud, in press). The location settings for YouTube are set to the United States to optimise the connection between the independent and dependent variable. Besides, videos are selected only when they deal with the specific car model, no comparisons of multiple models are included to confront the inclusion of car models outside our analysis. For every car model up to 52 videos are manually selected and scraped, in total 481 videos were scraped yielding 103.281 comments. The videos are collected in a 4 day period to confront the divergence in time of comment data. The separate data files for every video were merged into an overarching .csv file per car model. The Excel file is cleaned for any erroneous data, comments outside the selected timeframe are deleted and all comments are filtered on time in order to create a timeline of comments between June 2017 and April 2018 giving a total of 47.090 comments. In order to filter on data the Unix timecode, which represents the number of seconds elapsed since the 1st of January 1970 had to be translated to a human-readable date. This is done by using the formula =((cell/86400) + 25569 +(-5/24). The separate files for the car models are then uploaded into the Semantria analyser. In the Semantria analyser the comments are categorised to negative, positive and neutral comments. Subsequently, for every month the negative and positive comments are totalled, neutral comments are filtered and the aforementioned PNratio for every month is computed.

5. Analysis & Interpretation

5.1 Sample Characteristics

After extracting and summarising the data it was prepared for the statistical analysis. The data file is imported into the statistical analysis program SPSS which this research will use. In this research car sales is the dependent variable and Google Trends and PNratio are the independent variables of interest, with price and time as moderators. Descriptive statistics, skewness, kurtosis and normality tests were computed for the dependent variable. The measurement unit is car brand per week, this makes up for a sample of 15 car brands x 48 weeks (N = 720). The 48 week period matches with the maximum available time that YouTube holds the data available. The mean and standard deviation are $41,158.20 and $26,493 and there are no missing values in the dataset. The sample is equally divided into 7 low-priced car brands (n = 336) and 8 high-priced car brands (n = 384). In line with earlier research $35,000 is used as dividing line between low and high priced models (Voortman, 2015; Wijnhoven & Plant 2017). In addition another dummy variable is created for the different years (2017 and 2018). A Shapiro-Wilk’s test (p > .05), a Kolmogorov Smirnov test (p > .05) and a visual inspection of the histogram, normal Q-Q plots and box plots show that the sales numbers are not normally distributed. The skewness of the sample is .555 (SE = . 091) with the kurtosis being -.228 (.SE = .182). This is explained by the combination of sales number of exclusive high priced cars and high volume production car models.

5.2 Development of Sales

When looking at the different sales numbers in figure 3, the car brands differ structurally on the level of sales. In the figure, it can be seen that the best-selling brand (Chevrolet Silverado) has a level of 45.000 higher than the least selling brands at the bottom (BMW, Mercedes,

Porsche). This means that there is diversity in the sample with respect to sales numbers which relates to the exclusive high priced cars and high volume production car models. The diversity makes that there is variation in the dependent variable to explain. The consequence is that the sales numbers are not normally distributed which makes that there is a variation to explain in the analysis. In line with the diversity of sales, it can be seen that sales numbers are not normally distributed (p < .050). Nevertheless, the skewness and kurtosis are still under the threshold values of 3 and 10 that are commonly used to diagnose the degree of skewness.

"

5.3 Independent Samples T-test - Influence of Price

An independent-samples t-test was conducted to research whether the lower priced group scores significantly different on average than the higher priced group. This is done in order to find initial evidence for hypothesis 3, which states that higher priced car models show a weaker correlation to social media content. The p-values in table 3 were determined on either the basis of that the variances were assumed to be equal or not. The Levene’s test was used to determine the most appropriate choice in that respect. The T-test is tested for the variables Sales, Google Trends, PNratio, Total Comments, Positive Comments and Negative Comments. 0 10000 20000 30000 40000 50000 60000 70000 80000

Jun-17 Jul-17 Aug-17 Sep-17 Oct-17 Nov-17 Dec-17 Jan-18 Feb-18 Mar-18 Apr-18

Sales Selected Brands

Volvo XC90 Honda Civic Honda Accord Ford Fusion Ford Escape Chevrolette Corvette BMW i3 Toyota RAV4 Toyota Camry Porsche 911 Hyundai Elantra Mercedes AMG GT Chevrolet Silverado Ford Explorer Jeep Grand Cherokee

There was a significant difference in the means for Sales (t = 14,3, p<.001), Google Trends (t = 2.157, p = .031) Comments (t = 2.694, p = .007) and Positive Comments (t = 2.527, p = .012). These results suggest that price has an effect on the aforementioned variables. Specifically, our results suggest that sales are significantly higher in the lower priced group, which makes sense. In addition the low priced group shows significantly more comments, more positive comments and have a higher score on Google Trends. These results are initial traces of evidence for hypothesis 3. The results do not directly answer whether higher priced cars show weaker correlation values but do suggest that lower priced cars have a higher Google Trends score and attract more (positive) comments which would higher the PNratio.

Table 3: . T – Test with Means (M), Standard Deviation (SD), T-score and P-value

Variables Group N M SD T Score P value

Sales Group 1 336 25,6 E3 8,1 E3 14.3 < .001 Group 2 384 11,7 E3 16,8 E3 Google Trends Group 1 336 81.23 9.620 2.157 < .05 Group 2 384 79.34 13.73 Ratio Group 1 336 1.95 1,04 1.811 .071 Group 2 384 1.79 1,33 Comments Group 1 336 4430.57 3125.08 14.3 < .05 Group 2 384 3853.50 2540.63 Positive Group 1 336 98.87 126.64 14.3 < .05 Group 2 384 74.85 127.92 Negative Group 1 336 66.84 98.58 14.3 .296 Group 2 384 57.41 137.25

note N=720. Price was coded as 0 = low and 1 = high The lower priced group (1) included all cars below $35,000 and the higher group (2) included all cars above $35,000.

5.4 Regression Analysis

A bivariate analysis is conducted to research the separate correlations between the dependent and independent variables. In this research sales is the dependent variable and Google Trends and PNratio are the subjects of interest as independent variables.

The results in table 4 show that PNratio has a significant negative correlation with sales. This implies that a higher PNratio would lead to lower sales. This is in the opposite direction of the proposed research framework and proposed hypothesis 2. The effect is weak but significant (r = -.102, p < .010). Moreover, the sales for 2018 are significantly lower than 2017 (within a one-year period, so it might be a seasonal effect) and Google Trends has no significant relation with sales or the PNratio. PNratio on its turn has a stronger effect than the effect of year. In addition it is shown that price has a strong and significant effect on sales, which will be further researched in a deeper analysis with the usage of a dummy variable. The direction of the effect of Google Trends is in line with hypothesis 1, although the effect is insignificant and weak. At the highest level of correlation analysis hypothesis 1 and 2 are not supported: higher search queries index and more positive sentiment do not lead to higher sales.

Table 4: Means (M), Standard Deviations (SD) and Pearson correlations all variables

Variable M SD 1 2 3 4 5 1. Sales 18,2 E3 15,2 E3 1 2. Google Trends 80.22 12.015 .058 1 3. PNratio 1.86 1.21 -.102** .041 1 4. Year Dummy 0.35 0.48 -.082* .037 .031 1 5. Price 41,1 E3 26,4 E3 -.564** -.426** -.071 .000 1 Note. N = 720. Year was coded as 0 = 2017 and 1 = 2018. Price was coded as 0 = low and 1 = high * p < .05 ** p < .01 *** p < .001 (two-tailed)

When we split the dataset for price with the help of the dummy variable, a significant effect of Google Trends is detected for the low priced segment in comparison to no effect for the high priced segment. This would be the first evidence for a moderating effect of price. Looking at the results in table 5 we notice that higher priced cars show less correlation between social media content than lower priced cars. This is in line with the third hypothesis which states that more expensive car models show a weaker correlation between social media content and sales. Both on the highest level of analysis and at the price-split level it is noticeable that sentiment (PNratio) has a more substantial significant effect on sales than Google Trends. This is in line with the fourth hypothesis, which states that sentiment has a stronger correlation with sales than Google trends with sales. However, the effect is in the opposite direction of the proposed research model.

If we look further into the lower priced car models (table 6) we notice that there are significant results for the Hyundai Elantra, Honda Civic and Ford Escape. For the Hyundai Elantra we discover that Google Trends shows a negative significant correlation (r = -0,380, p

Table 5: Means (M), Standard Deviations (SD) and Pearson correlations dummy_price Dummy Price Variable M SD 1 2 3 4 Low (n=336) 1. Sales 25,6 E3 8,2 E3 1 2. Google Trends 81.23 9.620 -.201** 1 3. PNratio 1.95 1.04 -.356** -.099 1 4. Year Dummy 0.35 0.48 -.246* .060 .081 1 High(n=384) 1. Sales 11,6 E3 16,8 E3 1 2. Google Trends 79.34 13.726 .093 1 3. PNratio 1.78 1.33 -.086 .101* 1 4. Year Dummy 0.35 0.48 -.034 .025 -.002 1

Note. N = 720. Year was coded as 0 = 2017 and 1 = 2018. Price was coded as 0 = low and 1 = high 
 * p < .05 ** p < .01 *** p < .001 (two-tailed)

< .010) whereas the Honda Civic shows a positive correlation (r = 0.501, p < .050) just as the Ford Escape (r = 0.345, p < .050). Only the Ford Escape shows a significant correlation between PNratio and sales (r = -0.363, p < .050), although being opposite to the direction of the proposed researched model. The results vary to a great extent in effect and significance showing no linear trend.

5.5 Multiple Regression

A multiple linear regression was conducted in order to investigate what the strength of the proposed models is to predict sales based on price, year, Google Trends, PNratio and several moderations (Price x Ratio, Price x Trends and Trends x Ratio). Though no multicollinearity

Table 6: Means (M), Standard Deviations (SD) and Pearson correlations lower priced cars

Brand Low Variable M SD 1 2 3 4

Hyundai Elantra 1. Sales 15,6 E3 2,5 E3 1 2. Google Trends 86.13 6.135 -.380** 1 3. Ratio 2.56 1.38 -.029 .124 1 4. Year Dummy 0.35 0.48 -.141 .164 .121 1

Honda Civic 1. Sales 31,2 E3 3,9 E3 1 2. Google

Trends 84.73 7.010 .501** 1

3. Ratio 1.33 1.34 -.103 -.086 1

4. Year Dummy 0.35 0.48 -.676** -.260 -.028 1 Ford Escape 1. Sales 24,3 E3 2,79 E3 1

2. Google

Trends 89.94 5.847 .345* 1

3. Ratio 1.73 1.17 -.363* -.056 1

4. Year Dummy 0.35 0.48 -.585** .023 -.004 1

Note. N = 720. Year was coded as 0 = 2017 and 1 = 2018. Price was coded as 0 = low and 1 = high 
 * p < .05 ** p < .01 two-tailed)

existed (VIF < 2), there were two violations of the assumptions for a multiple regression. In line with the non-normality of the sales numbers, the residuals are to some extent not normally distributed. There is also some heteroscedasticity present (see the histogram and p-plot in the appendix). Because there are no extreme violations, the violations are taking into account in the discussion as a limitation to the validity of generalising the model. We start our first model with price and year. We add the predictors Google Trends and PNratio in the second model and conclude by looking for moderations. The results are presented in table 7. The results show that all three models are significant and explain parts of the dependent variable. The results of the regression indicated that the first model explained 21.3% of the variance and that the model was a significant predictor of sales, F(2,717) = 98.128, p = .000. Both Price dummy contributed significantly to the model (β = -.456, p < .001) as well as Year Dummy (β = -.082, p=.014). The second model inserted Google Trends and PNratio. The model explained 22.9% of the variance and the model was a significant predictor of Sales, F(4,715) = 54,295, p = .000. In the model Price Dummy (β = -.463, p < .0.001), Year Dummy (β = .079, p < .05) and Ratio (β = -.132, p < .001) contributed significantly to the model. It is noticeable that the effect of price is three times as strong as the effect of PNratio. The last and third model added several moderators to add explanatory power. The model explained 24.4% of the variance and the model was a significant predictor of Sales, F(8,711) = 34.099, p = .000. In the model the moderation between Price and Trends (β = 0.873, p < . 0.001) contributed significantly to the model.

The research proceeded towards a moderation analysis to determine whether the size of the effect of an independent variable on a dependent variable interacts with a moderator. The research discovered that there is a significant interaction between Price and Trends (figure 4). When the car model falls in the low priced category, a higher trends score leads to lower sales numbers. This conflicts with our first hypothesis which states that a higher Google Trends score should result in higher sales, no matter in which category the car is in. Contrary, car models in the high priced category with a higher Google Trends score leads to higher sales numbers, which strokes with the first hypothesis. Nonetheless, this conflicts with hypothesis 3 which states that higher priced cars should have a weaker correlation between the independent and dependent variable than lower priced cars. Concluding, we can say that the effect of price is negative for the lower priced car models and positive to neutral for the higher priced cars.

Table 7: Multiple Regression Analysis with Sales as Dependent Variable

Model 1 Model2 Model3

Estimates B SE β p B SE β p B SE β p Price dummy -13,9 E3 1007.95 .456*** .000 -14 E3 1002.88 .463*** .000 -44,2 E3 7637.03 -1.453*** .000 Year dummy -2602,2 1051,43 .082* .014 -2507,04 1041,94 .079* .016 -2480,7 1046,08 -.078* .018 Google Trends 48.295 42.235 .031 .354 -292,62 119,067 -.231** .014 Ratio -1663.24 8 412.340 -.132*** .000 -6915,6 3872,32 -.551 .075 Moderation Price & trends 335,78 90,99 .902*** .000 Moderation price & ratio 1513,12 873,12 .132 .084 Moderation trends & ratio 48,613 45,868 .341 .290 R-square .215 .233 .251 Adjusted R-square .213 .229 .244 F-value 98,128 54,295 34,099 Note. N = 720. * p < .05 ** p < .01 *** p < .001 (two-tailed)

"

5.6 Time Series

In order to test the influence of time the cross-correlation function in SPSS was used. This is performed for both independent variables PNratio and the Google Trends on the dependent variable Sales. All car models were separately researched by breaking up the data file into small subsets. The total number of cases is smaller in this subset since we aggregated the data to car brand per month (15 x 11, N =165) The maximum time lag was 11 months based on literature on time lags in the consumer purchase process for car buyers and data availability (Kandaswami & Tiwar, 2014; Putsis & Srinivasan, 1994). The cross-correlation function finds the strongest correlation between the dependent and independent variables by introducing a time lag. Due to the direction of the research model only time lags where the independent variable predates the dependent variable were included. The numbers in bold (table 8) indicate whether the correlation is significant. The results in the table show that when a time lag is introduced the correlation becomes stronger, but that it only is significant for a small number of car models. The fifth hypothesis, which states that the correlation will increase with the introduction of a time lag is supported for specific car models. The introduction of a

time lag yields more significant results for Google Trends than the sentiment mined from YouTube as reflected in the PNratio. This result is in conflict with hypothesis 4 which states that sentiment has a stronger correlation with than Google trends with car sales. The time lag for the PNratio only shows significant results with the imposition of a time lag of 1 month. The Ford Explorer shows a significant result with a time lag of 0 months which cannot be interpreted as a time lag. The time lags between Google Trends and sales range from 1 to 5 months. The results support that for different car models different consumer behaviour patterns might be applicable. Moreover, the results show that the inclusion of a time lag levers more significant correlations for the lower priced car models (with the exception of the BMW i3). This is additional support for hypothesis 3 which states that higher priced car models show a weaker correlation between social media content and sales.

Table 8: Best fitting Pearson’s correlations based on the influence of time

PNratio & Sales Google Trends & Sales

Model Lag Correlation Lag Correlation

BMW i3 1 .801** 1 .646** Chevrolet Corvette 4 .322 4 .322 Chevrolet Silverado 1 .529 5 .225 Ford Escape 4 .051 0 .431 Ford Explorer 0 .759** 5 .582* Ford Fusion 3 .422 2 .608* Honda Accord 1 .744** 2 .600* Honda Civic 2 .283 1 .857** Hyundai Elantra 2 .409 4 .364 Jeep Grand Cherokee 1 .042 8 .468 Mercedes AMG GT 8 .485 1 .120 Porsche 911 6 .102 5 .194 Toyota Camry 1 .358 3 .542 Model

Toyota RAV4 3 .473 2 .382

Volvo XC90 1 .611* 5 .595*

Note. N = 165. * p < .05 ** p < .01 *** p < .001 (two-tailed). Lag is in months.

PNratio & Sales Google Trends & Sales

Lag Correlation Lag Correlation

6. Discussion

This chapter will elaborate on the results of this research. A general summary and discussion on the central research question, hypotheses and results will be presented. In addition the theoretical and practical implications of the research are discussed. The primary purpose of this research was to research how car sales can be explained and predicted by using Google Trends data, YouTube data and sentiment analysis. In doing so this research intended to answer the central research question “What is the explanatory power of Google Trends and sentiment analysis mined from Youtube comments for car sales in the US?”. The central research question was researched by seeking for the most significant and relevant correlations between Google Trends scores and sentiment as reflected in the PNratio.

6.1 Key Findings

Five hypotheses were tested in order to test the relationship between Google Search volumes, YouTube sentiment and car sales. The main finding of the research was that the direction of the correlations did not fit the intended research models. At the highest level of analysis it was found that there are no significant correlations between Google Trends scores and car sales. The direction of the effect was in line with the proposed research framework but the effect was insignificant. When the dataset was split for price, a significant effect of Google Trends was found for the lower priced category but in the opposite direction of the intended research framework. The effect of Google Trends scores was only significant and positive for 2 car models. Therefore hypothesis 1 is rejected for 13 car models and supported for 2 car models (Honda Civic and Ford Escape). Furthermore, the sentiment as reflected in the PNratio gave weak results. At the highest level of analysis the effect was significant but weak and in the opposite direction of the intended research framework. When the dataset was split for price the lower priced category showed more significant results, although being in the opposite