用消費者行為改進銷售預測 - 政大學術集成

全文

(1)國立政治大學應用經濟與社會發展英語碩士學位學程英語碩士學位學程 International Master’s Program of Applied Economics and Social Development College of Social Sciences National Chengchi University. 政治碩士論文. 大. 立Master’s Thesis. ‧. ‧ 國. 學. 用消費者行為改進銷售預測. y. Nat. er. io. sit. Improved Sales Forecasting with Consumer Behavior. al. n. v. Student: 馬克斯 n i Muehlen C h MaximilianUzur e n Tsoyu g c h i Calvin Lin Advisor: 林左裕. 中華民國 106 年 6 月 2017 June.

(2) Abstract Purpose – The role of forecasting in a lean enterprise is immense. It is crucial for car manufacturers to have reliable information about the future to make important decisions and stay competitive. Developing markets and consumers provide new types of data that demand modern approaches to be handled. This paper aims to create reliable forecasting models through integration of Google Trends data from 2008 to 2016.. Design/methodology/approach – Building on the 5-stage-model of consumer. 政治大. buying behavior, the study identifies suitable Google keywords for this process.. 立. Autoregressive distributed lag models are used to examine the relationship between. ‧ 國. 學. sales and macro-economic variables as well as Google Trends. Predicted sales are used to test for accuracy.. ‧. Findings – Two most common evaluation measurements for forecasting accuracy suggest the use of Google Trends, as predictors for future sales, is outstanding. The. y. Nat. sit. finding concludes that macro-economic variables and seasonality are not as valuable as. er. io. Google Trends in short-term, up to one year, forecasting.. n. Value – Only little researcha on i v Google Trends and their l car sales forecasting takes. Ch. n engchi U. appropriate time lags into account. This analysis provides new insights into the linkage of consumer behavior and sales data.. 關鍵詞 – ADL, 銷售預測, Google 趨勢，消費者行為 Keywords – ADL, Sales Forecasting, Google Trends, Consumer Behavior. I.

(3) Table of Contents. Background ........................................................................................................ 1. 1.2. Statement of the Problem .................................................................................. 2. 1.3. Research Goal ................................................................................................... 3. Literature Review ..................................................................................................... 4 2.1. Decision Model for Purchases – Five-Stage Model............................................ 4. 2.2. Google Trends as a Source of Consumer Behavior ........................................... 8. 2.3. Empirical Findings ............................................................................................ 11. 立. Methodology .......................................................................................................... 13 3.1. Conceptual Model ............................................................................................ 13. 3.1.2. Autoregressive Distributed Lag Modeling .................................................. 15. y. sit. Data Collection and Scope............................................................................... 15. io. 3.2.1. al. Dependent Variable ................................................................................... 15. 3.2.3. Google Trends Keyword Selection ............................................................ 17. 3.2.4. Variable Correlation ................................................................................... 18. n. 3.2.2. iv n C Macro Data Variables ................................................................................ 16 hengchi U. Empirical Model...................................................................................................... 20 4.1. Adjusted R² and Akaike Information Criterion .................................................. 20. 4.1.1 4.2. Testing Residuals ...................................................................................... 24. Forecasting ...................................................................................................... 25. 4.2.1 5. ‧. Time Lag.................................................................................................... 13. Nat. 3.1.1. 3.2. 4. 政治大. 學. 3. 1.1. er. 2. Introduction .............................................................................................................. 1. ‧ 國. 1. Evaluation Measurements ......................................................................... 27. Results ................................................................................................................... 28 II.

(4) 6. 5.1. Data Analysis and Results of Empirical Analysis ............................................. 28. 5.2. Key Findings .................................................................................................... 31. Conclusion ............................................................................................................. 33 6.1. Discussion ........................................................................................................ 33. 6.2. Practical Application and Limitations ................................................................ 34. 6.2.1. Implementation in the Automotive Industry ................................................ 34. 6.2.2. Major Implications for Sales....................................................................... 34. 6.3. Limitation of this Research ............................................................................... 36. 政治大. References .................................................................................................................... 37. 立. ‧. ‧ 國. 學. n. er. io. sit. y. Nat. al. Ch. engchi. III. i n U. v.

(5) 1 Introduction 1.1 Background Globalization has created new markets over the last decades. The development and properties of those new markets are hard to capture, document and eventually to apply sufficient research on. The Tiger Cubs, Thailand, Malaysia, Indonesia, Vietnam and Philippines are part of the mentioned markets. This research is focusing on the. 政治大 driven by demand and increasing wealth, is quickly growing and penetrating several 立 countries of different development status. It is of high interest to optimize all kind of automotive industry in one particular Tiger Cub, Thailand. The automotive industry,. ‧ 國. 學. processes in those lucrative markets, to further boost profit and competitiveness. The low motorization rate and GDP growth of the last years show the car industry’s. ‧. potential for selected Tiger Cubs. Thailand has been chosen because it is the car. y. Nat. industry hub of South East Asia, with an excellent geographical location and many. io. sit. suppliers’ and car manufacturers’ factories set up there. Furthermore the brand that is. n. al. er. subject of this work, Mercedes-Benz, has a proven and positive development in. i n U. v. Thailand. Their presence in Thailand, in terms of sales volume, is the largest within the Tiger Cubs.. Ch. engchi. Table 1.1: Comparison of THA, IDN and MYS (The World Bank) (International Organization of Motor Vehicle Manufacturers). Population in Motorization Rate in 2015. Avg. GDP Growth of. 2014. the last 5 years. Thailand. 67,959,000. 232. 2.7%. Indonesia. 257,564,000. 83. 4.1%. Malaysia. 30,331,000. 405. 2.8%. 1.

(6) 1.2 Statement of the Problem To stay competitive, in terms of costs and supply, car manufacturers have to rely on demand forecasting. Knowing future sales is crucial for just-in-time deliveries and minimized storage costs while simultaneously satisfying the demand. Reducing costs and making the production as lean as possible are some of the biggest challenges of today’s car manufacturing industries (Karlsson & Annie, 2014). In addition, the supply chains are long and short time notice creates mistrust towards OEMs. This leads to suppliers increasing their price to hedge against the risk of short time changes (e.g. 08/09 during the financial crisis).. 政治大 are not state of art and not立 improving within the last 15 years (Rieg, 2010). Relying. Forecasting models to predict car sales are widely used in the industry, however they. ‧ 國. 學. solely on macro-economic data explains changes, is able to give long time trends but cannot catch up with the increasing haste of today’s life. The era of the internet has reached almost every country and especially emerging countries are becoming. ‧. immense future markets for almost every industry. To keep updated and ahead of the. y. Nat. competition it is necessary to adapt to new environments as fast as possible – survival. sit. of the fittest. The internet, especially Google, is a new way to obtain precious consumer. al. n. 2012).. er. io. data. Some industries already tested out Google Trends to predict demand (Pan. B.,. Ch. engchi. 2. i n U. v.

(7) 1.3 Research Goal The goal of this work is to find a modern forecasting approach that outperforms traditional approaches. To achieve this goal it is necessary to step back and consider what has been changed in the last decades that could make a forecasting model outdated? Among many changes there is one significant change, the role that the internet plays in a consumer’s purchasing behavior. With increased access to the internet, huge amounts of information, ways to communicate to people with the same interest and. 治政大 It is a public web facility, capture consumer behavior is through Google Trends. 立 launched in May 11, 2006 by Google Inc., which provides information about the relative gather experiences, the internet is a game changer in consumer behavior. One way to. 學. ‧ 國. volume of a search term.. A solid base to build on is made of existing forecasting models that are based on. ‧. macro-economic variables. To increase the accuracy of said models, the trend index of searched (googled) keywords will be implemented into the traditional forecasting model.. Nat. sit. y. It will be built up and executed for Mercedes-Benz in the Thai market. The approach of. io. er. using Google Trends and macro-economic data can be applicable to other, comparable, countries and products as well. The other Tiger Cubs, Indonesia, Malaysia, Vietnam. n. al. i n U. v. and the Philippines would be the first choice to apply this approach on, as well as close. Ch. engchi. competitors to Mercedes-Benz such as BMW and Lexus. The advanced models enable to stay ahead of the competition through the saving of avoidable and demonstrate state of art techniques not only inside of a car.. 3.

(8) 2 Literature Review This chapter introduces the theoretical structure of this work. The customer’s purchase decision plays a major role in all of the statistical approaches. The following section presents a well-known purchase decision model. The new information source Google Trends will be introduced detailed in this chapter. Google Trends provides a tool that does not only explain present consumer behavior, it also can predict future collective behavior (Goel, 2010).. 2.1 Decision Model for Purchases – Five-Stage Model. 立. 政治大. One of the main theories that are supporting the decision of using Google Trends is the. ‧ 國. 學. Five-Stage model of the consumer purchase process which describes the consumption behavior in defined stages. This model plays a major role in consumer behavior and has been previously used in the car industry context (Shende, 2014). The stages are. ‧. problem recognition, information search, evaluation of alternatives, purchase decision. y. Nat. and post purchase behavior. Not all consumers will go through the whole process, due. sit. to differences in products, applications, behaviors and costs. This model however. er. io. provides a good overall frame because it shows a full range of considerations that a. al. n. iv n C different steps within the Five-Stagehmodel, there willUbe an example that is related to engchi the substance of this research.. consumer will face when making a new purchase. Next to giving an explanation of the. Problem Recognition The whole process starts with the consumer realizing a problem or need for some purchasable good; it may have been triggered internally or externally. An internal stimulus would be the waste of time to commute to work, the uncomfortable circumstance in public transportation or the lack of freedom and independency relying on public transportation. An external stimulus would be advertising which is triggering the need for a certain product, in this case a car. It could rely solely on convenience and need in the technical sense or as an item that provides certain utility. Those kinds of 4.

(9) utility factors can be increased for some people with purchasing premium products (Kotler & Keller, 2016).. 立. 政治大. ‧. ‧ 國. 學. n. er. io. sit. y. Nat. al. Ch. engchi. i n U. v. Figure 2.1 Five-Stage Model of the Consumer Buying Process (Kotler & Keller, 2016). Information Search There are two levels of engagement in information search. One is heightened attention, which is considered a milder search state. The other one is active information search that includes looking for reading material, online research or visiting stores to learn about the product. Typical information sources for consumers are personal (family, friends), public (social media), commercials (salespeople, displays, packaging, websites and ads) and experiences that come from handling a certain product. Sources that are 5.

(10) Figure 2.2 Successive Set Involved in Consumer Decision Making (Kotler, et al., 2016). reaching the consumer most effective are personal, experiential and public sources that are known for being independent authorities. (Kotler & Keller, 2016) The process of information search is crucial for this work in order for the connection with Google Trends. Nowadays people, that do not have the access to experiential or personal information sources, rely on outside sources. Those have to be seen critical in behalf of distinguishing between an ad and an objective report. Talking to salespeople. 政治大. or getting information from commercials seem to be strongly biased since the instances involved have a direct profit from giving out positive information rather than negative. 立. information. Therefore the internet is a platform that offers various sources as well as. ‧ 國. 學. the direct exchange of experiences with other consumers. Most people realized this and are using the accessible information online. Nonetheless it has to be seen critical by virtue of internet sources that might be highly biased.. ‧. The first approach to find solutions for an existing problem will create a larger set that. y. Nat. contains different brands or products for a need. With further information gathering the. er. io. Evaluation of Alternatives. sit. sets will gradually decrease up to one product that fits the consumer’s needs the best.. al. n. iv n C The consumer has to compare competitive to make a final value judgment. h e n g brands i U h c Taking a look at the concept helps to understand this process. The consumer is trying to satisfy a need and is also looking for certain benefits. With different brands, the consumer sees bundles with varying abilities to deliver the benefits (Kotler & Keller, 2016). Price, quality, pleasure, social status, reliability, maintaining costs, safety are some of the factors that altogether create the utility a consumer gets for a certain product. The weighting of such factors varies among consumers nationally and internationally. Also the choice of factors will vary among countries, cultures, geographical locations etc. Imagining consumer A, who is looking for a car that brings him to work. He lives in Canada in an area with low density and long commuting routes. Consumer B lives in a 6.

(11) major city in South-East-Asia and needs a car to commute to work as well. Both consumers have the same “problem”, but they will rate their alternatives different. Consumer A might prefer bigger cars, which are running comfortably even for longer rides. It should have an installed heating system and probably a smooth suspension system combined with a larger engine with a 4 wheel-drive, to have a reliable vehicle even in winter. Consumer B on the other hand might not want any of those attributes in the first place. Small and convenient with a small turning circle to fit the parking situation in mega cities, automatic transmission to survive traffic jams without a cramp and low emissions to not fear any penalties within city areas. This simplified example of two different consumers shows how hard it is to generalize the utility function for consumers.. 政治大. Almost every attribute of a car, even the outside design, is varying among consumers.. 立. Purchase Decision. ‧ 國. 學. During the previous stage the consumer collects preferences and forms a choice set among brands that may also lead to an intention of buying the most preferred brand. An. ‧. important factor that can intervene between purchase decision and purchase intention is the attitude of others and unanticipated situational factors. Another person’s attitude. y. Nat. sit. might influence a consumer’s purchase decision heavily. Intensity of another person’s. er. io. negative attitude toward our preferred product and the consumer’s own motivation to comply with other person’s wishes are giving the magnitude. Unanticipated situational. n. al. Ch. i n U. v. factors such as losing a job or personal dislike of a sales person can erupt to change. engchi. the purchase intention as well. Risk aversion or risk loving will also influence the purchase decision. Not performing to expectations, social risks such as embarrassment in front of others, high maintenance costs, financial burden and other uncertainties are factors that are leading to the perceived risk of a purchase decision (Kotler & Keller, 2016). In the case of a car purchase all the mentioned factors under “Purchase Decision” are significant. Since cars are high-value goods, consumers are putting more time into the purchase decision and usually do not buy out of a shopping spree. With the internet it is enormously easy to get the opinion and experience of other people and groups about the preferred product. Those are much easier to obtain than many years ago. 7.

(12) Nowadays there are high amounts of reviews, tests and opinions about any car model so that consumers are much more likely to get influenced on the way to their purchase. Post purchase Behavior This is a crucial point for any product. A satisfied consumer is more likely to purchase the product again and also tends to give out positive referrals of that product or the product’s brand. Personal information sources are rated highly effective and therefore it is from highest importance to keep the consumer happy and satisfied even after the purchase. Dissatisfied consumer may not only abandon or return a product, they can also take public action or warning other consumers (Kotler & Keller, 2016). Bad news. 政治大 stronger than the satisfaction a product might give. Some consumers will explore 立 dislikes of a product of attribute that they have not recognized earlier. With the internet. usually spread wider than positive news due to the psychological fact that fear is. ‧ 國. 學. those information spread much faster, such as the recent cases of Samsung Galaxy Note 7 phones that are exploding. Even though this is a production error, it is now. ‧. known by almost every single person that is interested in buying smartphones from any brand. Those “small” mistakes in the construction cost Samsung’s smartphone sector. y. Nat. er. io. safe phones.. sit. customers and reputation, even after years and several product lines of successful and. al. n. iv n C consumer as long as he owns a certain h e nvehicle. i Usmall disadvantages that do not g c hAny In conclusion, with the background of a car purchase, it is important to further satisfy the. occur at the competition have huge impacts on the overall reputation of the brand and future consumer purchase decisions.. 2.2 Google Trends as a Source of Consumer Behavior Existing forecasting models usually combine survey based indicators and macroeconomic indicators. That way they can cover up economic and psychological aspects of consumer behavior. Problems and inaccuracies that come along with survey data are plenty, such as a sampling bias. In previous research remarkable results have been. 8.

(13) found, with a lack of time dimension. Out-of-sample forecasting experiments with that data had found of not significant value (Schmidt & Simeon, 2009). The Google Trends data is unbiased sample of Google search data. It is anonymized, categorized and aggregated. It allows to measure interest in a particular topic across search, from around the globe, down to the city-level. Real time trends data is a random sample of searches from the last 7 days and not subject of this work. The full Google dataset is non-real time and starts from 2004. The relative search interest of a topic, which can also be a specific set of search terms, is compared to itself, over time and where it is most-searched (Rogers, 2016).. 政治大 developed countries such as the U.S, where a broad range of data and observations 立 are available. Most researchers use this data-rich environment to conduct their research It is mentionable that most of the existing forecasting models have been conducted for. ‧ 國. 學. and prove/disprove theories. There is no doubt that those results are useful, also for other countries. However, it has to be realized that models cannot easily be generalized. ‧. and simultaneously maintaining a healthy amount of validity. Some of the most prosperous markets are the hardest to analyze, to obtain valid data and to request for. y. Nat. sit. further data. This is the point where Google Trends becomes handier. With a few. er. io. exceptions, most countries use Google as their primary tool for searching the internet. Globally the market share of Google’s search engine is 79.8% (Seach Engine Journal,. n. al. 2016).. Ch. engchi. i n U. v. In Thailand however the market share is 98%, which furthermore shows the significance and convenience of Google data in this country (Return On Now, 2015). The data created through searching, also called googling, gets delivered directly to Google, unfiltered by authorities. The data becomes publicly available and gives deeper insights of the internet searching behavior of people using Google. Internet access, which includes access to Google, is increasing. That makes working with Google Trends a modern and promising way of obtaining consumer data.. 9.

(14) A good approach to look at Google Trends data in relation with traditional data shows the following: Macro-economic variables indicate the consumer’s ability to purchase a good. Google Trends is providing a measure for the consumer’s preparatory steps to the purchase decision making.. How to use Google Trends? The tool is modern so that a large amount of people are not familiar with using it. For. 政治大 most important features are explained in the following: 立. this reason there is an introduction of how to utilize Google Trends to obtain data. The. ‧ 國. 學. Search Term is/are the keyword/s that consumers have entered into Google search. Several search terms can be used at the same time and different language inputs are. ‧. supported.. Country Selection provides all the countries where Google is being used.. y. Nat. io. sit. Time Range allows the user to choose between various options. Daily data provides. n. al. er. the results for every minute; Weekly data provides hourly results; Custom Range shows. i n U. v. weekly or monthly data, which depends on how far the time period ranges.. Ch. engchi. Category is responsible to create the index score. A category should be included to avoid ambiguities of search terms. The category used in this work is “Autos & Vehicles”. Type of Search allows the user to choose between “Web search”, “Image search”, “News search”, “Google shopping” and “Youtube search”. Furthermore the user is able to divide the country into sub regions and display related topics for the used search term. The output is presented via diagram and is exportable as a CSV file.. 10.

(15) 2.3 Empirical Findings The benchmark model that is used in this work originated from M. Hülsmann et. al. (2012). They used different methodology while maintaining a logical explainable variable selection that consists of macro-variables. Among their models or methods were Ordinary Least Squares, Quantile Regression, Support Vector Machine, Decision Trees, K-nearest-Neighbor and Random Forest. The variables were picked under certain criteria, which will be explained in 3.2.2. An important finding of this work is the importance of considering special effects, such as the car-scrap bonus in Germany and the irregular behavior of the US-American market. Since this thesis is mainly about the. 政治大 be dealt with in the discussion 立(Hülsmann, Borscheid, Friedrich, & Reith, 2012).. improvement of a forecasting model by Google Trends, estimates of special events will. 學. ‧ 國. The work of Choi & Varian (2012) is encouraging future researchers to engage with Google Trends as a complimentary tool for forecasting. Examples that are dealing with. ‧. simple seasonal AR models that include relevant Google Trends, in various fields. Models about Motor vehicles and parts, initial claims for unemployment benefits and. Nat. sit. y. consumer confidence that are including Google Trends variables tend to outperform. io. er. models, which exclude Google Trends, by 5% to 20%.. A seasonal AR-1 model is used as a baseline for car sales. The model can explain. n. al. i n U. v. 71.11% of the data variation (adj. R²). The next step of Choi & Varian’s (2012) work is to. Ch. engchi. include Google Trends categories in the regression. The in-sample fit can be significantly increased by doing so, reaching 80.8% adj. R². The execution of an out-ofsample forecast shows a 10% increase in forecasting accuracy when using Google Trends, measured by the Mean Absolute Error (MAE) during normal economic period (Choi & Varian, 2012) However, the research concentrates more on “contemporaneous forecasting” that predicts the present. The time difference between obtaining Google Trends and the car sales, accounts for only about 4 to 6 weeks. They used the search results from the first Saturday to Sunday, each month. The sales data was obtained 2 weeks after the end of each month. According to the reasoning of Kinski (2016) and the consumer purchase. 11.

(16) theory of Kotler & Keller (2016), the work does not accommodate appropriate time lags for Google Trends. The research of Fantazzini & Toktamysova (2015) found out that on average parsimonious bivariate models that only include Google Trends outperform more complex Bayesian VAR- and VEC models. Furthermore they found out that Google Trends-based models perform better in recessions. In addition they had better forecasting performances, with higher robustness, in various business cycles (Fantazzini & Toktamysova, 2015). The importance of a time lag has been discussed in detail in Kinski’s (2016) work. A. 政治大 Trends. In his findings he proved with a cross-correlation analysis that 33 out of 48 car 立 models show the need for significant time lags, which will increase the prediction part dealt with the importance of time lags in forecasting models which deal with Google. ‧ 國. 學. accuracy. The average adj. R² improvement of applying a time lag is 18.25%, according to his studies. Furthermore he showed the differences in optimal lag times vary among. ‧. cultures. The average time lag for car models in the U.S. is 4.04 months whereas the average time lag for Germany is only 2.33. He outlines the possibilities that can be. y. Nat. n. al. er. io. 2016).. sit. achieved with Google Trends and optimized time lags in terms of forecasting (Kinski,. Ch. engchi. 12. i n U. v.

(17) 3 Methodology 3.1 Conceptual Model It seems logical that many people consider a complex forecasting model to be more appropriate to solve a complex problem such as forecasting. People can be reluctant to believe simple forecasting solutions and often it is incentivized to provide model with high complexity to publish in highly ranked journals (Wharton, 2011). According to the definition for simple forecasting models of Green & Armstrong (2015) the model used in. 治政 Their studies reveal 大 that. this research is rather simple than complex. It is easily understandable by persons that are familiar with forecasting.. 立. complexity increases the. forecasting error by 27 percent on average, found in 25 papers with quantitative. ‧ 國. 學. comparisons. In addition to the probably better forecasting results there are other factors that show superiority of simple models (Green & Armstrong, 2015).. ‧. Models that are intended to be used in commercial forecasting often need to be approved by people who might not be experts in the field of forecasting and statistics.. Nat. sit. y. Presenting complex causalities and procedures will not achieve acknowledgment as. io. er. desired; it can turn out to cause confusion for the person who has to be persuaded. Not understanding a forecasting model correctly will not only quickly decrease the. n. al. i n U. v. audience’s interest, it will also often lead to rejection of that model/project. Furthermore. Ch. engchi. it can be more difficult to interact with experts of other fields that may give precious advice to more understandable models.. 3.1.1 Time Lag Before introducing the model it is crucial to highlight one of the features that have to be considered when working with Google Trends. As introduced earlier Google Trends are working closely with consumer behavior. The consumer purchase decision is separated in various steps that are executed along different time periods. To capture the relationship between the consumer behaviors, at certain points in the timeline, it is necessary to include time lags. It is assumed that the optimal time lag of Google Trends 13.

(18) variables will improve the forecasting performance. Not including time lags will only give an explanation of the present state of sales and will not provide logical results in terms of forecasting. Different keywords account for different phases along the customer journey. The following figure will illustrate the reasoning behind the implementation of time lags. A simplified example will be given afterwards.. 政治大. 立. ‧. ‧ 國. 學 er. io. sit. y. Nat. al. v. n. Figure 3.1: The modified 5-Step-Consumer Purchase Decision Model of (Kotler & Keller, 2016) shows the necessary of time lags. Ch. engchi. i n U. The keyword “luxury car” would fit in the category “Problem Recognition” where the consumer realizes the need for a luxury car, for various reasons. For the “Information Search” and “Evaluation of Alternatives” it is likely that certain brands will be googled, such as “Mercedes Benz” or “Lexus”. Furthermore consumers might search for promotions or financing options. During the “Purchase Decision” stage other factors will be more relevant for the final decision, such as the current situation of the economy in the consumer perspective (e.g. CPI) or the keywords that are related to the public opinion of a prospective purchase.. 14.

(19) 3.1.2 Autoregressive Distributed Lag Modeling The model used in this work is an autoregressive distributed lag model, short ADL model, which can include lagged dependent and independent variables. It allows the user to determine what effects variables have when a time lag is included. A simple ADL model: 𝑦𝑡 = 𝑐 + 𝛼1 𝑦𝑡−1 + 𝛽0 𝑥𝑡 + 𝛽1 𝑥𝑡−1 + 𝑢𝑡 In this model 𝑦 is the dependent variable and 𝑥 the independent variable. The term 𝑢𝑡 is. 治政 time lag of 1. In this model 𝛼 𝑎𝑛𝑑 𝛽 are called distributed 大 lag weight and reflect the 立 effect of changes in past appropriations, c is the intercept. The error term is considered referring to a white noise. 𝑦𝑡−1 and 𝑥𝑡−1 are lagged observations of the variable, with a. ‧ 國. 學. to be white-noise if each value in the sequence has a constant variance, a zero mean and is serially uncorrelated. So the following condition has to hold:. ‧. 𝐸(𝑢𝑡 ) = 𝐸(𝑢𝑡−1 ) = ⋯ = 0. y. sit. Nat. 2 𝐸(𝑢𝑡2 ) = 𝐸(𝑢𝑡−1 ) = ⋯ = 𝜎2. n. al. er. io. 𝐸(𝑢𝑡 𝑢𝑡−𝑠 ) = 𝐸(𝑢𝑡−𝑗 𝑢𝑡−𝑗−𝑠 ) = 0, 𝑓𝑜𝑟 𝑎𝑙𝑙 𝑢. i n U. v. If the error term has the desired properties, mentioned above, the data within the model. Ch. engchi. is stationary and can be estimated by ordinary least squares. (Hill, Griffiths, & Lim, 2010). 3.2 Data Collection and Scope 3.2.1 Dependent Variable The dependent variable in this work is monthly car sales of Mercedes-Benz cars in Thailand. The data originates from an automotive industry portal named Marklines, which obtained the data from press releases of Toyota Motor Thailand. The decision to choose this particular car manufacturer is motivated by its strong presence and. 15.

(20) development in recent years within the premium car market segment. Mercedes-Benz provides several model series in the upper class and luxury class segment. Their sales volume and market share has increased, from around 0.9% in 2012 to 4.44% in 2016 (Marklines). The importance of accurate sales forecasting for future success of this brand has been explained earlier. Monthly sales data is covering up the years 2008 to 2016, including a total of 96 observations.. 立. 政治大. ‧. ‧ 國. 學. n. er. io. sit. y. Nat. al. Ch. i n U. v. Figure 3.2: Monthly car sales of Mercedes Benz (thasalesmerc) (Marklines). 3.2.2 Macro Data Variables. engchi. A number of macro-economic variables that are related to car sales are also included in the model. These variables are chosen because they reflect the national economy’s state, which is assumed to have a large influence on the consumer’s decision to purchase a car. The collected data covers the time period from January 2008 to December 2016.. 16.

(21) Table 3.1: Economic variables: M – monthly data, Q – quarterly data, GEM – Global Economic Monitor and BOT – Bank of Thailand. Economic Variable. Frequency. Source. Explanation Measures the ratio of a price from a. Core Consumer. M. Price Index (CPI). GEM (The World Bank). fixed set of consumer goods and services in a current period to its price in a basic period, in Thailand. (2010 Base). Rate. Nat. Policy Interest. M. M. IndexMundi. Commodity price of crude oil. (IndexMundi). (petroleum), US$ per barrel. BOT (Bank of. SET Composite Stock Price Index. Thailand) BOT (Bank of Thailand). ‧. Index. ‧ 國. Stock Market. M. 學. Crude Oil Price. that are produced within Thailand. 政 Bank)治大 (Constant 2010 US$, in millions). (monthly average, local index) Policy Interest Rate of Thailand (end of. y. 立. Market value of all goods and services. period, % per annum). io. sit. Q. Product (GDP). GEM (The World. er. Gross Domestic. n. al. i n C 3.2.3 Google Trends Keyword Selection hengchi U. v. The most important preparation before starting to collect Google Trends data is to choose which potential keywords are advantageous for the model. Considering keywords relies on how representative they are as well as how good they fit into the logical framework of the model. Differences in culture, language and society are making it almost impossible to use the same words for different countries. After intense research and consultation of local Thai working professionals who were willing to assist in this project, a list of search terms has been created. The keywords possess high significance are listed below.. 17.

(22) Table 3.2: Summary of the Google Search Terms. Web Search Variable Name. Category. kwbenzcarthai. Autos & Vehicles. รถเบนซ ์. Benz Car. kwluxurycarthai. Autos & Vehicles. รถหรู. Luxury Car. kwpricebenzthai. Autos & Vehicles. ราคารถ Benz. Price Benz. kwbenzamg. Autos & Vehicles. Benz AMG. Benz AMG Mercedes Benz. Mercedes-Benz. Autos & Vehicles. Google Image Search. Mercedes-Benz. Benz Car. Nat. sit. y. ‧. รถเบนซ ์. kwbenzcarthaiiimg Autos & Vehicles. 學. kwm.b. Translation. 治政大Benz Autos & Vehicles Mercedes 立 ‧ 國. kwmb. Search Term. io. er. 3.2.4 Variable Correlation. al. n. iv n C U a higher correlation between correlation with the color red. Higherh color e nstrength g c h iequals The correlation plot indicates high positive correlation with blue color and high negative. the variables. The number inside the circles shows the percentage of correlation.. Noticeable correlations are the positive correlation between macro-variables cpi and stock and between kwbenzamg and salesmerc, cpi and stock. The negative correlation of kwm.b with cpi and stock are conspicuous as well.. 18.

(23) 立. 政治大. ‧. ‧ 國. 學. n. er. io. sit. y. Nat. al. Ch. engchi. Figure 3.3: Variable Correlation Plot. 19. i n U. v.

(24) 4 Empirical Model The statistical work and analysis have been executed with R version 3.3.1.. 4.1 Adjusted R² and Akaike Information Criterion To select the optimal combination of lags and variables, the Akaike Information Criterion (AIC) has been used. With a stepwise variable selection the goal is to achieve the highest adjusted R² while still having sufficient degrees of freedom as well as stationary residuals. The best models according to the R² and AIC achieve an adj. R² of 0.7 to almost 0.8.. 立. 政治大. Seasonal dummy variables are included in the macro model, clearly pointing out a. ‧ 國. 學. seasonal pattern in the data. The models that are including Google Trends, do not require seasonal dummy variables. Google Trends are representing the seasonality. ‧. since the consumer behavior reflects seasonality and is the direct cause of it.. Google Trends Macro + Google Trends. y. sit. Degrees of Freedom. er. Adj. R². a0.7345 76 i v l C n U h i e 0.706 n g c h 83. n. Macro. io. Model. Nat. Table 4.1: Summary of the predicted models. 0.7901. 82. 20.

(25) Macro Model Table 4.2: Macro-economic variables and seasonal dummies, outcome. Dependent variable: Mercedes Benz Cars - Sales Volume Interest rate, no lag -0.128** (0.060) Crude Oil Price, no lag 0.338** (0.142) Consumer Price Index, 4.lag 15.437*** (4.126) Sales Volume, 1.lag 0.470*** (0.112) February 0.412*** (0.155) March 0.398*** (0.150) April 0.334** (0.151) May 0.362** (0.146) June 0.364** (0.146) July 0.076 (0.146) August 0.340** (0.146) September 0.384** (0.145) October 0.045 (0.146) November 0.193 (0.147) December 0.264* (0.146) Constant -69.647*** (18.759). 政治大. Ch. engchi. y. sit. er. n. al. 92 0.778 0.735 0.280 (df = 76) 17.786*** (df = 15; 76). ‧. io. Note:. Nat. Observations R2 Adjusted R2 Residual Std. Error F Statistic. 學. ‧ 國. 立. i n U. v. *. p<0.1; **p<0.05; ***p<0.01. The model with macro variables is including 11 seasonal dummy variables to deal with seasonality. It can be seen that the Consumer Price Index (cpi), with a lag of 4 periods, and the previous sales volume of Mercedes-Benz cars (salesmerc) with one lagged period are the most significant variables. They both have positive coefficients. The policy interest rate (intrate) is heavily influencing the bank’s deposit and lending rate and is therefore likely to be one of the last decision criteria upon a purchase, lag 0. It has a negative coefficient which implies that increasing interest rates decreases the 21.

(26) purchase/investment of higher priced Mercedes-Benz cars, following the common macro-economic theory. The seasonal dummies play an important role in the model as well as Google Trends Model: Table 4.3: Google Trends variables, outcome. Dependent variable: Mercedes Benz Cars - Sales Volume Benz Car (Thai), no lag -0.173 (0.223) Mercedes-Benz, 2.lag -0.897** (0.361) Mercedes Benz, 2.lag 0.577** (0.244) Price Benz (Thai), 5.lag -0.047 (0.051) Luxury Car (Thai), 4.lag 0.062 (0.053) Benz Car (Thai) [Image search], 4.lag 0.261*** (0.099) Sales Volume, 1.lag 0.576*** (0.081) Constant 3.607** (1.540). 立. *. v. p<0.1; **p<0.05; ***p<0.01. n. al. er. sit. y. ‧. ‧ 國. 學. io. Note:. 91 0.729 0.706 0.295 (df = 83) 31.874*** (df = 7; 83). Nat. Observations R2 Adjusted R2 Residual Std. Error F Statistic. 政治大. Ch. engchi. i n U. The adjusted R² of the model that is almost solely consisting of Google Trends, with the exception of one period lagged Mercedes-Benz car sales, is slightly lower than the previous model. Keywords that include general information about cars, such as “รถหรู” (kwluxurycarthai) which stands for luxury cars in the Thai language is likely to gain consumer’s first interest in the brand. It is followed by images of Mercedes-Benz cars that can be found on Google picture search. The consumer is likely to use this tool to create a visualized overview of this brand’s different products. Both have positive coefficients. It seems that using the Keyword “Mercedes-Benz” (kwm.b) with one period lag, has a negative impact on the car sales.. 22.

(27) Google Trends & Macro Model: Table 4.4: Macro-economic variables combined with Google Trends, outcome. ‧. y. sit. *. p<0.1; **p<0.05; ***p<0.01. n. al. The highest adjusted R² can. er. io. Note:. 92 0.816 0.793 0.247 (df = 81) 35.904*** (df = 10; 81). Nat. Observations R2 Adjusted R2 Residual Std. Error F Statistic. ‧ 國. 立. 政治大. 學. Consumer Price Index, 1.lag Crude Oil Price, 2.lag Sales Volume, 1.lag Benz Car (Thai), 1.lag Benz Car (Thai), 2.lag Mercedes Benz, no lag Mercedes Benz, 2.lag Benz Car (Thai) [Image search], no lag Price Benz (Thai), 4.lag Benz AMG, 1.lag Constant. Dependent variable: Mercedes Benz Cars - Sales Volume 23.538*** (4.035) 0.204** (0.090) 0.314*** (0.087) -0.362* (0.211) -0.293 (0.196) 0.619*** (0.174) 0.623*** (0.186) -0.185** (0.084) 0.105*** (0.040) 0.193* (0.098) -108.631*** (18.561). iv n C be h achieved e n g cwith h ia Ucombination. of macro-economic. variables, Google Trends and one period lagged car sales of Mercedes-Benz. It is mentionable that the search term “Mercedes Benz” (kwmb) is highly significant and positive in both of the models that include Google Trends. Whereas a different name for the brand, which is more common among the majority of Thai people, called “รถเบนซ”์ (kwbenzcarthai), with a translated meaning of “Benz Car” has a negative coefficient. Interest rate does not play a significant role in this model, it seems that the used Google Trends can account for some of the effects that the interest rate causes.. 23.

(28) 4.1.1 Testing Residuals As stated in chapter 3.1.2 it is necessary for the model to have the desired properties of constant variance, zero mean and serially uncorrelated error terms, in order to keep classical properties of ordinary least squares (OLS) estimation. The Augmented DickeyFuller test (ADF) uses non-stationarity as the null hypothesis and stationarity as the alternative hypothesis. The Augmented Dickey-Fuller statistic is a negative number, which strengthens the rejection of the null hypothesis the more negative it is (Wikipedia, 2017).. 政治大. Table 4.5: Outcome of the Augmented Dickey Fuller Test (ADF) for the model residuals. 立. Outcome of ADF test. 學. ‧ 國. Model. Dickey-Fuller = -3.6889, Lag order = 4,. Macro. p-value = 0.02973. alternative hypothesis: stationary. ‧. Dickey-Fuller = -4.5992, Lag order = 4,. Google Trends. y. Nat. p-value = less than 0.01. n. al. Dickey-Fuller = -5.0772, Lag order = 4,. er. io. sit. alternative hypothesis: stationary. Macro & Google Trends. i n U. v. p-value = less than 0.01. Ch. alternative hypothesis: stationary. engchi. It can be observed that the residuals of each model are stationary, which allows using OLS estimation. However, the outcome of the ADF test for the macro model shows the weakest stationarity in comparison with the other two models. A combination of macro variables and Google Trends provides the highest test results in terms of a negative Dickey-Fuller statistic.. 24.

(29) 4.2 Forecasting The designated forecasting period is from 01.2016 to 12.2016, in total 12 periods. Having actual data available gives the opportunity to execute an out-of-sample comparison of the predicted values and the actual values. In total there are 5 forecasting models. Each model creates a vector with predicted values that will be compared to the actual data vector. It is to mention that those vectors have different lengths, according to the n-step-ahead forecasts. Table 4.6: Vector length according to the forecasting model. Forecasting model. 立. 12. 學. ‧ 國. 1-step-ahead. 政治大 Vector length. 2-step-ahead. 10. ‧. 3-step-ahead. 11. 4-step-ahead. 9. y. Nat. 7. n. al. er. io. sit. 6-step-ahed. Ch. i n U. v. The forecasting is structured as a loop that repeats itself until the prescribed vector. engchi. length is reached. This model building period is done with 01.2008 to 12.2015 data. To increase the understanding of this method, an example will follow that uses the 4step-ahead forecasting model. At first a 4-step-ahead prediction is executed that creates 4 predicted values (𝑦̂1 , 𝑦̂2 , 𝑦̂3 𝑎𝑛𝑑 𝑦̂4). Only the 4th of the predicted values (𝑦̂4 ) will be transferred into the final prediction vector. After that the next time period’s actual data will be incorporated into the model, so that the new model is built from the data of 01.2008 to 01.2016. The loop repeats itself and adds another 4 th predicted value to the vector.. 25.

(30) The following chart gives a better overview on how the forecasts work. The 4-step-ahead forecasting model has been used in the visualization. Table 4.7: The forecasting process is building a predicted value vector. time. 8. 𝑦̂1−4. 𝑦̂2−1. 𝑦̂2−2. 𝑦̂2−3. 𝑦̂2−4. 𝑦̂3−1. 𝑦̂3−2. 𝑦̂3−3. 𝑦̂3−4. …. …. …. …. …. …. …. …. y. …. …. sit. al. n. 7. 𝑦̂1−3. io. 6. 𝑦̂1−2. 立. 5.2016. Ch. 8.2016. i n U …. engchi. er. 5. 𝑦̂1−1. ‧ 國. 4. 4.2016. Nat. Forecast loop. 3. 3.2016. 9.2016. 10.2016. 11.2016. 12.2016. 學. 2. 2.2016. ‧. 1. 治 7.2016 政 6.2016 大. 1.2016. v. 9 𝑦𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑒𝑑 = [𝑦̂1−4 𝑦̂2−4 𝑦̂3−4 … 𝑦̂9−4 ]. 26. …. …. …. …. …. …. …. …. …. 𝑦̂9−1. 𝑦̂9−2. 𝑦̂9−3. 𝑦̂9−4.

(31) 4.2.1 Evaluation Measurements The measurement of forecasting accuracy is the final tool that evaluates forecasts. A forecast is usually a set of projections selected as the most likely representation of the future. With this definition the core of measurements would be the forecasting error, which is the difference between predicted value and observed value. 𝜀𝑡 = 𝑦̂𝑡 − 𝑦𝑡 To measure the forecast accuracy this work will ignore the direction of the error. The. 政治大. used measures are non-scale-dependent and get their positive value by deriving from the absolute value or squared value. The percentage error is calculated as follow:. 立. ‧ 國. 𝑦̂𝑡 − 𝑦𝑡 ∗ 100 𝑦𝑡. 學. 𝑃𝐸𝑡 =. The most common, non-scale-dependent measure for forecasting is the Mean Absolute. ‧. Percentage Error (MAPE) and the Mean Square Percentage Error (MSPE). Due to the. y. Nat. underlying error distributions of only positive values and no upper bound, percentage. io. larger (Swanson, Jeff, & T.M., 2011).. n. al. ∑𝑛𝑡=1|𝑃𝐸𝑡 | 𝑀𝐴𝑃𝐸 = 𝑛. Ch. engchi. 𝑀𝑆𝑃𝐸 =. i n U. ∑𝑛𝑡=1(𝑃𝐸𝑡 )2 𝑛. 27. er. sit. errors are highly prone to be skewed to the right, which can make the error substantially. v.

(32) 5 Results 5.1 Data Analysis and Results of Empirical Analysis The forecasting and evaluation of the results has been executed with all three models that were built in 4.1. One model consists of macro variables, one uses Google Trends and another one combines both. The following chart gives an overview of the built models. Table 5.1: The three forecasting model’s summary. 立. Macro. -0.128 *. kwbenzcarthai(0). -0.173. 0.338 *. kwm.b(2). -0.897 *. 0.470 ***. kwmb(2). 0.398 **. kwbenzcarthaiimg(4). al. kwluxurycarthai(4). 0.334 *. Ch. y. kwpricebenzthai(5). iv. 0.576 ***. 0.577 * -0.047. sit. 0.412 **. n. Apr. salesmerc(1). io. Mar. 15.437 ***. Nat. Feb. 3.607. er. salesmerc(1). Intercept. ‧. crudeoil(0). -69.647. 學. intrate(0). ‧ 國. Intercept cpi(4). 政治大 Google Trends. 0.261 ** 0.576. Macron & Google Trends U i e n g cIntercept h -103.091. May. 0.362 *. Jun. 0.364 *. Jul. 0.076. cpi(1). 22.337 ***. Aug. 0.340 *. crudeoil(2). 0.183 *. Sep. 0.384 *. salesmerc(1). 0.331 ***. Oct. 0.045. kwbenzcarthai(1). -0.543 **. Nov. 0.193. kwmb(0). 0.589 **. Dec. 0.264 .. kwmb(2). 0.561 **. kwbenzcarthaiimg(0). -0.175 *. kwpricebenzthai(4). 0.108 **. kwbenzamg(1). 0.168 .. 28.

(33) The forecasting techniques introduced in 4.2 include 1-step-ahead, 2-step-ahead, 3step-ahed, 4-step-ahead and 6 step-ahead forecasting. All the different forecasts have been evaluated by both measurements that were introduced in 4.2.1. Table 5.1: Results of the evaluation with MAPE and MSPE. adj. r². DF. 1-step. 2-steps. 3-steps. 4-steps. 6-steps. ahead. ahead. ahead. ahead. ahead. 33.309. 37.217. 40.101. 39.583. 51.940. MSPE 治 20.159 18.879 政 % 大. 38.295. 33.959. 54.727. MAPE %. 25.820. 24.117. 29.193. 23.505. 32.720. MSPE %. 11.538. 11.820. 17.424. 8.125. 21.024. MAPE %. 32.654. 30.974. 34.611. 35.744. MSPE %. 18.045. 21.371. 22.396. 40.165. in units MAPE %. 0.7345. 76. 82. n. al. Ch. 30.499. y. 0.7901. io. Trends. Nat. Google. 83. ‧. Macro &. 0.706. sit. Trends. 學. Google. ‧ 國. 立. 24.651. er. Macro. engchi. i n U. v. This informative chart provides a good summary of all evaluations. The percentage error, measured by MSPE, is lower in all the cases. The difference between the various n-step-ahead methods is remarkable. Tendency is that the more periods a forecast is predicting, the higher the percentage error will be. Results that are worth noticing are especially those starting from 3-steps-ahead to 6-steps-ahead. In this period there are the lowest percentage errors observable. Furthermore is this time period more applicable to actual business planning since it delivers results for several months in advance. The 1-step-ahead approach is rather short sighted.. 29.

(34) MAPE 60. Macro GT. 50. GTMacro MAPE/%. 40 30 20 10. 立. 0 2. y. sit er. al. n. 40. io. GTMacro. 30. 6. ‧. GT. 50. MSPE. Nat. Macro. 4. 學. 60. MSPE/%. 3 Step-ahead. ‧ 國. 1. 政治大. Ch. engchi. i n U. v. 20 10 0 1. 2. 3. 4. Step-ahead. Figure 5.1: MAPE and MSPE. The x axis denotes the n-step-ahead forecasting method. 30. 6.

(35) The macro model and the macro & Google Trends model both have higher accuracy in the short sight forecasting. They lose much of their accuracy when the predicted time period increases. According to both evaluation measurements the accuracy can be increased by including Google Trends.. 5.2 Key Findings As the literature states out there are problems with forecasting accuracy in the automobile industry that might be due to outdated techniques (Dharmani, Anand, & Dr. Demirci, 2013).. 政治大. The approach of using Google Trends as a predictor for sales forecasting is the core of. 立. this work. There is literature which shows that including Google Trends can increase the. ‧ 國. 學. forecasting accuracy. The models that were used in those literature however did either not include time lags (Choi & Varian, Predicting the Present with Google Trends, 2012). ‧. or did not include a full range of different Google Search terms that are related to the sales (Kinski, 2016).. y. Nat. sit. Key finding of this work is the relevance of Google Trends in modern sales forecasting. al. er. io. under the circumstances that are given, the time period, the monthly data and the. n. industry. Logic behind the use of Google Trends in forecasting is to capture consumer. Ch. i n U. v. behavior and show how it influences the purchase decision. The five stages model of. engchi. Kotler & Keller (2016) was used as a foundation to understand the importance of time lags when working with Google Trends and consumer behavior. The direct connection between purchase and consumer behavior offers an alternative way to traditional forecasting which mostly consists of macro-economic variables and seasonal patterns (Hülsmann, Borscheid, Friedrich, & Reith, 2012). The initial approach in this work was to include Google Trends as a complimentary tool for sales forecasting to find out if it improves forecasting accuracy in an emerging country, Thailand. Research that deals with Google Trends in terms of sales forecasting mostly deals with developed countries, USA and Germany for example. The amount of data available, macro data as well as internet data, in those countries is substantially 31.

(36) higher than for emerging countries like Thailand, Indonesia etc. However, there is sufficient Google Trends data in almost every emerging country that allows the Google service. Data analysis showed that the implementation of Google Trends improves forecasting accuracy. Furthermore it allows omitting seasonal dummy variables, forasmuch as it can deal with seasonality on its own. The use of evaluation measurements, introduced in 4.2.1, is recommended to proof the improving forecasting accuracy. Including Google Trends in a macro model, that has been modified to not include seasonal dummy variables, outperforms the pure macro model.. 政治大. Table 5.2: Forecasting accuracy improvement. Green color indicates an improvement. avg. MSPE. 40.430%. 32.894%. 33.204%. 25.326%. 0.7345. 0.7901. Change -18.6% -23.7% +7.04%. Nat. sit. y. ‧. Adj. R². Macro & GT. 學. avg. MAPE. Macro 立. ‧ 國. Measurement. io. er. Furthermore is has been proven that leaving out the macro-economic variables completely, will further increase the accuracy. The improvement that can be achieved. n. al. Ch. i n U. v. with only Google Trends, in comparison to the other two models, is substantial.. engchi. Table 5.3: Forecasting accuracy improvement. Red color indicates a worsening. Measurement. Macro & GT. GT. Change. avg. MAPE. 32.894%. 27.071%. -17.7%. avg. MSPE. 25.326%. 13.986%. -44.8%. Adj. R². 0.7901. 0.706. -10.6%. 32.

(37) 6 Conclusion 6.1 Discussion The implementation of estimators that deal with special events are crucial to good forecasting models, according to Hülsmann et. al. (2012). The sales data shows suspicious behavior that is neither explainable with seasonality nor with the macroeconomic variables that has been used in this work. Starting in 2014, sales in December are drastically decreased.. 立. 政治大. ‧. ‧ 國. 學. n. er. io. sit. y. Nat. al. Ch. engchi. i n U. v. Figure 6.1: Sales drops that start in December,2014. Even though there is a major event that has happened in 2014, which is the military coup, it is hard to find evidence that directly links to this behavior. Nonetheless to still catch the effects that are caused by it there have been tests with special event estimators included in the model. The special event estimator is a dummy variable that. 33.

(38) becomes active for every December after 2013. Doing so leads to a significant increase in the forecasting accuracy. The Google Trends model is able to lower its MAPE by around 10% with including this special event estimator. Nevertheless it is a very specific estimator that does not have evidence for a longer time period.. 6.2 Practical Application and Limitations 6.2.1 Implementation in the Automotive Industry The main area of appliance for this work is in the automotive industry which is. 政治大 construction of infrastructure. Since this work is dealing with one of the Tiger Cubs, it 立. eventually affecting a broad range of industries, from the supplier industry to the. has its biggest value for other emerging markets such as Indonesia, Malaysia, Vietnam. ‧ 國. 學. and the Philippines. Those countries share a lot of properties that would differ significantly to other emerging countries in Asia or South America.. ‧. A logical and optimized selection of Google Trends will be a challenge. For a successful forecasting model is it unavoidable to contact local experts as well as consumers. With. y. Nat. sit. the consumer being the main player in this model it is highly important to understand. al. er. io. their purchase decision researching behavior. Different countries will have different. n. approaches and therefore it is impossible to find one universal set of Google Trends that fits every country.. Ch. engchi. i n U. v. This also shows up the sheer unlimited number of possibilities for further improvement. Applying to countries with a substantial different culture, it is suggested to use Hofstede’s Cultural Dimensions, which are dealing with differences in the culture and society, to obtain a broad idea to start with. It has been used to provide an insight in time lags in different countries (Kinski, 2016).. 6.2.2 Major Implications for Sales There are several industries that can benefit from the use of Google Trends in their forecasting. In general all the industries, which products are part of a consumer 34.

(39) purchase process that includes information research over the internet, can be considered. Durable goods that are linked to an attentive research before purchase are in focus for this approach. Inexpensive goods that are subject to spontaneous decisions or needs do not require extensive research. For those products the selection of Google Trends keywords that were used in this thesis are inappropriate. This being said, it does not mean that Google Trends will not improve forecasting models for those products. It is strongly suggested for researchers to focus on those kinds of products as well, using a different selection of Google Trends keywords. The brand itself will probably be searched for less, because. 治政 look up right before a purchase, which could be related大 to the problem that this product 立 might be a relation between promotions, coupons, is solving. Furthermore there the effect that those products give is desired. There are search terms that people will. ‧ 國. 學. discounts and the search volume.. Global or national events that can trigger a certain need or fear in the consumer’s mind. ‧. should also be considered as it can influence the product sales. A subjective increase in criminality, induced by the increasing amount of negative news, might cause people to. y. Nat. sit. rethink their safety and lead to higher sales in safety equipment. Sport events, New. al. er. io. Year’s resolutions or increasing reports of obesity might lead to people informing their. n. selves about exercising and will eventually lead in an increase in gym memberships or. Ch. fitness related products to be sold.. engchi. i n U. v. Finally this work emphasizes on the enormous potential that Google Trends is contributing to the forecasting field. Almost unlimited data can be obtained without the effort that comes along from doing surveys. In addition to mostly first degree related keywords, which were mostly used in this work, there are also second degree related keywords that are widening the horizon for possible forecasts even further.. 35.

(40) 6.3 Limitation of this Research There is a clear limitation of Google Trends data so far, which is the long term prediction possibility. Macro-economic data has several times proved its significance to identify economic movements and trends in the long run. There are many areas applicable to macro-economic data. So far there is no research done for long term forecasting with Google Trends. One of the valid reasons that this has not been conducted yet is the availability of Google Trends. With just about 10 years of data since launch of Google Trends, there is just not enough information to run any models for a long period. The long term significance of Google Trends has not been validated and will probably. 政治大. need another 10 years to get sufficient data for it.. 立. ‧. ‧ 國. 學. n. er. io. sit. y. Nat. al. Ch. engchi. 36. i n U. v.

(41) References. Carrière-Swallow, Y., & Labbé, F. (2010). Nowcasting with Google Trends in an Emerging Market. Central Bank of Chile. Choi, H., & Varian, H. (2009). Predicting the Present with Google Trends. Google Inc. Choi, H., & Varian, H. (2012). Predicting the Present with Google Trends. Economic Record, Vol. 88, No. 1, 2-9.. 政治大. Crain Communication Inc. (2013). Top Suppliers. Automotive News.. 立. Dharmani, S., Anand, D., & Dr. Demirci, M. (2013). Shifting Gear - Capacity. ‧ 國. 學. management in the automotive industry. Ernst & Young Global Limited. Fantazzini, D., & Toktamysova, Z. (2015). Forecasting German Car Sales Using Google. ‧. Data and Multivariate Models. Internal Journal of Production Economics, Vol. 170,. sit. y. Nat. Part A, 97-135.. Goel, S. H. (2010). Predicting consumer behavior with Web search. Proceedings of the. io. n. al. er. National Academy of Sciences, Vol. 107, No. 41, 17486 – 17490.. Ch. i n U. v. Gredenhof, M., & Karlsson, S. (1997). Lag-length Selection in VAR-models Using Equal. engchi. and Unequal lag-length procedures. 177.. Green, K. C., & Armstrong, J. S. (2015). Simple vs. complex forecasting: The evidence. Journal of Business Research, Vol. 68, 1678–1685. Hill, C. R., Griffiths, W. E., & Lim, G. C. (2010). Principles of Econometrics. John Wiley & Sons, Incorporated. Hofstede, G. (n.d.). Cultural Dimensions. (ITIM International) Retrieved March 12, 2017, from https://geert-hofstede.com/. 37.

(42) Hülsmann, M., Borscheid, D., Friedrich, C. M., & Reith, D. (2012). General Sales Forecast Models for Automobile Markets and their Analysis. Vol. 5, No. 2 (2012) 65-86. Karlsson, E., & Annie, R. (2014). Development of a Sales and Operation Planning Process. Göteburg, Sweden. Kinski, A. (2016). Google Trends as Complementary Tool for New Car Sales Forecasting: A Cross-Country Comparison along the Customer Journey. University of Twente.. 政治大. Kotler, P., & Keller, K. L. (2016). A Framework for Marketing Management. Pearson Education Limited.. 立. Marklines. (n.d.). Automotive Industry Portal Marklines - Vehicle Sales. (MarkLines Co.,. ‧ 國. 學. Ltd.) Retrieved February 17, 2017, from Automotive Industry Portal Marklines: https://www.marklines.com/en/vehicle_sales/index. ‧. Pan. B., W. D. (2012). Forecasting hotel room demand using search engine data.. sit. y. Nat. Journal of Hospitality and Tourism Technology, Vol. 3, No. 3, 196-210.. al. er. io. Return On Now. (2015). Search Engine Market Share. Retrieved May 27, 2017, from. n. Returnonnow: http://returnonnow.com/internet-marketing-resources/2015-search-. Ch. engine-market-share-by-country/. engchi. i n U. v. Rieg, R. (2010). Do Forecasts Improve over Time? A case study of the accuracy of sales forecasting at a German car manufacturer. IJAIM, Vol.18, No.3, 222-236. Rogers, S. (2016, July 2). Retrieved May 22, 2017, from https://medium.com/googlenews-lab/what-is-google-trends-data-and-what-does-it-mean-b48f07342ee8 Schmidt, T., & Simeon, V. (2009). Forecasting Private Consumption: Survey-based Indicators vs. Google Trends. Journal of Forecasting, Vol. 30, No. 6, Ruhr economic papers, No. 155, 565-578.. 38.

(43) Shende, V. (2014). Analysis of Research in Consumer Behavior of Automobile Passenger Car Customer. International Journal of Scientific and Research Publications, Vol. 4, No. 2, 1-8. Swanson, D. A., Jeff, T., & T.M., B. (2011). MAPE-R: A rescaled measure of accuracy for cross-sectional forecasts. Springer. The World Bank. (n.d.). The World Bank - Global Economic Monitor. (The World Bank Group) Retrieved February 18, 2017, from The World Bank IBRD IDA: http://data.worldbank.org/data-catalog/global-economic-monitor. 治政 The World 大. The World Bank. (n.d.). World DataBank. (The World Bank Group) Retrieved February 23,. 2017,. from. 立. Bank. IBRD. IDA:. http://databank.worldbank.org/data/home.aspx. ‧ 國. 學. Wharton. (2011, September 15). Wharton Magazine. Retrieved April 25, 2017, from http://whartonmagazine.com/blogs/importance-of-simple-forecasting-methods/. ‧. Wikipedia. (2017). Augmented Dickey Fuller Test. Retrieved April 10, 2017, from. sit. y. Nat. Wikipedia:. io. al. n. zur Muehlen, M. (2016). Automotive Industry Thailand.. Ch. engchi. 39. er. https://en.wikipedia.org/wiki/Augmented_Dickey%E2%80%93Fuller_test. i n U. v.

(44)