Price setting behavior of web-shops.
Using data from a German price comparison websiteStudent: Niels Westenberg
Student ID: 10264264
Supervisor:
dhr. dr.
A.P. Kiss
1
Abstract
This thesis aims to provide an answer to the question if firms are adopting a mixed pricing strategy. Theoretical models show that there is no pure strategy in finite games. Having a look at different information search models, one can conclude that if there is a third-party that list prices, a mixed strategy will yield the best outcome for a firm. However, this has not been tested empirically in the case of web-shops. Using data on prices, web-shops and service from www.billiger.de I calculated first-order autocorrelations and cross-correlations to test whether other relative prices can indicate a firm’s relative price of today. I find significant first-order autocorrelations and almost 60% significant cross-correlations within the same firm. Also the service premium has been researched. I do find significant results when a firm is trusted and when a product has positive consumer ratings. However, these estimates are too small to explain a relative price and the price gap. I conclude that web-shops are not using the mixed pricing strategy.
2 Statement of Originality
This document is written by Student Niels Westenberg 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.
3
1. Introduction
When firms are competing on prices with a homogeneous good, there is no pure strategy that will lead to a Nash equilibrium (Shilony, 1977). One might expect that undercutting your rivals is the best strategy. However, when all firms are setting prices at marginal cost, they do not make any profit. Firms need to adopt a mixed strategy, such that prices are random. Different theoretical papers about information search models with different underlying conditions show this (Rosenthal, 1980; Varian, 1980; Baye & Morgan, 2001). The common assumption is that there is a third-party involved, which lists prices. This is never tested empirically for firms. This paper will answer the question if web-shops are indeed using a mixed pricing strategy.
To see if firms actually set prices at random, I will test if one price can explain another price. I make use of a new dataset with billiger.de as the third-party. This dataset includes prices of different TV’s of different web-shops offered on billiger.de. I calculate the relative price of a certain product of a certain web-shop for each day, by dividing the listed price of a TV by the average price of that TV of that day. This relative price will be used to test whether prices are random. Previous prices should not be correlated with today’s prices.
Furthermore, there should be no correlation between the relative prices of different products listed by a web-shop if prices are random. Also the consumer rating and trusted label should not give any information regarding the relative price. Regressing the relative price on the consumer rating and trusted label should not return any significant results.
Testing for first-order autocorrelation returns significant results with an estimate of around 0.80. That means that yesterday’s relative prices give a good indication of today’s relative prices. Also the seventh lagged relative price is significant. However, this estimate is much smaller and is almost no bigger than 0.10. If we look at the correlations between the products of the same firm, I find that around 60% of all the relative prices of the TV’s are correlated with each other. This is significantly more than the 5% cross-correlations you find by chance even if there is no correlation at all. Looking at the service premium, we find that when a firm is trusted the relative price will increase for a small and medium TV, but
decreases for a big TV. On the other hand, a positive consumer rating, a rating of 70 and more, will decrease the relative price. When regressing on all TV’s, all the estimators are significant. However, the “trusted” dummy does not have an effect of more than 5
4
percentage points. Also the consumer ratings are small. Summarizing these results, I find that web-shops are not randomizing prices.
2. Literature Review
The use of Internet has become a bigger part of our daily routine. We check our emails, watch videos and search for information for example. The Internet is expected to increase competition, because information about products and services is more accessible. Firms can see prices of their competitors faster and they can adapt to it. But also consumers can see the information provided on such websites more easily and consumers are more informed about prices they are looking for. However, since users of the Internet can see all the information in one click, you expect that firms want to attract consumers by undercutting each other (Baylis & Perloff, 2002). If consumers are rational they will buy the cheapest TV listed. Web-shops are expected to compete Bertrand style with a homogenous product and undercut if they can. But is this really the strategy they use? If we take a look at picture 2.1, we see that the same TV has a price between €255.59 and €358.22.
Picture 2.1
Prices of a random TV selected from billiger.de
5
One possible explanation for the price gap is that the willingness to pay differs among consumers (Homburg et al., 2005). Consumers will pay more if they are satisfied. Satisfaction can be achieved by excellent service, or by a good price/quality ratio for example. This service can be influenced by many different factors, like warranty, expertise of the seller, kindness in a face-to-face conversation. However, the last two factors are absent when purchasing via a web-shop. Baylis and Perloff (2002) expected that there was a trade – off between prices and service in the electronics market, using the camera as sample. They found that firms who have unfriendly services for the consumers charge higher prices. On the other hand, there is a paper that researches this question in the hospital market (Mukamel & Mushlin, 1998). From that paper can be concluded that a good quality report increases the prices they charge. Last but not least, Pan et al. (2002) also tried to answer the question if there is the so-called service premium. They found in general that this effect is really small and hardly explains the price dispersion among e-retailers.
The big price gap may indicate that firms are not setting prices at marginal costs. From a game theoretical point of view, setting prices at marginal cost is not the best strategy. The so called information search models, because consumers have to find the correct information about different goods, prove this. I will briefly recap some well-known models, each with different assumptions. Rosenthal (1980) assumed that there was a part of consumers that were loyal to a firm. These consumers buy the good of a firm, no matter if it is more expensive than offered by another firm. To have your firm listed on the third-party was at zero costs. Another model (Varian, 1980) made a distinction between consumers as well. There are uninformed consumers, who have no idea of other prices and informed consumer who are perfectly informed about all prices. That is because it was time
consuming, and therefore costly, to search for information. Also in this model it was costless to have your price listed by a third-party. Baye and Morgan (2001) assumed that consumers did not incur any cost to search for information, but that a firm had to pay to get the price listed by a third-party. However, all these different models led to the same conclusion. When there is a third-party involved, there is only a mixed strategy equilibrium and no pure
strategy that will lead to Nash equilibrium.
Even more general, when there is a finite game form there is always a mixed strategy equilibrium (Azar & Bar-Eli, 2011). Sometimes there can be a pure strategy equilibrium as well, but this does not necessarily occur. This is tested empirically in another setting with
6
penalty kicks (Chiappori et al., 2002). They divided the goal into three direction parts, namely right, center and left. Data from the France and Italian elite league is used of a timespan of three years. They found that shooting penalty kicks to the right, center or left at random was the most successful and the penalty kick takers should use a mixed strategy.
A research that is closer to this subject is done by Kaplan et al. in 2016. Kaplan et al. find that some supermarkets set a high price and a low price for two complementary goods, like tea and sugar. In this way they attract both type of consumers, the consumers who buy everything in the same store, but also the consumers who go to different shops. The latter group will now also buy the cheaper product. The biggest difference between web-shops and tangible super-markets is the availability of information. Also these web-shops are selling solely TV’s and therefor have no complementary good.
3. Data
Data is collected of prices, different TV’s and vendors for a period of nine months. The data was retrieved in April 2002 and the process ended in February 2013. The data is collected each day from a German price comparison website, called billiger.de. At first, this dataset had over 2.500.000 observations, but this dataset had to be cleaned. One of the vendors was ebay.de, a website where everybody can offer their products. Since the actual seller can be different for each TV when offering a product on ebay.de, this vendor had to be dropped out of the dataset. It is almost impossible to track down the real seller on this website. Also, some vendors were offering their TV multiple times a day for a complete different price. I only included the lowest price offered by that vendor for that day. The much higher price by the same vendor could be a bundle. This research is only about homogenous goods, and bundling is a method to differentiate. Billiger.de is also listing price of table TV’s. These TV’s are TV’s that are built in a table and or mostly used for educational purposes. They are relatively expense compared to a TV’s of the same size and are not used for watching
television. I decided to remove these observations from the dataset as well. Also some prices were displayed wrong in the dataset. Some TV’s had a price of €7.20 or €300100.95. Since I can in no way find out the real price, these corrupted records have to be removed. In total, after fixing the dataset, there are 1034701 observations and 1944 different TV’s, offered by 191 different vendors. A nice overview can be seen in Table 4.1.
7
Table 4.1.
Small overview of variables.
VARIABLES Minimum value Maximum value
Price €129.99 €7999.99
TV Size 15 Inch 99 Inch
Consumer Rating 20 100
Trusted 0 1
Not only prices are displayed in the dataset, but consumer rating, number of ratings and a trusted label is also included. I use these variables to get an indication about the service. I only take the consumer ratings into account when they are rated five times or more. There are 1593 TV’s that have a consumer rating which is rated five times or more. The number of web-shops that have a trusted label is 73.
According to Birghthub (2010), TV’s can be categorized into three groups based on their size in inches. Small TV’s are less than 32 inch, medium TV’s are between 32 inch and 46 inch and big TV’s are equal or bigger than 46 inch. There are 271 TV’s that are small, 1065 medium sized TV’s and 608 big sized TV’s.
Table 4.2.
Small overview of the groups.
VARIABLES Minimum value Maximum value Mean
Price small TV €129.99 €873.46 €297.34
Price medium TV €169 €2699 €569.10
Price big TV €199 €7999.99 €1256.04
The cheapest TV’s of each group do not differ that much in price, but the maximum price of the big TV is way bigger than the medium TV. Also the average prices of each group are different. This is important, because it may explain difference in results between groups if there are found any.
8
4. Relative price model
To empirically test the theory of randomizing prices, I make use of a relative price model. Data on prices is gathered from the German price comparison website billiger.de. In order to answer my sub-questions, I will calculate the relative price of each TV, listed by the firms, on a daily base. The relative price is calculated by dividing the listed price of a TV on a day by the average price of that TV on that day. The following formula will calculate the relative price of a TV on a certain day:
𝑅𝑒𝑙𝑎𝑡𝑖𝑣𝑒 𝑃𝑟𝑖𝑐𝑒𝑖𝑗𝑘 =
𝑃𝑟𝑖𝑐𝑒𝑖𝑗𝑘 𝑀𝑒𝑎𝑛𝑖𝑘
𝑤𝑖𝑡ℎ 𝑇𝑉𝑖, 𝑣𝑒𝑛𝑑𝑜𝑟𝑗 𝑎𝑛𝑑 𝑑𝑎𝑦𝑘
This relative price will be used instead of the listed price, because it might be that the underlying market where these firms buy their TV’s from have sudden price increases or decreases. By dividing the price by the mean I correct for such shocks and the relative price
might not change because of exogenous effects.
To test whether firms are randomizing prices, I will cross-correlate ten randomly selected TV’s offered by the same firm. Not only the products are randomly selected, but also the firms to avoid biased results. Then these correlations need to be tested on being significantly different from zero. If firms are really setting prices at random, I expect not to see any significant correlations across the products of the same firm. Not only will cross-correlations be tested, but if price setting is random, the lagged (previous) value is not correlated with the present value. So I will perform a test for autocorrelation and see if this is indeed the case.
There might also be another indicator which reveals the relative price of a product. As mentioned before, service might be important and therefore change the relative price. Additional regressions will be done with on the left hand side the relative price and on the right hand side the variables that define the service. These variables include a dummy that is true when the firm has the trusted label and the consumer ratings. To check if there is another relation besides a linear one between consumer rating and relative price, I will include the quadratic consumer rating as well. I will add dummies for the months and all product as control variables. In the regressions I will distinguish between all TV’s, the groups based on their size and individual sizes.
9
5. Results
In this chapter, both sub-questions will be answered empirically with use of the relative price model described above.
The first and most logical possibility is that the previous listed prices can give a good indication of today’s prices. If prices are totally random, previous listed prices should not have a significant effect on today’s prices. To see if this is true, I first tested for serial correlation. Since this panel data is unbalanced, I used the Wooldridge test for
autocorrelation. This will test the null hypothesis that there is no first-order autocorrelation in the panel data. The test calculated a F-statistic of 538.085 with distribution 𝐹(1, 15017). For this distribution, the critical value is 6.63, at a 1% level (Stock & Watson, 2012). Since 538.085 is bigger than 6.63, we have to reject the null hypothesis. This means that there is first-order autocorrelation.
Not only yesterday’s price might be a good estimator for today’s price, but it might be the case that prices that are listed more days ago may give a good indication of today’s prices. For example, a vendor might have the strategy to set a price every three days that he wants and let the price float the other days of the week. A regression on past daily prices for a week is needed to test this. The first regression includes all TV’s, regression (2) includes only TV’s when the relative price is larger than one, regression (3) includes only TV’s when the relative price is smaller than one, regression (4) includes only small sized TV’s, regression (5) includes only medium sized TV’s and the last regression (6) includes only big sized TV’s. The output can be seen in table 5.1.
From this table we can also see that there is a first-order autocorrelation for all TV’s, for all the significance levels. Also after dividing the TV’s into groups because of certain characteristics, we still see that the first lagged relative price is significant and positive. The value is also close to one, around 0.80. If we go into detail by focusing on regression (1) we find that the fifth and seventh lag are also significant. The fifth lag is only significant at the 10% level, while the seventh lag estimator is significant for all levels. However, both
estimators are not bigger than 0.10. Regression (2), including TV’s when the relative price is larger than one, has no other significant estimators. This regression has also the lowest R2. In
regression (3), including TV’s when the relative price is smaller than one, shows significant estimators for the third, fifth, sixed and seventh lag. Only the estimator of the seventh lag is
10
significant for all levels. Regression (4), including only small sized TV’s, has also multiple significant estimators. The estimators of the second, fourth, sixed and seventh lag are all significant, the first two at all levels, the last two at 5% level respectively. Also here we have to put a note that these estimators are not larger than 0.10. This regression has also the least observations of all regressions in this table. Regression (5) has no significant estimators at the 1% level. There is only one significant estimator at the 5% level, which is for the third lag. The estimators of lag five and six are significance at the 10% level. Again we see that none of the estimators is larger than 0.10. The last regression, including all big sized TV’s, has a significant seventh lag estimator of 0.0845, for all levels. Also the estimator of the sixed lag is significant, but on a lower level. It is the only significant negative estimator we find in all regressions. If we take a look at the general picture, we see that the estimator of the seventh lag is significant in most cases. In regression (2) and regression (5) this estimator is not significant different from zero at any level. However, this estimator is in all regressions but one smaller than 0.10.
11
Table 5.1.
Regression on seven lagged Relative Prices.
(1) (2) (3) (4) (5) (6)
VARIABLES Relative Price Relative Price>1 Relative Price<1 Relative Price (small TV’s) Relative Price (medium TV’s) Relative Price (big TV’s) Lags(1) 0.811*** 0.824*** 0.715*** 0.804*** 0.754*** 0.829*** (0.0198) (0.0202) (0.0426) (0.0182) (0.0340) (0.0234) Lags(2) 0.0225 0.0334 0.00666 0.0976*** -0.0200 0.0375 (0.0269) (0.0214) (0.0529) (0.0135) (0.0727) (0.0234) Lags(3) 0.0258 -0.000562 0.0622** 0.00529 0.0872** 0.00124 (0.0205) (0.0222) (0.0279) (0.0111) (0.0378) (0.0218) Lags(4) 0.0118 0.00766 0.0192 0.0477*** 0.0260 0.00683 (0.0154) (0.0201) (0.0222) (0.00956) (0.0204) (0.0193) Lags(5) 0.0217* 0.0127 0.0384** -0.00539 0.0293* 0.0217 (0.0120) (0.0153) (0.0172) (0.0107) (0.0171) (0.0153) Lags(6) -0.0295 -0.0110 -0.0575* 0.0218** 0.0347* -0.0476* (0.0214) (0.0231) (0.0337) (0.00852) (0.0205) (0.0262) Lags(7) 0.0692*** 0.0329 0.113*** 0.0155** 0.0134 0.0845*** (0.0185) (0.0226) (0.0320) (0.00735) (0.0148) (0.0226) Constant 0.0678*** 0.121*** 0.0852*** 0.0134*** 0.0750*** 0.0674*** (0.00579) (0.0101) (0.0133) (0.00168) (0.0145) (0.00608) Observations 524,945 225,751 271,672 65,925 295,059 163,961 R-squared 0.815 0.749 0.825 0.958 0.814 0.812
Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1
12
Besides that the price of previous days may have an effect on today’s prices, it might also be the case that prices of other products that a firm is offering have a significant correlation. Therefore, I will calculate correlations of relative prices across products within the same firm. Only products with a lifespan of more than 50 days will be selected to have enough observations.
The first firm that will be researched if prices are random is xxx-deals.de. In figure 5.1 the evolution of the average price of all products is displayed. The average price of all products was in the beginning around €700. After a month, 30 days, the average price was below €600 and was back at the original level after 100 days. From that moment the average price started to grow up to €800. From this figure we can see that xxx-deals.de was very active in setting prices, since they are far from constant over time.
Figure 5.1.
Plot of the average price of all products combined offered by xxx-deals.de.
5 0 0 6 0 0 7 0 0 8 0 0 9 0 0 Ave ra g e p ri ce o f a ll p ro d u ct s o ff e re d 0 50 100 150 200
13
To test whether these prices are random a correlation matrix is displayed in table 5.2. Ten randomly selected products from xxx-deals.de are used to calculate the cross-correlation using pair-wise correlations. All the correlations are tested on being significantly different from zero at the 5% level. We can see that in most cases a TV of a certain brand is
correlated with a TV of that same brand. The only exception is the Samsung TV, the
SAMSUNG-22INCH is uncorrelated with the SAMSUNG-37INCH. Also if we look at TV with the same size we see that both 46-inch TV’s are correlated. There is also a significant correlation between both 32-inch TV’s. However, the LG-42INCH and PANASON-42INCH are
uncorrelated. The biggest correlation is between PHILIPS-40INCH and PANASONIC-32INCH. This cross-correlation is 0.9547, almost close the one. There is no TV which is correlated with all other TV’s. The GRUNDIG-32INCH has the most cross-correlations with the other TV’s, with seven significant correlations. If we count the total significant results, we find 25 significant correlations. There are 39 different correlations in total. That means that out of all products that xxx-deals.de are offering, 64% of the products are correlated with each other.
The next firm that is used for this research is redcoon. Again is the evolution of the average price of all products plotted in graph 5.2. In contrast to xxx-deals.de, the average price does not fluctuate that much. It seems to be that the prices are more stable with some outliers. This can be due that the same set of products is not offered all the time. Sometimes a product is not offered and will be replaced by another one. That might be the case here. The average price seems to float around €720. It is too early to draw any conclusions with
respect to the price setting behavior of redcoon. Based on this graph prices do not look random, but further research is needed.
14
Figure 5.2.
Plot of the average price of all products combined offered by redcoon.
Again I will select ten products from redcoon randomly and use these products to calculate the cross-correlations. This correlation matrix is displayed in table 5.2. The same sized TV’s are all correlated with each other. GRUNDIG-26INCH is correlated with PHILIPS-26INCH, PANASONIC-32INCH and SAMSUNG-32INCH are correlated but the 46-inch TV’s do not show any significant cross-correlations. Interesting to see is that the TV’s of the same brand are correlated, except the Samsung TV’s. The Grundig TV’s show a correlation of 0.3448 and the LG TV’s show a correlation of 0.7798. The LG-37INCH has the most significant
cross-correlations with the other products. There are seven cross-correlations which are significant. The total amount of significant cross-correlations is 25 out of 39, which again results in 64% of all correlations being significant.
6 5 0 7 0 0 7 5 0 8 0 0 8 5 0 9 0 0 Ave ra g e p ri ce o f a ll p ro d u ct s o ff e re d 0 100 200 300
15
Figure 5.3.
Plot of the average price of all products combined offered by getgoods.de.
The last firm that will be used to answer the question is getgoods.de. As we can see from figure 5.3 the average price of all the goods combined are decreasing over time. It starts at a price of almost €1000 and after 300 days of listing the prices the average price drops to almost €600. This is a decrease of almost 40%. However, since day 100 the average price seems to be stable towards the end.
In table 5.4 the correlation matrix is presented of ten different products which are offered by getgoods.de. If we have a look at the correlation between TV’s of the same brand we find a significant between the TV of Philips. The 32-inch and 42-inch TV of Philips are correlated. The 32-ich TV of Samsung is only correlated with the 55-inch TV of Samsung and not with the other Samsung TV’s. The 55-inch TV of Samsung is correlated with all the other Samsung TV’s. The same sized TV’s show some correlations, but not for all. The 32-inch TV of Philips is correlated with both other 32-inch sized TV’s. The 32-inch TV of Samsung and Grundig are not correlated. There are also three 55-inch TV’s in the matrix, from Samsung, Sony and Toshiba. The Toshiba TV has a significant correlation with both other TV’s. The Sony TV and
4 0 0 6 0 0 8 0 0 1 0 0 0 1 2 0 0 Ave ra g e p ri ce o f a ll p ro d u ct s o ff e re d 0 100 200 300
16
Samsung TV are uncorrelated. We find 27 significant cross-correlations, which results in 59% of all correlations being significant.
The other sub-question of interest gives an insight if firms charge a service premium. If firms charge a service premium, they still give a small indicator for their prices. In order to check if there is a service premium, I include a dummy variable that is true if the firm is trusted and the variable “consumer rating” in the regression. In figure 5.4 you can see how many percent each consumer rating occurred. We can see a top at a rating of 90. Almost 20 percent of all ratings were a rating of 90. There are a few people rating below 60 out of a scale 1 to 100.
Figure 5.4
Distribution of consumer ratings
0 5 10 15 20 Pe rce n t 20 40 60 80 100 rating
17
Table 5.5.
OLS regression on dependent variable “Relative Price”.
(1) (2) (3) (4)
VARIABLES Relative Price Relative Price Relative Price Relative Price
Trusted 0.0150*** 0.0162*** 0.0160*** (0.0000) (0.0000) (0.0000) Consumer Rating -0.0016*** -0.0086*** -0.0088*** (0.0000) (0.0000) (0.0000) Consumer Rating2 0.0000*** 0.0000*** (0.0000) (0.0000) Constant 0.9920*** 1.1323*** 1.3950*** 1.4158*** (0.0000) (0.0000) (0.0000) (0.0000) Observations 1,034,701 781,484 781,484 781,484 R-squared 0.0008 0.0062 0.0068 0.0058 P-value in parentheses *** p<0.01, ** p<0.05, * p<0.1
In order to test if there is a service premium charged, I include dummy variables for each month as control variables. Another set of dummies for each product is added, to correct for the fixed effects. From regression (1) in table 5.5 we can see that the value of the dummy “trusted” is 0.015 and is significantly different from zero, at all significance levels. If I include the variable consumer rating, we get the second regression. Now the dummy “trusted” has a value of 0.0162 and still significant. Also consumer rating is significant at all levels. Consumer rating is negative, with a value of -0.0016.
To check for another relationship besides a linear one between consumer rating and relative price, the quadratic version of “consumer rating” is also included. It might be the case that the impact of the consumer rating decreases as the consumer rating increases. Consumers care more about a positive rating than a negative rating, but they do not value a rating of 90 more than a rating of 70. Adding the quadratic term gives us regression (3) in table 5.5. The “trusted” dummy has not changed that much. The estimator is now 0.016 and still significant at all levels. “Consumer Rating” has a value of -0.0086 now and “Consumer Rating2” has value of 0.0000. Both are significant at the 1% level. The last regression (4) is
only regressed on the consumer ratings. We still have a negative value for consumer rating and an estimator of 0.0000 for consumer rating2. Important to note is the low R2 for all
18
regressions.
To see if there is a difference between a positive rating, a rating of 70 and above, and a negative rating with respect to the relative, I included three dummies. This can be seen in Table 5.6. Dummy 1 is true for consumer ratings between 70 and 80, dummy 2 is true for consumer ratings between 81 and 90 and dummy 3 is true for consumer ratings above 90. Again the dummies for each month and each product are included as control variables.
Table 5.6.
OLS regression on dependent variable “Relative Price”. (1)
VARIABLES Relative Price
Trusted 0.0137*** (0.0000) Rating[70-80] -0.0306*** (0.0000) Rating[81-90] -0.0400*** (0.0000) Rating[91-100] -0.0405*** (0.0000) Constant 1.0210*** (0.0000) Observations 1,034,701 R-squared 0.0058 P-value in parentheses *** p<0.01, ** p<0.05, * p<0.1
The dummy “trusted” decreased a little, it has a value of 0.0137, and is still significant at all levels. If we check the dummies for the consumer rating intervals, we find that all estimators are significant at the 1% level and are also all negative. The consumer rating between 70 and 80 has a value of -0.0306, the one between 81 and 90 has a value of -0.0400 and the one for the highest ratings is -0.0405. The values of the rating between 81 and 90 and 91 and 100 look almost the same. Testing of they are the same turns out to be that this is not the case, as they are significantly different from each other.
Bigger TV’s are most of the time more expensive than small sized TV’s. One can argue that when a product is more valuable you may care more about good delivery. If you order a
19
pair of socks online and they get lost you will not be as much upset as when a car that you ordered online arrives with scratches. From Table 5.7 we can see that there are some small differences between the groups. For all groups, the dummy “trusted” is significantly
different from zero. However, only for the big sized TV’s this estimator is negative. Also for big sized TV’s this estimator is significant at 5%, while for small sized TV’s and medium sized TV’s there is a 1% significance level. The rating estimators are more similar. They are all negative and significant at all levels, no matter which group.
Table 5.7.
OLS regression on dependent variable “Relative Price”.
VARIABLES Relative Price
Small TV's Relative Price Medium TV's Relative Price Big TV's Trusted 0.0322*** 0.0194*** -0.0035** (0.0000) (0.0000) (0.0217) Rating[70-80] -0.0643*** -0.0351*** -0.0254*** (0.0000) (0.0000) (0.0000) Rating[81-90] -0.0729*** -0.0469*** -0.0298*** (0.0000) (0.0000) (0.0000) Rating[91-100] -0.0745*** -0.0490*** -0.0266*** (0.0000) (0.0000) (0.0000) Constant 1.0126 1.0313*** 1.0299*** (0.9983) (0.0000) (0.0099) Observations 133,311 581,264 320,128 R-squared 0.1290 0.0161 0.0009 P-value in parentheses *** p<0.01, ** p<0.05, * p<0.1
There are some differences between the groups, but do they also differ within a group? Can we recognize a trend by regressing on each particular TV? Bigger TV’s are more expensive than smaller TV’s, but also within these groups there are price differences.
If we have a look at table 5.8, the table with all small TV’s, we see that the “trusted” dummy is positive and significant at the 1% level for all different inches. Again all “consumer rating” dummies are negative, except the ones that are not significant. They are also all very small. However, when regressing the 23-inch TV I find relative big estimators, compared to the others. Also the R2 is bigger than those of the other small sized TV’s.
20
In table 5.9 are the medium sized TV’s listed. These are TV’s with a size of 32 inch up to 45. The “trusted” dummy is again positive when the estimate is significant. For the 39-inch TV’s this estimate is not significant and for the 42-39-inch TV’s this dummy is even negative, but not significant. The dummies for the consumer ratings are significant at all levels, for all TV’s. They are also all negative, but one. The dummy for ratings between 70 and 80 for 39-inch TV’s is the only one being positive. This size has the fewest observations.
The big TV’s are very well represented. Table 5.10, 5.11 and 5.12 display the
regression output for the big sized TV’s. If we first focus on the “trend” dummy, it is hard to spot a trend. The estimates are flipping between being negative and positive. The bigger the TV’s the more times this dummy will have a positive sign, but still with random negatives. This dummy is significant at all levels in most cases. For 58-inch TV’s and 80-inch TV’s this dummy is not significant at all and for the 67-inch TV’s it is significant at the 5% level. The dummy for the consumer rating from 70 up to 80 seems to go from positive for the “smaller” big sized TV’s to negative for the biggest ones. Starting at the 55-inch TV’s until the 94-inch TV’s this estimator is negative only and significant at all levels. The dummy for the ratings from 81 up to 90 seems to follow the same trend. However, as the TV size increases there are still some positive values. For example, 72-inch and 69-inch are positive and significant, while in between they are negative. The dummy for the highest ratings has a stronger
pattern, like the dummy for ratings from 70 up to 80, starting with positive, significant values and ending with negative, significant values.
From the first regressions we see that there is a negative relationship between
consumer rating and relative price. By regressing on all TV’s I find that being trusted as a firm will increase the relative price, because of the positive and significant estimate. However, the R2 is around 0.02. Table 5.2 shows again a positive estimator of the “trusted” dummy.
The positive interval rating dummies show a decrease in relative price compared to a negative or absent rating. Also between the different groups of TV’s, according to their size, there is hardly any difference. The “trusted” dummy is positive for all three groups and the consumer rating dummies are all negative, for three groups. By regressing on each size of TV, the signs of the all estimators are flipping from positive to negative or vice versa.
21
6. Conclusion
Regressing the lagged prices of a product on today’s price of that product returned
significant results. For all regressions the first lag is significant at the 1% level. That means if you know yesterday’s price you can make a good estimate of today’s price of that product. The lowest estimate of the first lag is 0.715. Besides that, the seventh lag is also significant in most cases. This estimate is much smaller and reaches hardly 0.10. So this estimate is not that strong. If we look at the correlations between the products of the same firm, I find that around 60% of all the relative prices of the TV’s are correlated with each other. This is significantly more than the 5% cross-correlation you find by chance even if there is no correlation at all. Looking at the service premium, we find that when a firm is trusted the relative price will increase for a small and medium TV, but decreases for a big TV. There is no information on the website how firms get the trusted label. It might be that if you have a certain percentage of complaints of your total transactions you lose the trusted label. If this is true, then firms selling only expensive TV’s have more chance losing this label. They often make less transactions, so fewer complains are needed to lose the label compared to the other firms. One possible explanation for the negative sign for big TV’s is that non-trusted firms are using the price as a signal. They will charge even more than a trusted firm to let the consumers think that they have extraordinary services. We can also see that consumers do care about the price. A good price/quality ratio seems to be important in order to get a positive rating. A positive consumer rating, a rating of 70 and more, will decrease the relative price. When regressing on all TV’s, all the estimators are significant. However, the “trusted” dummy does not have an effect of more than 5 percentage points. Also the
consumer ratings are small. They hardly change the relative price and do not cause the price gaps. Summarizing these results, I find that web-shops are not randomizing prices and do not adopt a mixed pricing strategy. However, the sample used for this research is relatively small. There are only used three firms out if 191 firms total. The total percentage of cross-correlations are very similar across the three firms that I do not think that using more firms for this research will change the outcome.
22
7. References
AFM. Retrieved from: https://www.afm.nl/~/profmedia/files/
doelgroepen/vergelijkingssites/artikel-consumenten-vergelijkingssites.ashx on 11 April 2016.
Azar, O. H., & Bar-Eli, M. (2011). Do soccer players play the mixed-strategy Nash equilibrium?. Applied Economics, 43(25), 3591-3601.
Baye, M. R., Morgan, J., & Scholten, P. (2004). Price dispersion in the small and in the large: Evidence from an internet price comparison site. The Journal of Industrial
Economics, 52(4), 463-496.
Baylis, K., & Perloff, J. M. (2002). Price dispersion on the Internet: Good firms and bad firms. Review of industrial Organization, 21(3), 305-324.
Borenstein, S., & Rose, N. L. (1991). Competition and price dispersion in the US airline
industry (No. w3785). National Bureau of Economic Research.
Brighthub. Retrieved from: http://www.brighthub.com/electronics/home-theater/articles/67265.aspx on 15 June 2016.
Brynjolfsson, E., & Smith, M. D. (2000). Frictionless commerce? A comparison of Internet and conventional retailers. Management science, 46(4), 563-585.
Chiappori, P. A., Levitt, S., & Groseclose, T. (2002). Testing mixed-strategy equilibria when players are heterogeneous: The case of penalty kicks in soccer. American Economic
Review, 1138-1151.
Ellison, G., & Ellison, S. F. (2009). Search, obfuscation, and price elasticities on the internet. Econometrica, 77(2), 427-452.
Finance and economics: Frictions in cyberspace. Economist, 353, 8146 (November 20, 1999), 94.
Graddy, K. (1995). Testing for imperfect competition at the Fulton fish market. The RAND
Journal of Economics, 75-92.
Hackl, F., Kummer, M. E., Winter-Ebmer, R., & Zulehner, C. (2014). Market structure and market performance in e-commerce. European Economic Review, 68, 199-218.
23 Homburg, C., Koschate, N., & Hoyer, W. D. (2005). Do satisfied customers really pay more? A
study of the relationship between customer satisfaction and willingness to pay. Journal of Marketing, 69(2), 84-96.
International Telecommunication Union. Retrieved from: https://www.itu.int/en/ITU-D
/Statistics/Documents/facts/ICTFactsFigures2015.pdf on 11 April 2016.
Kaplan, G., Menzio, G., Rudanko, L., & Trachter, N. (2016). Relative price dispersion: evidence
and theory (No. w21931). National Bureau of Economic Research.
Mukamel, D. B., & Mushlin, A. I. (1998). Quality of care information makes a difference: an analysis of market share and price changes after publication of the New York State Cardiac Surgery Mortality Reports. Medical care, 36(7), 945-954.
Narasimhan, C. (1988). Competitive promotional strategies. Journal of Business, 427-449. Pan, X., Ratchford, B. T., & Shankar, V. (2002). Can price dispersion in online markets
be explained by differences in e-tailer service quality? Journal of the Academy
of Marketing Science, 30(4), 433-445.
Pepall, L., Richards, D., & Norman, G. (2011). Contemporary Industrial Organization. Danvers: Wiley.
Pereira, P. (2005). Do lower search costs reduce prices and price dispersion? Information
Economics and Policy, 17(1), 61-72.
Rosenthal, R. W. (1980). A model in which an increase in the number of sellers leads to a higher price. Econometrica: Journal of the Econometric Society, 1575-1579.
Salop, S., & Stiglitz, J. (1977). Bargains and ripoffs: A model of monopolistically competitive price dispersion. The Review of Economic Studies, 44(3), 493-510.
Shilony, Y. (1977). Mixed pricing in oligopoly. Journal of Economic Theory,14(2), 373-388. Spulber, D. F. (1995). Bertrand Competition when Rivals' Costs are Unknown. The Journal of
Industrial Economics, 43(1), 1–11.
Stock, J. H., & Watson, M. M. (2012). Introduction to Econometrics (pp. 309-312). Essex: Pearson.
24
Attachments
Table 5.2.
Cross-correlations of ten products offered by xxl-deals.de.
P-value under correlation coefficient * p<0.05 GRUNDIG-46 INCH GRUNDIG-32INCH LG-42 INCH PANASONIC-32INCH PANASONIC-42INCH PHILIPS-40INCH PHILIPS-46INCH PHILIPS-47 INCH SAMSUNG-22INCH SAMSUNG-37INCH GRUNDIG-46 INCH 1.0000 GRUNDIG-32INCH 0.3448* 1.0000 0.0011 LG-42INCH 0.2117 0.7798* 1.0000 0.0808 0.0000 PANASONIC-32INCH -0.6704* -0.7849* -0.3885* 1.0000 0.0170 0.0000 0.0034 PANASONIC-42INCH -0.3831 -0.2734* 0.1004 0.7640* 1.0000 0.0534 0.0119 0.3230 0.0000 PHILIPS-40INCH . -0.5271* 0.1552 0.9547* 0.5919* 1.0000 . 0.0000 0.2099 0.0000 0.0000 PHILIPS-46INCH 0.3308* 0.5972* 0.7937* -0.0705 -0.3400 . 1.0000 0.0063 0.0000 0.0000 0.8189 0.0827 . PHILIPS-47INCH -0.3271* -0.1915 -0.4530* -0.5275* -0.5308* -0.5410* -0.3891* 1.0000 0.0450 0.0976 0.0000 0.0016 0.0000 0.0002 0.0252 SAMSUNG-22INCH 0.2365* 0.1395 0.6216* . 0.2791 . 0.5553* -0.1586 1.0000 0.0303 0.2358 0.0000 . 0.3137 . 0.0000 0.4202 SAMSUNG-37INCH 0.1496 -0.2743* -0.5417* . 0.2471 . -0.6788* 0.0173 -0.1011 1.0000 0.1457 0.0247 0.0000 . 0.3389 . 0.0000 0.9423 0.3721
25
Table 5.3.
Cross-correlations of ten products offered by redcoon.
P-value under correlation coefficient * p<0.05 GRUNDIG-26INCH LG-37INCH LG-72INCH PANASONIC-32INCH PANASONIC-55INCH PHILIPS-26INCH SAMSUNG-32INCH SAMSUNG-40INCH SONY-40INCH TOSHIBA-40INCH GRUNDIG-26INCH 1.0000 LG-37INCH 0.2859* 1.0000 0.0000 LG-72INCH 0.2236* 0.3400* 1.0000 0.0071 0.0000 PANASONIC-32INCH -0.0423 0.6345* 0.6384* 1.0000 0.5437 0.0000 0.0000 PANASONIC-55INCH 0.3412* -0.3864* -0.2085* -0.1941* 1.0000 0.0000 0.0000 0.0128 0.0049 PHILIPS-26INCH 0.3468* -0.0074 0.0489 -0.0958 -0.1479 1.0000 0.0000 0.9297 0.5719 0.2656 0.0846 SAMSUNG-32INCH -0.1840* 0.1739* 0.1754* 0.5390* -0.0683 -0.1934* 1.0000 0.0030 0.0045 0.0349 0.0000 0.3154 0.0230 SAMSUNG-40INCH 0.1282 0.5322* -0.1508 0.0643 -0.0324 -0.2045 -0.0321 1.0000 0.2569 0.0000 0.1996 0.5733 0.7741 0.0726 0.7790 SONY-40INCH 0.0382 -0.0661 0.2887 0.4380* 0.0704 -0.2810 0.4496* 0.5341 1.0000 0.7556 0.5897 0.2306 0.0007 0.5775 0.3764 0.0001 0.2169 TOSHIBA-40INCH -0.0230 0.6005* 0.1740 0.7614* -0.2216* -0.0679 0.4253* 0.2328 0.0820 1.0000 0.7326 0.0000 0.0773 0.0000 0.0036 0.5087 0.0000 0.1429 0.5333
26
Table 5.4.
Cross-correlations of ten products offered by getgoods.de.
P-value under correlation coefficients * p<0.05 GRUNDIG-32INCH PHILIPS-32INCH PHILIPS-42INCH SAMSUNG-32INCH SAMSUNG-37INCH SAMSUNG-40INCH SAMSUNG-55INCH SHARP-60INCH SONY-55INCH TOSHIBA-55INCH GRUNDIG-32INCH 1.0000 PHILIPS-32INCH 0.3245* 1.0000 0.0009 PHILIPS-42INCH 0.1005 -0.4936* 1.0000 0.4408 0.0000 SAMSUNG-32INCH -0.1854 -0.3590* -0.2392* 1.0000 0.0735 0.0002 0.0462 SAMSUNG-37INCH 0.1256 0.5642* -0.0953 0.1533 1.0000 0.1973 0.0000 0.3066 0.0854 SAMSUNG-40INCH 0.3744* 0.6155* -0.1086 0.1037 0.2590* 1.0000 0.0027 0.0000 0.5286 0.3509 0.0029 SAMSUNG-55INCH 0.0977 0.0422 0.2449* 0.3726* 0.4889* 0.1964* 1.0000 0.3189 0.6204 0.0086 0.0000 0.0000 0.0146 SHARP-60INCH 0.3449 0.4475* -0.6413* -0.0045 0.8731* 0.8449 0.4087* 1.0000 0.0723 0.0021 0.0000 0.9842 0.0000 0.1551 0.0016 SONY-55INCH 0.3294 0.8500* -0.2271 0.0786 0.4744* -0.6295* -0.0365 0.7881* 1.0000 0.1447 0.0000 0.0712 0.6690 0.0000 0.0000 0.7445 0.0000 TOSHIBA-55INCH -0.1794 -0.5569* 0.3693* 0.3013* -0.4538* -0.6345 -0.7086* -0.6201* -0.6521* 1.0000 0.2223 0.0000 0.0017 0.0469 0.0000 0.1759 0.0000 0.0000 0.0000
27
Table 5.8.
OLS regression on dependent variable “Relative Price” with only small sized TV’s.
VARIABLES 19 Inch 22 Inch 23 Inch 24 Inch 26 Inch
Trusted 0.0380*** 0.0458*** 0.0134*** 0.0048*** 0.0267*** (0.0000) (0.0000) (0.0000) (0.0053) (0.0000) Rating[70-80] 0.0013 -0.0371*** -0.1937*** -0.0073* -0.0933*** (0.7393) (0.0000) (0.0000) (0.0858) (0.0000) Rating[81-90] 0.0012 -0.0565*** -0.1575*** -0.0634*** -0.0960*** (0.6473) (0.0000) (0.0000) (0.0000) (0.0000) Rating[91-100] -0.0222*** -0.0719*** -0.1744*** -0.0588*** -0.0791*** (0.0000) (0.0000) (0.0000) (0.0000) (0.0000) Constant 0.9968*** 1.0125 1.1755*** 1.0418*** 1.0004*** (0.0000) (0.9996) (0.0000) (0.0000) (0.0000) Observations 13,216 43,858 5,660 11,624 58,704 R-squared 0.0425 0.1236 0.4554 0.0871 0.1668 Table 5.9.
OLS regression on dependent variable “Relative Price” with only medium sized TV’s.
VARIABLES 32 Inch 37 Inch 39 Inch 40 Inch 42 Inch
Trusted 0.0308*** 0.0220*** 0.0017 0.0162*** -0.0015 (0.0000) (0.0000) (0.6022) (0.0000) (0.4208) Rating[70-80] -0.0361*** -0.0303*** 0.0168*** -0.0114*** -0.0308*** (0.0000) (0.0000) (0.0069) (0.0000) (0.0000) Rating[81-90] -0.0475*** -0.0394*** -0.0332*** -0.0272*** -0.0359*** (0.0000) (0.0000) (0.0000) (0.0000) (0.0000) Rating[91-100] -0.0518*** -0.0460*** -0.0210*** -0.0343*** -0.0244*** (0.0000) (0.0000) (0.0000) (0.0000) (0.0000) Constant 1.0202*** 1.0185*** 1.0151*** 1.0130*** 1.0200*** (0.0000) (0.0000) (0.0000) (0.0000) (0.0000) Observations 240,786 89,841 2,940 148,164 109,823 R-squared 0.0853 0.0600 0.0407 0.0062 0.0022 P-value in parentheses *** p<0.01, ** p<0.05, * p<0.1
28
Table 5.10.
OLS regression on dependent variable “Relative Price” with only big sized TV’s.
VARIABLES 46 Inch 47 Inch 48 Inch 50 Inch 52 Inch 55 Inch
Trusted -0.0398*** -0.0200*** 0.0829*** -0.0708*** -0.0627*** 0.0518*** (0.0000) (0.0000) (0.0000) (0.0000) (0.0000) (0.0000) Rate[70-80] 0.0233*** -0.0333*** 0.0295*** 0.1610*** 0.0248 -0.0803*** (0.0000) (0.0000) (0.0022) (0.0000) (0.5452) (0.0000) Rate[81-90] 0.0163*** -0.0080 0.0766*** 0.0086 0.0617*** -0.0967*** (0.0000) (0.1093) (0.0000) (0.6963) (0.0010) (0.0000) Rate[91-100] -0.0040 0.0289*** 0.0348*** 0.0548*** 0.0596 -0.0996*** (0.1226) (0.0000) (0.0000) (0.0053) (0.1353) (0.0000) Constant 1.0105 1.0300** 0.9225*** 1.0491* 0.9773*** 0.9460** (0.9996) (0.0202) (0.0000) (0.0640) (0.0000) (0.0116) Observations 121,568 76,037 301 11,081 1,135 74,606 R-squared 0.0038 0.0017 0.5467 0.0044 0.0346 0.0178 P-value in parentheses *** p<0.01, ** p<0.05, * p<0.1
29
Table 5.11.
OLS regression on dependent variable “Relative Price” with only big sized TV’s.
VARIABLES 58 Inch 60 Inch 65 Inch 67 Inch 69 Inch 70 Inch
Trusted -0.0081 0.0421*** 0.0449*** 0.0632* 0.1573*** -0.0359*** (0.2542) (0.0000) (0.0000) (0.0949) (0.0000) (0.0000) Rating[70-80] -0.1457*** -0.0875*** -0.1055*** (0.0000) (0.0000) (0.0000) Rating[81-90] 0.0112 -0.1427*** -0.0805*** -0.0016 0.4180*** -0.1542*** (0.4672) (0.0000) (0.0000) (0.9648) (0.0000) (0.0000) Rating[91-100] -0.1274*** -0.1123*** -0.0892*** -0.1337*** (0.0000) (0.0000) (0.0000) (0.0000) Constant 1.0289*** 0.9992*** 1.0270*** 0.9545*** 1.0000*** 1.1332*** (0.0000) (0.0000) (0.0000) (0.0000) (0.0000) (0.0000) Observations 88 16,545 2,823 15 380 4,310 R-squared 0.7036 0.0888 0.2408 0.2559 0.6638 0.3035 P-value in parentheses *** p<0.01, ** p<0.05, * p<0.1
30
Table 5.12.
OLS regression on dependent variable “Relative Price” with only big sized TV’s.
VARIABLES 72 Inch 75 Inch 80 Inch 84 Inch 90 Inch 91 Inch 94 Inch
Trusted 0.1010*** -0.3167*** -0.0049 -0.0001*** 0.3527*** 0.1043*** 0.0377*** (0.0000) (0.0000) (0.3111) (0.0030) (0.0000) (0.0000) (0.0081) Rating[70-80] -0.1311*** -1.3721*** -0.1593*** (0.0000) (0.0000) (0.0000) Rating[81-90] 0.0566*** -1.3961*** -0.0942*** 0.0001*** -0.0052 -0.0151 0.0954*** (0.0000) (0.0000) (0.0000) (0.0000) (0.1947) (0.3743) (0.0000) Rating[91-100] -0.0259*** -0.7108*** -0.1478*** 0.0006 -0.0564*** (0.0000) (0.0000) (0.0000) (0.8920) (0.0002) Constant 0.9349*** 1.9796*** 1.0549*** 0.9999*** 0.6467*** 0.9905*** 0.9483*** (0.0000) (0.0000) (0.0000) (0.0000) (0.0000) (0.0000) (0.0000) Observations 869 171 1,117 32 812 7,762 18 R-squared 0.2482 0.3258 0.3820 0.8604 0.7511 0.0129 0.8578 P-value in parentheses *** p<0.01, ** p<0.05, * p<0.1
