Is The Daily Deal Social Shopping?

: An Empirical
Analysis of Purchase and Redemption Time of
Daily-Deal Coupons

Minjae Song

Eunho Park

Byungjoon Yoo

Seongmin Jeon

September 2012
Abstract
Shortly after Groupon started its business in 2008, daily-deal sites became new
shopping places for online shoppers. Daily deals are oered in a coupon format with
huge price discounts and have social shopping features that involve collective actions.
In this paper we use individual-level panel data from one of the major daily-deal sites in
Korea to analyze how online shoppers change their purchase and redemption behaviors
as they repeat purchases and redemptions. We nd evidence that social shopping
features deter inexperienced shoppers from buying deals early on. We also nd evidence
that inexperienced shoppers drive congestion at participating businesses.

We are grateful to participants at the 2012 Marketing Science Conference for their comments. All errors
are ours.

Corresponding author. Assistant Professor of Economics and Marketing at the Simon Graduate School
of Business, University of Rochester, minjae.song@simon.rochester.edu.

Graduate Student, College of Business Administration, Seoul National University, marrno@snu.ac.kr

Associate Professor of Management Information System, College of Business Administration, Seoul Na-
tional University, byoo@snu.ac.kr

Ph.D. Candidate, College of Business Administration, Seoul National University, smjeon@yahoo.com

1
1 Introduction
In 2008 Groupon, an internet startup company, experienced an instant success by selling
daily deals with heavy discounts, and shortly afterwards daily-deal sites became new shop-
ping places for online shoppers. Daily deals are oered in a coupon format and have social
shopping features that involve collective actions among shoppers; many daily-deal sites in-
cluding Groupon require a certain number of shoppers to commit to purchase to make deals
active. If a deal does not attract the minimum number of shoppers, it does not go on sale
and those who are already committed receive their money back. If it succeeds in attracting
enough shoppers, the deal is tipped, which means it has reached a tipping point whereby
the minimum number of consumers have committed to purchase. For example, if the tipping
point is 100, the deal is not on sale until at least 100 consumers are committed to buy.
1
Once
a deal has tipped, all buyers, whether they buy before or after it is on, enjoy the same price
discount and have the same time window to redeem it.
Daily-deal sites provide platforms that bring merchants, especially small businesses,
and online shoppers together. On its website Groupon describes its business model as follows.
Groupon negotiates huge discountsusually 50-90% owith popular businesses. We send
the deals to thousands of subscribers in our free daily e-mail, and we send the businesses a
ton of new customers. Thats the Groupon magic. Daily-deal sites make money by keeping
approximately half the money customers pay. For example, suppose a daily-deal site oers
an $80 massage at 50 percent discount and equally splits the revenue with the merchant.
The consumer pays $40 to the daily-deal site and gets a massage valued at $80 from the
massage shop, while the shop gets $20 from the daily-deal site.
Daily-deal sites, however, provide more than platforms for merchants and customers
to interact. Because of the minimum purchase feature, a consumer cannot buy a deal that
she likes unless other consumers also buy it. This means that consumers depend on one
1
We use consumers, customers, and (online) shoppers interchangeably.
2
another in their purchases, which makes buying daily deals social shopping.
2
Wikipedia.org
categorizes the daily-deal site as one of the major social shopping methods, where social
shopping is dened as a method of e-commerce where shoppers friends become involved in
the shopping experience.
3
This aspect distinguishes the daily-deal site from other discount
sites such as Restaurant.com that just function as platforms in the two-sided market context.
This social shopping or group shopping feature may result in, as a by-product, un-
pleasant redemption experiences as consumers try to redeem their coupons at the same time.
There are anecdotes about retailers complaining about dealing with too many customers right
after deals are sold and consumers complaining about receiving poor quality service due to
too many customers. For example, a person purchased a voucher for a spa from Groupon
and went to book the treatment, but was told that there were no more spots left as too many
people had already scheduled appointments. She could not use the voucher after all.
4
One may argue that rational individuals should not care how many people are com-
mitted to buy when they make purchase decisions because they receive a full refund if deals
fail to attract the minimum number of purchases. On the other hand, consumers may enjoy
contributing to tipping deals. They may send a link to their friends or even buy deals that
they are not excited about otherwise. This would be what daily-deal sites expected to hap-
pen when they adopted the minimum purchase feature. It is also plausible, however, that
consumers may want to wait and see if deals are tipped to avoid wasting time, although it
only takes a few clicks to buy deals.
In this paper we use individual-level data from one of the major daily-deal sites
in Korea to analyze how online shoppers change their purchase and redemption behaviors
as they become more experienced. In particular, we test whether their purchase and
redemption times change as they buy more deals and redeem more coupons. For the purchase
time we analyze over 370,000 purchases on about 500 deals sold in a three month period.
2
LivingSocial, another major daily-deal company, does not have the minimum purchase feature but gives
deals for free when consumers send a link to contacts and three of them buy the oer through the link.
3
See http://en.wikipedia.org/wiki/Social shopping for details.
4
http://groupon-reviews.measuredup.com/Complaint-Voucher-Redemption-40928
3
For the redemption time we analyze over 250,000 redemptions for the deals sold during that
three month period. For each individual we know how many deals she bought and redeemed
up to the sample period and what deals she bought during the sample period. For the deals
she bought during the sample period we observe at what time she bought them, whether
they were tipped at the time of purchase, and when she redeemed them.
Since about 40 percent of consumers bought multiple times during the sample period,
we can control for each consumers purchase and redemption experiences separately from
their characteristics (such as gender, age, location, etc.) and various deal characteristics.
Using variables that account for the number of past purchases and past redemptions, we
analyze whether and how online shoppers past experiences aect their current purchase and
redemption times.
We nd that typical consumers in their rst few purchase occasions tend to buy deals
that are already tipped. Experienced shoppers are more likely to buy deals before they are
tipped but we do not nd evidence that they do so to contribute to tipping deals. They
rather tend to buy deals right after they are announced. We also nd that typical consumers
redeem their rst few coupons right after their purchases but tend to wait longer as they
redeem more coupons.
Our ndings on the purchase time suggest that the social shopping feature of the daily
deal deters inexperienced shoppers from buying deals early on, rather than encouraging them
to do so. As they buy more deals, however, they seem to gure out that there is no cost
associated with buying deals before they are on and, hence, buy deals as early as they can.
Our ndings on the redemption time provide evidence that inexperienced shoppers drive
congestion at participating businesses right after deals are sold but try to avoid crowds in
their next redemptions. When all these ndings are put together, it is hard to claim that
the success of the daily-deal business is due to the social shopping feature. Rather, it causes
problems for merchants who have to handle an unusually large number of customers after
deals are sold.
4
Our paper is the rst to analyze daily-deal shoppers behaviors using individual-level
data and the very rst paper to analyze their redemption behaviors. Existing studies on
daily deals use data compiled from monitoring daily-deal websites, so their analyses are
limited to purchase patterns across deals. Ye, Wang, Aperjis, Huberman, and Sandholm
(2011), for example, use data obtained from the websites of Groupon and LivingSocial and
compare the dynamics of purchasing times on these two deal sites. Byers, Mitzenmacher,
and Zervas (2011) use similar data to analyze the relationship between deal sales and deal
characteristics. An exception is Dholakia and Kimes (2011), who use survey data from daily-
deal users as well as non-users, but they focus on examining consumer perceptions of daily
deals. Subramanian (2012) develops an analytical model of the daily deal where real-time
information on ongoing coupon sales inuences consumers belief about the quality of a deal
and consumers can strategically wait to check this information before they buy. He analyzes
a situation where setting up the tipping point hurts the daily-deal business.
There are also studies that focus on whether merchants benet from selling goods
and services on daily-deal sites. Kumar and Rajan (2012) use an analytical model and data
from merchants to show that oering deals do not necessarily yield prots for merchants.
Edelman, Jae, and Kominers (2011) use an analytical model to show that oering deals is
more protable for merchants who are relatively unknown or merchants with low marginal
costs. Dholakia (2011) examines the determinants of protable Groupon promotions using
survey data from merchants.
The rest of the paper is organized as follows. Section 2 describes the data set. Section
3 analyzes purchase time and Section 4 analyzes redemption time. Section 5 concludes.
2 Industry and Data
While a number of new daily-deal sites such as WagJag and BuyWithMe have entered the
market since Groupons initial success, Groupon has been dominant in the US market and
5
has expanded to more than 40 countries around the world. Groupons remarkable growth
continues today. In the rst quarter of 2011, it made over $644 million revenue, which is
about 200 times larger than its revenue in the same quarter in 2009. Groupons subscriber
base grew from about 2 million in the beginning of 2010, to over 83 million by March 2011.
5
The Groupon-like daily deal business did not exist in Korea until March 2010 when
a start-up company launched a daily-deal site following Groupons business model.
6
Shortly
afterwards a series of daily-deal sites quickly sprang up with the identical business model.
As of December 2011 there are over 300 competing daily-deal sites in Korea with industry
annual sales reaching KW 500 billion ($470 million) by the end of 2011. The top three daily-
deal sites dominate the market with their combined sales reaching KW 45 billion ($42.21
million).
7
According to a report by Rankey.com, these three sites are listed among the top
10 e-commerce websites in Korea as of August 2011.
We have detailed information on deals as well as registered customers from one of
the major daily-deal sites for the rst year of its business. The data set includes over 3,000
deals that generated almost 3 million coupon sales in total from over 1 million registered
customers. From the entire data, we limit our attention to deals sold between October 1,
2010 and December 31, 2010. During this period 679 deals were sold to 253,151 customers
and the total number of coupons sold is 508,593. Since we have data from one company, we
avoided a period when the market is competitive. Our sample period is before the daily-deal
market exploded in Korea. In August 2010 there were only about 30 daily-deal sites and the
company we have data on was dominating the market. However, we also avoided the rst
six months after the industry was born in order to obtain a large sample size of repeating
purchases and redemptions.
Table 1 summarizes the characteristics of the 253,151 consumers who bought deals
5
See Groupons S-1 ling with the SEC at http://sec.gov. The article by Amy Lee on June 2, 2011 at
the Hungton Post provides a nice summary. See http://www.hungtonpost.com/2011/06/02/groupon-
ipo n 870652.html
6
http://www.wikitree.co.kr/main/news view.php?id=3801
7
http://www.cnngo.com/seoul/shop/social-commerce-sites-869074
6
during the three-month period. About 65 percent are female and the median age is 30. The
age and gender distribution, not reported in the paper, shows that people in their twenties
and thirties are major customers of this site, and among those in their twenties the number
of females is almost twice as large as that of males. Over 50 percent of the shoppers in our
sample receive either mobile texts (SMS) or e-mails on daily deals from the site.
8
Although
not reported in the table, we also know when they made an account on this deal site and
which regions are their preferred markets.
9
The median number of purchases during the
sample period is 1, and about 40 percent of shoppers purchased more than 1 deal and about
10 percent purchased more than 3 deals.
10
The distribution of purchase frequency is heavily
skewed and has a long right tail. Out of the 10 percent who purchased more than 3 deals,
about 40 percent bought more than 5 deals and about 7 percent bought more than 10 deals.
11
In addition to knowing how many deals were purchased, we also know when they were
bought. Thus, given a deal, we know how many deals those shoppers bought from the time
they made accounts to the date of that deal, and we use this information to account for their
purchase experiences. In the beginning of the sample period about a quarter of consumers
bought at least once previously. We also know what time they bought a deal, so we can
determine whether they bought before a deal was tipped. In the following section we link
this information to consumers purchase experiences and test whether their purchase time
changes as they buy more deals. In addition, we have information on when they redeemed
the deals that they bought. Although each purchase almost always leads to redemption,
consumers have dierent patterns of redemption. For example, some redeem their coupons
immediately after their purchases, while others wait for several weeks before redeeming them.
Table 2 summarizes the characteristics of 679 deals in our sample. Product Sales is
the number of coupons sold per each deal. The mean is 1,590 with the standard deviation
8
The company sends emails at 7 oclock in the morning.
9
Consumers indicate which region is their preferred market but do not always buy in their preferred
markets.
10
If a consumer buys two dierent deals in two dierent markets, we count it as two purchases. But if she
purchases multiple coupons of the same deal, we count it as one purchase.
11
For the 1 year we have data on the median number of purchases is 4 and the mean is 6.67.
7
3,588. The deal with the highest sales sold 49,995 coupons. Original Price is the original
price of deals and ranges from 700 KW to 1 million KW with the median 35,000 KW.
12
This large variation in prices is mainly due to having very dierent products. Discount
Rate refers to the discount rate and varies widely, but over 95 percent of deals oer at least
50 percent discount. The price consumers actually pay (the original price multiplied by the
discount rate) ranges from 0 to 599,000 KW with the median price 14,900 KW. Sale Period
denotes the number of days that a deal is available. A majority of deals are on sale for 24
hours but some deals, mainly those that are put up on Fridays, are sold for multiple days.
Tipping Point is the number of sales required to turn on a deal. The median tipping point is
100 but as low as 1 and as high as 5,000. Maximum Sale is the number of coupons available
for each deal and ranges from 30 to 60,000 with the median 1,500. Both the tipping point
and the maximum sale are predetermined and revealed to consumers when deals appear on
the site. Redemption Period refers to the number of days that purchasers have for redeeming
coupons they bought. This period is usually 60 or 90 days. Deals with zero redemption days
are concerts or performances on xed dates. Soldout indicates whether a deal was sold out,
meaning the number of coupons sold reached the maximum sale. About 30 percent of the
deals in our sample were sold out. Figure 1 shows deal types in a pie chart. Restaurants
make up the largest category, followed by beauty/healthcare, concert/exhibition and so on.
The deal site sets the discount rate, the tipping point, and the redemption period,
presumably jointly with vendors, to sell all coupons available. A simple correlation among
them, however, does not show any clear pattern in how they set these values. The most
correlated characteristics are the tipping point and the maximum sale. They are positively
correlated with the correlation coecient 0.60. The original price is positively correlated
with the discount rate with the correlation coecient 0.31, suggesting that more expensive
products are usually more heavily discounted. The tipping point is negatively but weakly
correlated with the discount rate with the correlation coecient -0.11. Although it is not
12
1,200 Korean Won is approximately equal to one US dollar.
8
high, this suggests that a heavier price discount tends to accompany a lower tipping point.
The correlation between the tipping point and the original price is even weaker with the
correlation coecient -0.04.
To see whether or how these attributes aect coupon sales, we run a simple OLS with
the log of sales as the dependent variable and all other attributes as covariates. In addition
to the variables reported in Table 2, we include the dummy variables for each day of the
week, 6 category dummy variables, and 13 region dummy variables. We rst run OLS for all
deals, and then selected two subsamples and run OLS for each sample. The rst subsample
excludes 173 deals that lasted for longer than 24 hours, and the second subsample excludes
206 deals that were sold out.
Table 3 reports results for these three regressions. Results are very similar across the
three samples. The adjusted R-squared is 0.61 for all deals, which is high considering the
fact that every deal is unique. The log of the (original) price variable has a negative and
statistically signicant coecient, suggesting that more expensive products sell less. The
discount rate is positive and statistically signicant, suggesting that a deal with a heavier
discount is more popular. The tipping point is also positive and statistically signicant,
suggesting that a deal with a higher tipping point is more popular. The maximum sale is
positive and signicant, indicating the deal site sets a higher maximum sale for more popular
deals. The positive coecients for the tipping point and the maximum sale suggest that the
deal site sets them at higher values for deals that they expect to be popular. The redemption
period variable has a positive and signicant coecient, showing that a deal with a longer
redemption period is more popular. D salesday is a dummy variable for deals that lasted
longer than a day, and has a positive and signicant coecient, showing that a deal that
lasted longer sold more coupons.
The day variables do not show any signicant eects. Deals sold on Wednesdays seem
more popular, but its eect is not signicant. The insignicant day eects are consistent
with what the deal site describes about their business during the sample period: they did
9
not have a large inventory of deals to select for each day. They admit that the hardest
part in an earlier period was to convince vendors to sell their products as the daily deal
business was new and the vendors margin was extremely low.
13
This means that the deal
site did not have much freedom to select more appropriate days for each deal. The category
variables do not show much signicant eects other than positive eects for restaurants and
concerts/exhibitions.
In analyzing purchase time in the following section we use deals that were sold for 24
hours but do not exclude deals that were sold out. 506 deals satisfy these criteria, and the
number of consumers goes down to 195,553 but the consumer characteristics hardly change.
The results in Table 3 show that the relationship between the coupon sales and the deal
characteristics does not change substantially with these criteria. The most notable change is
seen in the category dummy variables; three category dummy variables including restaurants,
beauty/health care, and concerts/exhibitions have larger coecients, and the pub and cafe
category variables now have positive yet insignicant coecients.
3 Purchase Time
The feature of the daily-deal site that makes it social commerce or social shopping is the
(predetermined) minimum number of purchase commitments, which is called the tipping
point. If a deal does not attract a certain number of committed buyers, consumers cannot
buy it no matter how much they like it. One of our main research questions is how consumers
behave towards the tipping point. Given that consumers nd deals that they like, are they
more likely to purchase before deals have tipped? In particular, we link consumers purchase
behaviors to their purchase experiences and test if their purchase time changes as they buy
more deals. For this we analyze 370,616 purchases made by 195,553 consumers on 506 deals.
To control for purchase experiences we create ve dummy variables for the rst pur-
chase through the fth purchase. Given a deal the dummy variable for the rst purchase
13
Due to a condentiality agreement we cannot disclose the names of the people we talked to.
10
takes 1 if the number of previous purchases is 0, the dummy variable for the second purchase
takes 1 if the number of previous purchases is 1, and so on. These ve dummy variables
cover about 84 percent of all observations. Note that these dummy variables control for
the number of past purchases from the date that consumers made an account. For previous
redemptions we create three dummy variables for the rst two redemptions. Given a deal
the dummy variable for no redemption takes 1 if a consumer did not redeem any coupon up
to that point of time, the dummy for one redemption takes 1 if she redeemed one coupon
before, and so on. These three dummies cover about 90 percent of all observations.
Figure 2 shows an empirical distribution of purchase times for all deals that were sold
for 24 hours. The gure shows that 9 to 10 am has the highest frequency, about 11 percent,
closely followed by midnight to 1 am. 10 to 11 am has the third highest peak followed by 11
am to the noon. The frequency is quite evenly distributed in the afternoon hours and is lower
in the evening. Not surprisingly, 3 to 8 am has the lowest frequencies. In order to see who
buys at what time and if purchase times change as consumers buy more deals, we rst run the
multinomial logit regression where we divide 24 hours into 4 equal length intervals (6 hour
intervals) and use the second interval (from 6 am to the noon) as the base. Table 4 reports
results. In addition to the variables reported in the table, we include the deal characteristics
reported in Table 2, deal categories, weekday dummies, 13 dummies for market locations,
12 dummies for weeks that deals were sold, and 32 dummies for weeks that consumers made
accounts. The 12 week dummies control for any time trend in purchase behaviors during
the sample period, and the 32 week dummies control for consumer heterogeneity associated
with the timing of becoming customers at the deal site.
Results show that most of the purchase/redemption experience dummies are statis-
tically signicant. To interpret these estimates we compute the predicted probability of
buying the average deal in each time interval for the average consumer. For the average
buyer who buys a deal for the rst time the probability of buying it in each time interval
is 0.0000052 for the rst 6 hours of the day (from midnight to 6 am), 0.67 for the next 6
11
hours, 0.12 for between noon and 6 pm, and 0.22 for the last 6 hours. When this consumer is
buying the second deal, the probability goes up by 21 percent for the rst time interval and
11 percent for the second time interval, while it goes down by 17 percent for the third interval
and 25 percent for the fourth interval. In other words, given that the average consumer is
buying the second deal, she is more likely to buy it in the rst 12 hours of the day (between
midnight and the noon) and less likely to buy it in the last 12 hours of the day compared
to her last purchase. A purchase-time change is more dramatic when she is buying her sixth
(or higher) deal. The probability of buying a deal in the rst 6 hours of the day is three
times higher compared to her rst purchase while the probability of buying it in the last 12
hours is more than 40 percent lower.
One concern here is, however, that we do not perfectly control for individual dier-
ences other than with a few customer characteristics and the 32 dummies for weeks that
they made an account at this deal site. It is possible that the result above is mainly driven
by those who bought no more than one or two deals by the end of the sample period; if
those who bought only one deal bought it in the afternoon, the rst purchase dummy may
not distinguish the rst purchase eect from the eects of (unobserved) individual charac-
teristics that made them buy only once. Because of this concern, we also run the xed-eect
regression with purchase times as the dependent variable. In particular, we divide 24 hours
into 2,400 units and transform purchase times into these units.
14
In Table 5 we compare the
xed eect regression results with the OLS results. The OLS results show that the purchase
time of the rst purchase is about one hour later than that of the second purchase and about
one hour and thirty minutes later than that of the third purchase. When consumers make
their sixth (or higher) purchase, they are likely to buy deals three hours earlier than their
rst purchase. The xed eect results show that the eects of previous purchases become
smaller but are still consistent with those of the OLS results. Now the purchase time of the
14
This linear regression may not be the best choice because the dependent variable is bounded and there
are a lot of observations near the bounds, but it may still show how results change with the individual xed
eects.
12
rst purchase is about 40 minutes later than that of the second purchase and 1 hour and
30 minutes later than that of the sixth purchase, but the results still suggest that the more
deals consumers have purchased in the past, the earlier they buy the next deal.
Why would consumers buy earlier as they buy more deals? Put dierently, why
would those who buy for the rst time or the second time buy in the afternoon or later? One
implication from the purchase time results is that consumers who are buying for the rst
time or who have only bought a couple of deals want to wait until deals have tipped. Figure
3 shows when deals reach the tipping point. The gure shows that about 40 percent of deals
reach the tipping point in the rst hour, but about 20 percent of deals reach it between
7 and 10 oclock in the morning and over 10 percent of deals reach it after 10 oclock in
the morning. It may well be that those consumers do not want to buy deals that have not
attracted enough consumers.
We rst test if the purchase time around the tipping point is associated with the
purchase experience. In particular, we estimate the probability of purchasing a deal before
it has tipped with the same regressors as in the multinomial logit regression above. Table 6
reports results from the binary logistic regression as well as the xed eect logit regression.
Note that the number of observations goes down from 370,616 to 71,402 in the xed eect
logit regression as it only uses consumers that change their purchase times over dierent deal
purchases. With the xed eect logit regression we can control for unobserved individual
heterogeneity, although we cannot learn much about the average partial eect.
In the binary logistic regression all ve of the previous purchase dummies are statisti-
cally signicant and negative and their absolute magnitude goes down with more purchases,
indicating that the more deals consumers bought previously, the more likely they are to buy
before deals reach the tipping point. In particular, compared to those who are buying their
rst deal, the probability of buying a deal before it reaches the tipping point is about 30
percent higher for those who bought one deal before, about 68 percent higher for those who
bought two deals before, and about three times higher for those who bought ve deals be-
13
fore.
1516
These eects are much larger than the eects of any consumer characteristics such
as age, sex, and whether they are receiving emails. In the xed eect logit regression the
coecients of the previous purchase dummies are lower in the absolute term but the relative
magnitude among the purchase experience dummies does not change much, indicating that
individual dierences do matter but the main implication still holds.
These results show that inexperienced shoppers tend to wait until deals have tipped
while experienced ones buy before they reach the tipping point. One may interpret these
results as evidence that there is not much group shopping or social shopping going on with
inexperienced shoppers. The idea of having the tipping point is that it may trigger social
interaction among online shoppers; if I see a deal that I like, I may call up my friends to join
me buying it. However, the results do not seem consistent with such a behavior at least for
inexperienced shoppers.
These results also suggest that experienced buyers are the main driving force behind
tipping daily deals; they may receive so much utility from turning deals on that they even
buy deals that they would not normally purchase if they have already tipped. Instead of the
dummy variables, we also use the number of previous purchases and the number of previous
purchases squared in the binary logistic regression and nd that this eect persists up to 42
previous purchases which covers almost everyone in the sample. However, one may interpret
the same results as consumers learning that they do not lose a penny even if a deal they
are committed to buy fails to be on; all they lose is a few minutes spent on the deal site.
In this interpretation experienced shoppers early-purchase behavior is not associated with
social shopping.
In order to shed more light on motives behind the early purchase, we analyze purchases
before deals reach the tipping point. In particular, we estimate the probability of being part
of the last 10 percent of purchases before deals have tipped as well as the probability of being
15
As above we compute the probability of the average consumer buying the average deal.
16
The probability of buying a deal before it reaches the tipping point is 0.049 for consumers buying their
rst deal.
14
part of the very rst 10 percent of purchases. Suppose a deal needs 100 purchases to be on,
and when a customer visits the deal site, only 5 customers are committed to buy.
17
If she
receives a utility from tipping a deal, her purchase has a marginal impact as compared to
when 95 or 99 customers are committed to buy. On the other hand, if she does not receive
such a utility and does not want to buy a deal that is unlikely to tip, she is more likely to
buy when 95 customers are committed to buy.
We use the binary logistic regression as well as the xed eect logit regression with
the same regressors as in Table 6. Because we only use purchases before deals have tipped,
the number of observations goes down to 32,076. Results are reported in Table 7. On the
left panel (under Last 10%) we estimate the probability that purchases are part of the last
10 percent before deals reach the tipping point, and on the right panel (under First 10%)
we estimate the probability that purchases are part of the rst 10 percent after deals appear
on the site. The binary logistic regression results show that those who bought many deals
in the past are less likely to buy right before deals have tipped but are more likely to buy
right after deals appear than those who bought no deal or a couple of deals in the past. In
the xed eect logit regression most of the past purchase dummies are no longer statistically
signicant, but the dummy for the rst purchase is still signicant and even has a larger
coecient than in the binary logistic regression, implying that the dierence between those
who bought many deals in the past and those who bought no deal in the past does not stem
from individual heterogeneity.
These results show that the more deals consumers buy, not only are they more likely
to buy before deals are turned on but also more likely to buy right after deals start. This
implies that those frequent buyers buy deals that they like without paying much attention
to whether deals are going to be on. The typical consumer implied by our results tends to
wait until a deal has tipped in their rst few purchases but after that she pays less attention
to the deal status and buys deals that she likes. It is important to note that our results
17
The deal site we obtain data from shows the number of purchase in real time, while Groupon in the US
shows a rough status like over 530.
15
show strong state dependence in purchase behaviors. In other words, our results suggest
that consumers do change their purchase behaviors as they buy more deals.
One may interpret these results in the context of observational learning (Banerjee,
1992; Cai, Chen and Fang, 2009); shoppers learn about the quality of deals from others. In
particular, Subramanian (2012) shows that in such a situation shoppers are willing to wait
until a sucient number of coupons are sold. In this perspective our results suggest that
inexperienced buyers are those who wait and learn from others while experienced ones do
not care much about whether other shoppers like a deal or not. However, the observational
learning motivation does not explain why experienced buyers no longer care about other
shoppers decisions. The deal site oers dierent deals every day, so buying deals in the past
does not necessarily provide more information on a deal oered today.
4 Redemption Time
Once a deal has tipped, everyone who buys it has the same amount of time to redeem their
coupons. The redemption period ranges from 50 to 100 days for about 60 percent of deals in
our sample and the median is 90 days. Because of this limited time window, businesses who
successfully sold many coupons on the daily-deal site serve an unusually large number of
customers, and sometimes have diculty in satisfying customers expectation. For example,
consumers may have to wait for hours to be seated at restaurants. This congestion is an
inevitable cost of daily deals and presents a challenge to the industry unless it nds a way
to coordinate consumers redemption times.
18
In this aspect it is important to know when consumers redeem the coupons and if
their redemption time changes as they redeem more coupons. In Figure 4 we select 121
deals that have a 90 day redemption period and show the frequency of redemption time
18
Margie Fishman writes in her blog at www.openforum.com on September 19, 2011 that Living-
Social staers assist small business owners during peak redemption periods, such as right after a
coupon is released or just before its expiration date, but does not explain how they do it. See
http://www.openforum.com/articles/the-top-5-group-buying-sites-love-em-or-hate-em
16
for consumers who bought those deals. The gure shows that a large number of customers
redeem their coupons right after purchases and right before they are expired. 28 percent of
coupons were redeemed in the rst 10 days, 44 percent were redeemed in the rst 20 days,
and 14 percent were redeemed in the last 10 days.
For the redemption-time analysis we select deals whose redemption period is at least
20 days long and track redemptions for those deals. With this criterion deals that can
be redeemed only on xed dates such as concerts or plays are excluded. The new sample
includes 267,143 observations and their redemption times range from 1 to 183 days where
1 day means redeeming a coupon on the next day of purchase. Also, note that redemption
periods for many deals in our sample, at least for deals sold in the last 20 days of our sample
period, end later than the end of our sample period.
We rst estimate the probability of redeeming coupons in the rst 10 days of the
redemption period using the binary logistic regression and the xed eect logit regression,
and report results on the left panel of Table 8. We use the same regressors as in the previous
section, but the dummy variables for days now represent days that coupons were redeemed,
not days that deals were purchased. Also, the dummy variable for the rst purchase is taken
out because all consumers in this sample inevitably made at least one purchase in the past.
To avoid confusion, we re-label the previous purchase dummies as One Purchase instead of
Second Purchase, Two Purchasesinstead of Third Purchase, etc. One Purchasemeans
that consumers made one purchase up to that point, Two Purchasesmeans that consumers
made two purchases up to that point, etc. We also re-label the previous redemption dummies
as First Redemption instead of No Redemption, Second Redemption instead of One
Redemption, etc.
Results show that consumers are less likely to redeem in the rst 10 days as they
redeem more. In the binary logistic regression the probability of redeeming the average
coupon in the rst 10 days is 0.32 for the average consumer who is redeeming for the rst time.
This probability goes down to 0.17 for those redeeming for the second time and 0.09 for those
17
redeeming for the fourth time.
19
In the xed eect logit regression the number of observations
goes down to 96,892 because consumers who always redeemed in the rst 10 days or always
redeemed later than that are not used. Results show that the redemption experience eects
are now much larger. One implication of these results is that experienced consumers try to
avoid the crowds that made their visits unpleasant in the previous redemption occasions.
Similarly, we also estimate the probability of redeeming coupons in the last 10 days
of the redemption period and report results on the right panel of Table 8. Results show
that consumers are more likely to redeem their coupons in the last 10 days as they redeem
more, and the eects become larger in the xed eect logit regression. We may interpret
this result as experienced consumers waiting to redeem their coupons as much as possible,
but it is also possible to interpret it as their procrastinating using coupons or having too
many coupons to redeem on time.
20
In addition to estimating the probability model, we also run linear regressions with
the redemption time as the dependent variable. Table 9 shows both the OLS and the FE
regression results. The OLS results show that consumers who are redeeming for the rst time
redeem about 12 days earlier than those who are redeeming for the second time and about
30 days earlier than those who redeemed more than ve times. These dierences become
much larger in the xed eect regression. The dierences are more than 20 days between
those who are redeeming for the rst time and those who are redeeming for the second time
and more than 90 days between the former and those who redeemed more than ve times.
These results are consistent with those presented in Table 8.
These results imply that typical consumers in our data set redeem their coupons
right away after the rst purchase but then wait longer as they redeem more. Although
19
To be more specic the former probability is for those who bought two deals and are redeeming their
second coupon and the latter probability is for those who bought four deals and are redeeming their fourth
coupon.
20
Using redemption data for deals that have at least 30 days as the redemption period, we divide it into
three periods: the rst 10 days, the last 10 days, and the rest, and use the multinomial logit model to
estimate the probability of choosing a redemption time out of these three periods. Results and implications
are similar as those of Table 8.
18
we can only speculate why they are changing their redemption time this way, it may very
well be that they do so to avoid the congestion they experienced in the previous redemption
occasions. Considering the fact that small businesses are the main vendors on the daily-
deal sites, smoothing out the redemption time is a crucial for the success of the daily-deal
industry.
5 Conclusions
In this paper we empirically analyze consumers purchase and redemption time of daily-deal
coupons using individual-level data from one of the major daily-deal sites in Korea. The
daily-deal site we analyze is a social shopping site where each deal requires a certain number
of committed purchases to be on sale. We nd that typical consumers in their rst few
purchase experiences buy deals that are already on but tend to buy earlier as they buy more
deals. Experienced shoppers are more likely to buy before deals are on but we do not nd
evidence that they do so to turn on deals. Rather they tend to buy deals as soon as deals
start. We also nd that typical consumers redeem their rst few coupons right after their
purchases but tend to wait longer as they redeem more coupons.
Our ndings on the purchase time suggest that the social shopping feature of the
daily deal does not encourage inexperienced shoppers from buying deals early on. As they
buy more deals, however, they seem to learn that there is no cost of buying deals before they
are on. Our ndings on the redemption time provide evidence that inexperienced shoppers
drive congestion right after deals are sold and that they soon try to avoid crowds in their
subsequent redemptions.
When all these ndings are put together, it is hard to claim that the success of the
daily-deal industry is due to the group shopping feature. Rather it causes problems for small
businesses that have to handle an unusually large number of customers after deals are sold.
19
References
[1] Banerjee, A. (1992), A Simple Model of Herd Behavior, Quarterly Journal of Eco-
nomics, 107, 797-817.
[2] Byers, J. W., M. Mitzenmacher, and G. Zervas (2011), Daily Deals: Prediction, Social
Diusion, and Reputational Ramications, Working Paper, Boston University.
[3] Cai, H., Y. Chen, and H. Fang (2009), Observational Learning: Evidence from a Ran-
domized Natural Field Experiment, American Economic Review, 99, 86482.
[4] Dholakia, U. M. (2011), What Makes Groupon Promotion Protable For Businesses?,
Working Paper, Rice University.
[5] Dholakia, U. M. and S. E. Kimes (2011), Daily Deal Fatigue or Unabated Enthusiasm?
A Study of Consumer Perceptions of Daily Deal Promotions, Working Paper, Rice Uni-
versity.
[6] Edelman, B., S. Jae, S. D. Kominers (2011), To Groupon or Not to Groupon: The
Protability of Deep Discounts, Working Paper, Harvard Business School.
[7] Kumar, V. and B. Rajan (2012), Social Coupon as a Marketing Strategy: A Multifaceted
Perspective, Journal of the Academic Marketing Science, 40, 120-136.
[8] Subramanian, U. (2012), A Theory of Social Coupons, Working Paper, University of
Texas at Dallas.
[9] Ye, M., C. Wang, C. Aperjis, B. A. Huberman, and T. Sandholm, (2011) Collective
Attention and the Dynamics of Group Deals, Working Paper, Pennsylvania State Uni-
versity.
20
Table 1: Consumer Summary Statistics
Mean Stdev Median Min Max
Gender (Male=1) 0.36 0.48 0 0 1
Age 31.4 7.04 30 8 97
SMS (Yes=1) 0.56 0.50 1 0 1
e-mail (Yes=1) 0.52 0.50 1 0 1
Purchase

1.96 1.94 1 1 158

Volume of Order

2.01 1.47 2 1 40
Note: The summary statistics on 253,151 consumers who bought at least one deal from October
1, 2010 to December 31, 2010.

Purchase indicates the number of deals that these consumers bought during the sample period.

Volume of Order indicates the number of coupons consumers bought per deal.
21
Table 2: Product Summary Statistics
Mean Stdev Median Min Max
Product Sales 1,590 3,588 909 2 49,995
Original Price (in KW

) 69,384 120,110 35,000 700 1,000,000

Discount Rate (%) 56.50 8.45 54 15 81
Sale Periods (in days) 1.52 0.90 1 1 5
Tipping Point 112 292 100 1 5,000
Maximum Sale 2,984 6,249 1,500 30 60,000
Redemption Periods (in days) 83.85 46.59 90 0 366
Soldout

0.302 0.459 0 0 1
Note: The summary statistics on 696 deals that were sold from October 1, 2010 to December 31,
2010.

KW denotes Korean Won.

Soldout indicates whether the number of coupons sold reached the maximum number of coupons
available.
22
Table 3: Sales Analysis: OLS
All One-day No Soldout
Discount Rate 0.021** 0.032** 0.026**
(0.004) (0.005) (0.005)
log(Price) -0.613** -0.553** -0.669**
(0.038) (0.042) (0.047)
Tipping Point (in 100) 0.023** 0.024** 0.029*
(0.011) (0.011) (0.016)
Max. Sale (in 1,000) 0.060** 0.068** 0.057**
(0.006) (0.007) (0.007)
Redemption Periods 0.004** 0.004** 0.004**
(0.001) (0.001) (0.001)
D salesday 0.346** - 0.220**
(0.069) - (0.087)
Monday -0.024 -0.068 -0.013
(0.084) (0.087) (0.107)
Tuesday -0.065 -0.141 -0.015
(0.087) (0.088) (0.109)
Wednesday 0.097 0.068 0.140
(0.086) (0.088) (0.112)
Thursday -0.121 -0.091 -0.160
(0.084) (0.087) (0.108)
Restaurant 0.408** 0.693** 0.412**
(0.109) (0.119) (0.131)
Pub -0.300** 0.055 -0.265*
(0.131) (0.137) (0.151)
Cafe -0.235* 0.117 -0.313*
(0.142) (0.146) (0.175)
Beauty/Health Care 0.119 0.248* 0.110
(0.116) (0.127) (0.141)
Leisure 0.104 0.190 0.232
(0.142) (0.174) (0.168)
Concert/Exhibition 0.201* 0.475** 0.220
(0.122) (0.139) (0.137)
Adj. R-squared 0.614 0.611 0.629
N 679 506 473
* p-value < 0.1, ** p-value < 0.05
Note: The dependent variable is log(Sales). We also included the dummy variables for weeks that
deals were sold and market locations.
23
Table 4: Purchase Time: Multinomial Logit
0 - 6 12 - 18 18 - 24
First Purchase -0.632** 0.653** 0.792**
(0.021) (0.022) (0.028)
Second Purchase -0.547** 0.361** 0.401**
(0.021) (0.022) (0.027)
Third Purchase -0.417** 0.237** 0.237**
(0.021) (0.022) (0.028)
Fourth Purchase -0.336** 0.169** 0.155**
(0.022) (0.023) (0.029)
Fifth Purchase -0.240** 0.094** 0.080**
(0.023) (0.025) (0.032)
No Redemption -0.256** 0.064** -0.028
(0.022) (0.024) (0.030)
One Redemption -0.302** 0.079** -0.034
(0.021) (0.023) (0.029)
Two Redemptions -0.282** 0.076** -0.064**
(0.021) (0.024) (0.031)
Volume of Order -0.025** 0.017** 0.012**
(0.005) (0.004) (0.004)
Age -0.040** -0.004** -0.028**
(0.0008) (0.0006) (0.0008)
Male 0.451** 0.127** 0.370**
(0.010) (0.009) (0.011)
e-mail -0.319** -0.075** -0.194**
(0.015) (0.014) (0.016)
sms -0.156** 0.036** -0.015
(0.014) (0.013) (0.015)
e-mailSMS 0.268** 0.019 0.145**
(0.020) (0.018) (0.022)
Preference Area -0.176** -0.042** -0.106**
(0.011) (0.010) (0.012)
* p-value < 0.1, ** p-value < 0.05
Note: We also control for the deal characteristics reported in Table 2, days of the week, weeks that
deals were sold, deal categories, market locations, and weeks that consumers made accounts.
24
Table 5: Purchase Time: Linear Regression
OLS Fixed Eect
First Purchase 3.150** 1.513**
(0.050) (0.076)
Second Purchase 2.053** 0.890**
(0.048) (0.067)
Third Purchase 1.439** 0.560**
(0.049) (0.062)
Fourth Purchase 1.105** 0.398**
(0.051) (0.059)
Fifth Purchase 0.748** 0.179**
(0.055) (0.058)
No Redemption 0.584** 0.106
(0.052) (0.078)
One Redemption 0.662** 0.095
(0.050) (0.067)
Two Redemptions 0.587** 0.042
(0.052) (0.063)
Volume of Order 0.056** -0.029**
(0.009) (0.013)
Age 0.022** -
(0.002)
Male -0.184** -
(0.021)
e-mail 0.216** -
(0.032)
sms 0.312** -
(0.030)
e-mailSMS -0.247** -
(0.043)
Preference Area 0.089** -0.091**
(0.023) (0.036)
Adj. R-squared 0.147 0.119
N 370,616 370,616
* p-value < 0.1, ** p-value < 0.05
Note: We also control for the deal characteristics reported in Table 2, days of the week, weeks that
deals were sold, deal categories, market locations, and weeks that consumers made accounts.
25
Table 6: Purchase Before Deals are On
Binary Logistic FE Logit
First Purchase -0.987** -0.487**
(0.028) (0.054 )
Second Purchase -0.697** -0.307**
(0.027) (0.048)
Third Purchase -0.465** -0.187**
(0.027) (0.044)
Fourth Purchase -0.361** -0.143**
(0.028) (0.041)
Fifth Purchase -0.280** -0.113**
(0.030) (0.034)
No Redemption -0.201** -0.097*
(0.028) (0.054)
One Redemption -0.234** -0.052
(0.026) (0.046)
Two Redemptions -0.169** 0.010
(0.027) (0.042)
Volume of Order -0.051** -0.037**
(0.006) (0.010)
Age -0.019**
(0.001)
Male 0.219**
(0.013)
e-mail -0.118**
(0.020)
sms -0.083**
(0.018)
e-mailSMS 0.107**
(0.026)
Preference Area -0.009 0.081**
(0.014) (0.026)
N 370,616 71,402
* p-value < 0.1, ** p-value < 0.05
Note: We also control for the deal characteristics reported in Table 2, days of the week, weeks that
deals were sold, deal categories, market locations, and weeks that consumers made accounts.
26
Table 7: First and Last 10% of Sales before Deals are On
Last 10% First 10%
Binary Logit FE Logit Binary Logit FE Logit
First Purchase 0.334** 0.554** -0.720** -0.510**
(0.090) (0.275) (0.091) (0.237)
Second Purchase 0.208** 0.298 -0.394** -0.297
(0.087) (0.250) (0.085) (0.211)
Third Purchase 0.164* 0.078 -0.388* -0.263
(0.087) (0.225) (0.086) (0.193)
Fourth Purchase 0.034 0.038 -0.232** -0.308*
(0.093) (0.211) (0.087) (0.176)
Fifth Purchase 0.160* 0.331 -0.212** -0.136
(0.095) (0.202) (0.084) (0.167)
No Redemption -0.044 -0.507* 0.002 -0.017
(0.090) (0.263) (0.084) (0.230)
One Redemption -0.015 -0.097 -0.062 -0.085
(0.084) (0.223) (0.078) (0.184)
Two Redemptions -0.136 -0.041 -0.103 -0.017
(0.089) (0.206) (0.080) (0.168)
Volume of Order 0.100** 0.078* -0.111** -0.128**
(0.017) (0.046) (0.022) (0.049)
Age 0.002 -0.005
(0.003) (0.003)
Male -0.022 -0.004
(0.039) (0.041)
e-mail 0.017 0.006
(0.061) (0.063)
sms 0.014 -0.050
(0.057) (0.061)
e-mailSMS 0.007 0.085
(0.081) (0.085)
Preference Area -0.032 0.022 0.029 0.183*
(0.044) (0.135) (0.046) (0.110)
N 32,076 2,720 32,076 3,715
* p-value < 0.1, ** p-value < 0.05
Note: We also control for the deal characteristics reported in Table 2, days of the week, weeks that
deals were sold, deal categories, market locations, and weeks that consumers made accounts.
27
Table 8: Redemption Time
First 10 Days Last 10 Days
Binary Logistic FE Logit Binary Logistic FE Logit
One Purchase 0.205** 2.284** 0.682** -4.287**
(0.023) (0.091) (0.031) (0.254)
Two Purchases -0.070** 1.121** 0.664** -2.542**
(0.022) (0.070) (0.028) (0.196)
Three Purchases -0.161** 0.557** 0.583** -1.539**
(0.022) (0.059) (0.027) (0.162)
Four Purchases -0.168** 0.279** 0.455** -0.908**
(0.023) (0.052) (0.026) (0.138)
Five Purchases -0.162** 0.074 0.319** -0.545**
(0.024) (0.048) (0.027) (0.116)
First Redemption 1.898** 7.253** -2.094** -7.890**
(0.027) (0.072) (0.031) (0.094)
Second Redemption 1.334** 5.561** -1.468** -5.788**
(0.026) (0.062) (0.028) (0.077)
Third Redemption 0.966** 4.280** -1.027** -4.273**
(0.026) (0.055) (0.026) (0.065)
Fourth Redemption 0.666** 3.198** -0.668** -2.993**
(0.028) (0.051) (0.026) (0.057)
Fifth Redemption 0.481** 2.300** -0.482** -1.973**
(0.032) (0.050) (0.028) (0.050)
Volume of Order 0.138** 0.129** -0.169** -0.214**
(0.004) (0.010) (0.006) (0.015)
Age -0.008** 0.007**
(0.001) (0.001)
Male 0.198** -0.123**
(0.010) (0.012)
e-mail -0.046** 0.101**
(0.015) (0.019)
sms 0.015 0.061**
(0.014) (0.017)
e-mailSMS 0.022 -0.056**
(0.020) (0.024)
Preference Area 0.087** 0.024 -0.176** -0.121**
(0.011) (0.030) (0.013) (0.034)
N 267,143 96,892 267,143 68,867
* p-value < 0.1, ** p-value < 0.05
Note: We also control for the deal characteristics reported in Table 2, days of the week, weeks that
deals were sold, deal categories, market locations, and weeks that consumers made accounts.
28
Table 9: Redemption Time: Linear Regression
OLS Fixed Eect
One Purchase 8.376** -6.399**
(0.296) (0.535)
Two Purchases 9.764** 0.113
(0.275) (0.446)
Three Purchases 9.052** 3.163**
(0.269) (0.394)
Four Purchases 7.383** 3.759**
(0.272) (0.362)
Five Purchases 5.710** 3.187**
(0.282) (0.345)
First Redemption -39.816** -85.460**
(0.308) (0.320)
Second Redemption -28.918** -64.943**
(0.288) (0.286)
Third Redemption -20.694** -48.577**
(0.282) (0.266)
Fourth Redemption -14.176** -35.131**
(0.292) (0.261)
Fifth Redemption -9.754** -24.175**
(0.323) (0.270)
Volume of Order -2.388** -1.784**
(0.054) (0.067)
Age 0.113**
(0.009)
Male -2.769**
(0.126)
e-mail 1.163**
(0.191)
sms 0.072
(0.177)
e-mailSMS -0.332
(0.250)
Preference Area -2.466** -1.185**
(0.136) (0.181)
Adj. R-squared 0.199 0.106
N 267,143 267,143
* p-value < 0.1, ** p-value < 0.05
Note: We also control for the deal characteristics reported in Table 2, days of the week, weeks that
deals were sold, deal categories, market locations, and weeks that consumers made accounts.
29
Figure 1: Product Category
30
Figure 2: At What Time Consumers Buy Deals
31
Figure 3: When Deals are Tipped On
32
Figure 4: When Coupons are Redeemed
33