Sie sind auf Seite 1von 26

NET Institute*

www.NETinst.org

Working Paper #13-22


September 2013
Optimal Design of Two-Sided Market Platforms: An Empirical Case Study of eBay
Aaron Bodoh-Creed
University of California at Berkeley
Jorn Boehnke
University of Chicago
Brent R. Hickman
University of Chicago

* The Networks, Electronic Commerce, and Telecommunications (NET) Institute,


http://www.NETinst.org, is a non-profit institution devoted to research on network
industries, electronic commerce, telecommunications, the Internet, virtual networks
comprised of computers that share the same technical standard or operating system, and
on network issues in general.

Electronic copy available at: http://ssrn.com/abstract=2336395

OPTIMAL DESIGN OF TWO-SIDED MARKET PLATFORMS: AN EMPIRICAL


CASE STUDY OF EBAY
******** PRELIMINARY DRAFT ********
AARON BODOH-CREED, JRN BOEHNKE, AND BRENT R. HICKMAN
Abstract. While much is known about optimal design of auctions within the context
of a single item for sale, little is known about optimal design of large platform markets
like eBay and auto auction houses that house large numbers of concurrent auctions. We
attempt a macro-level empirical market design exercise by combining a unique dataset on
tablet sales collected from eBay over the course of a year with methodologies developed
by Bodoh-Creed (2011) and Backus and Lewis (2013). The former proposes a tractable
approach to studying dynamic auction markets when the number of participants on both
sides is sufficiently large. The model also delivers predictions on optimal fee schedule
design specifically, in terms of listing fees and percentage charges for a sale for an
auction house wishing to maximize profits by attracting the appropriate mix of buyers
and sellers into the market. We begin by empirically investigating the key assumptions
of the model which deliver (computational and empirical) tractability, and find that they
are reasonable. We then estimate consumer demand, market supply, and the distributions
of market entrants (this part still in progress). These figures are plugged into the BodohCreed (2011) framework in order to compute optimal fee schedules and draw comparisons
to actual fee schedules, as well as to make policy prescriptions.

Date: Current version: September 2013.


Key words and phrases. Online Auctions, eBay, Market Design, Large Markets.
Bodoh-Creed: University of California - Berkeley, Haas School of Business, S450 SSB #1900. Berkeley,
CA 94720; acreed@haas.berkeley.edu.
Boehnke: Department of Economics, University of Chicago, 1126 E 59th Street, Chicago, IL 60637;
jboehnke@uchicago.edu.
Hickman: Department of Economics, University of Chicago, 1126 E 59th Street, Chicago, IL 60637;
hickmanbr@uchicago.edu
.
We gratefully acknowledge financial support from the NET Institute (www.NETinst.org).

Electronic copy available at: http://ssrn.com/abstract=2336395

TWO-SIDED MARKET PLATFORMS

1. Introduction
This paper uses economic theory and statistical analysis to allow for computation of
the optimal pricing schedule for eBay under a variety of different assumptions on buyer
and seller behavior. We employ a unique eBay dataset which we have compiled over
the last year. It includes information on 100% of item postings on eBay within two
product lines: Kindle Fire and Microsoft Surface tablets. Our computations leverage
Bodoh-Creed (2013)s framework to study the effect of platform fees on the revenue
generated in each auction, the total market size, buyer-seller mixture, and the efficiency
of the platform (i.e., the total volume of trade). We propose to perform the first-ever
applied macro-level market design analysis for an entire platform market. We show
that our eBay data strongly support Bodoh-Creeds assumptions of a nonatomic stationary model, where matching between buyers and sellers follows a Poisson process.
These market characteristics vastly simplify the computation and empirical analysis of
the model.
eBay is a prime example of a market platform, or a business that specializes in facilitating trade between buyers and sellers. Market platforms are increasingly important in
todays economy: other major examples include StubHub, wholesale automobile auction
houses, art and wine auction houses (e.g., Christies and Sothebys), and Amazon.com.
These firms are all major players in their respective economic spheres: American car
dealers sourced 30% of their used inventory from auction houses in 2007 (see Roberts
(2012)); Amazon, eBay, and Christies each individually account for billions of dollars
in transactions annually; and Sothebys and StubHub account for hundreds of millions.
All of them (with the exception of Amazon) also have another peculiar characteristic in
common: despite their size and breadth of product offerings, none of them maintain any
of their own inventories. In other words, each one is a pure market maker, existing not
to produce or retail goods, but for the sole purpose of creating a forum where buyers
and sellers can trade.
Economists have learned much about the optimal design of a single auction in isolation, that is, how to configure the rules of the auction to generate maximal revenues
from a fixed number of bidders. On the other hand, despite the increasing importance of
market platform firms like eBay, little is known about optimal design of platform markets that host many concurrent auctions. To maximize profit, the firm obviously would
like to attract both sellers and buyers to the market, but achieving the optimal buyerseller mixture is also a primary concern. For example, when there is a large number
of auctions for a given product category, potential buyers will have high option values

TWO-SIDED MARKET PLATFORMS

which will lead them to submit low bids within a given auction. If the platform becomes
a buyers market with too many sellers relative to the number of buyers, then sellers
will begin to exit because their profits will be low. Conversely, if the platform becomes a
sellers market with few sellers relative to the number of buyers, then buyers will see too
much of their surplus being extracted through competition and they will begin to exit.
Thus, achieving both entry into the market and the optimal mix of buyers and sellers is
paramount.
In addition, revenue generation requires prices and/or fees be applied to some set of
the agents, but the magnitude of these prices can affect the mix of agents participating.
For example, if sellers face high charges when they sell a good, they will generate little revenue from the auctions and leave the market. This will, in turn, cause buyers to
abandon the market since there is little hope of purchasing a good at a low price, and
the loss of buyers will in turn cause more sellers to leave the market. Previous market
analyses have by and large ignored these effects. Recent advances in economic theory
have begun to explore these issues. Bodoh-Creed (2013) develops a theory model of
endogenous entry into the market place by individual buyers and sellers, and qualitatively explores macro-level market design questions. For example, to what extent should
platform fees (i.e., fees charged by the market platform firm) be assessed directly to buyers or sellers (or both)? Should fees be assessed at the point of entry into the market or
upon consummation of a trade (or both)? And, should fees be fixed or percentage based?
Bodoh-Creed also shows that it is possible to accurately approximate the model in a computationally tractable way as the number of buyers and sellers becomes large, so long as
the market satisfies certain stationarity conditions. In large market settings that satisfy
these conditions, the researcher is able to avoid various difficulties, including the curse of
dimensionality and computation of complex, history-dependent equilibrium strategies.
This paper provides empirical evidence that the assumptions underlying Bodoh-Creed
(2013) are met within the eBay marketplace.
From a business perspective, the power of combining economic theory with statistics
comes from the potential to forecast the effect of untested business strategies without engaging in potentially costly experimentation. For example, if one can estimate consumer
preferences and seller costs for a given product within a full economic model of the marketplace, one can then simulate shifts in prices and sales volume under experimental fee
schedules not observed in the data.1 Our (forthcoming) results provide estimates of how
1One example of this was an empirical analysis of timber auctions by the Canadian government in

Paarsch (1997). Paarsch estimated a model of bidder valuations based on a theoretical economic model
of bidding. Using his estimates, Paarsch computed the revenue maximizing reserve price and discovered

TWO-SIDED MARKET PLATFORMS

far the current eBay pricing structure is from the optimal structure. In other words, how
much could eBay increase revenues by altering its fee schedule? Although this question
could be addressed through a price experiment, the costly nature of changes to such a
basic element of the eBay platform suggests that our numerical computations will have
significant added value as a first cut at the problem. From an academic perspective,
our structural approach will be able to quantify the importance of eBay in developing
secondary markets for the goods we study.
The remainder of this paper has the following structure. In section 2, we survey the
related literature. In section 3, we describe and explore various aspects of our eBay
dataset. In section 4, we perform an explanatory analysis to investigate the appropriateness of assuming market demand and supply are stationary, and matching of bidders to
auctions is Poisson. In section 5, we outline continuing research and conclude.

that the Canadian government was setting reserve prices that were too low. They subsequently raised
their auction reserve prices and and saw significant increases in average revenue.

TWO-SIDED MARKET PLATFORMS

2. Literature Review
Several authors have investigated profit maximization from the perspective of a single
sale listing in isolation. Myerson (1981) showed that the optimal reserve price strictly
exceeds the sellers valuation for the item, and showed that the optimal selling mechanism with a fixed number of bidders can be implemented using a second-price auction
with an appropriately chosen reserve. The structural IO literature subsequently investigated this issue empirically, with papers such as Haile and Tamer (2003) and Paarsch
(1997). This literature has not considered, however, the task of optimally designing a
large market place like eBay that facilitates many sale listings simultaneously and across
time.
Bodoh-Creed (2013) develops a theory model of endogenous entry into the market
place by individual buyers and sellers, and explores macro-level market design questions such as the following: to what extent should platform fees (i.e., fees charged by the
market platform firm) be assessed directly to buyers, or sellers (or both)? Should fees
be assessed at the point of entry into the market, or upon consummation of a trade (or
both)? And, should fees be fixed, or percentage based? Bodoh-Creed (2013) also shows
that it is possible to accurately approximate the model in a computationally tractable
way as the number of buyers and sellers becomes large. Bodoh-Creed builds on previous work employing tractable large-market approximations, such as models studied by
Pesendorfer and Swinkels (2000), Swinkels (2001), Cripps and Swinkels (2006), McLean
and Postlewaite (2002), and Kojima and Pathak (2009).
Lewis and Backus (2013) explore identification and estimation within a fully dynamic
auction market in which entry is exogenous, multiple goods are auctioned simultaneously and over time, and private valuations are persistent for each individual bidder. In
related work, Nekipelov (2008) provides a model of eBay auctions as continuous time
games with repeated bidding within a single auction. In his framework, each single
auction is a dynamic game, but the market studied is not dynamic in our sense since the
agents do not enter new auctions in successive periods. Finally, Peters and Severinov
(2006) develop a model of multi-unit auction environment similar to eBay with the goal
of studying the sorting of buyers into sellers auctions. Their work provides a model of
selection into simultaneous auctions in a static setting. Bodoh-Creed (2013) emphasizes
the dynamics of the market and provides a relatively simple model of how buyers and
sellers match within each period.
Hickman (2010) constructed a theoretical model of an individual auction with the
same pricing rule as the one employed on eBay. Most researchers have assumed that

TWO-SIDED MARKET PLATFORMS

eBays pricing rule is a straightforward application of the common second-price format,


but Hickman (2010) showed that it is actually a hybrid of a first-price and a second-price
mechanism due to the role of minimum bid increments. The sale price of an item on
eBay is usually the second-highest bid plus a fixed increment, but when the top two
bids are close enough (i.e., closer than the increment), then the sale price is set equal
to the winners bid. Thus, a rational eBay bidder accounts for the fact that her bid
may affect the sale price, and the equilibrium lay strictly between the first-price and
second-price equilibrium. Hickman, Hubbard, and Paarsch (2013) explore the empirical
implications of the non-standard pricing rule on eBay. They perform a Monte Carlo
analysis to measure the magnitude of bias introduced by falsely assuming eBay uses
a simple second-price mechanism. They then explore nonparametric identification and
estimation of bidder valuations within a (static) eBay auction and estimate demand for
used laptops, while developing a simple test for when the hybrid nature of the pricing
rule can safely be ignored for simplicity.
Boehnke (2013) investigates why gift certificates often sell for more than their face
values on eBay. About 51.6% of Amazon gift certificates sold on eBay, for example, sell
S on average 104.1% of face value. Researchers have often
for prices that are too high A
attributed this apparent overpayment to bidding fever or pseudo-endowment effects.
However, using a novel dataset of eBay transactions, Boehnke showed that for about half
of all Amazon gift certificates overpayment occurs in Buy-it-now sales, and thus does not
involve bidding, much less bidding fever. Moreover, the data reveal that overpayment
is not random; rather, it is highly cyclic. These patterns appear to be rationalized by
institutional features such as eBay reward programs and special offers. These features
affect bidding and buying behavior in meaningful ways and should thus be taken into
consideration in future research using eBay data.

TWO-SIDED MARKET PLATFORMS

3. Data
We have monitored information on 100% of sales both auction and fixed price in
certain eBay product lines, including Video Games, Video Game Consoles, iPads,
Tablets, and e-book Readers, and Cell Phones and Smartphones over an unbroken
time-frame beginning in June 2012 and continuing through the present. Since last year
we have downloaded information on more than 1.5 million eBay transactions, amounting to 5TB of raw data. This dataset now constitutes an unparalleled asset for research.
We have collected product characteristics, prices, shipping / return policies, bidding behaviors, seller information, and many other variables. A powerful aspect of this dataset
is its time span: observing continuous sales over an extended period renders detailed
information on market conditions before, during, and after episodes of change. E.g.,
eBay alters its own pricing schemes such as listing fee offers during the holiday season.
Other details of data scraping, including ongoing efforts to acquire additional variables,
are discussed in the appendix.
Table 1 provides summary statistics on several key variables in our data. The three
products that our preliminary analysis will concentrate on are the Kindle Fire, an ebook reader developed and marketed by Amazon.com, the Microsoft Surface, and the
Microsoft Surface Pro, a relatively new line of PCs with the look and feel of a tablet,
but also having much of the capability of a traditional laptop computer. Each of these
three products is available in different conditions, being new, used / refurbished, and
damaged, meaning that the device does not function properly and is being sold for
repair or to use for spare parts. In total there are 38,577 transactions described in the
table, or an average of roughly 130 per day during the sample period (about 10 months)
for these products.
Other figures reported in the table include start price, or the reservation price set by
an individual seller, shipping price, duration of the auction listing (in days), sellers
return policy, number of bidders, buyer feedback score for the winning bidder, and
seller feedback score. Feedback scores for buyers and sellers are a count variable of the
number of past transactions for which the individual received positive feedback from
the other party to the transaction. The averages in the table are quite high mostly due to
a relatively small group of buyers and sellers on eBay who engage in very large amounts
of trade and thus have large feedback scores. For example, the median buyer feedback
and median seller feedback for all Kindle Fire sales are 58 and 457.5, respectively.

8741

18405

120.79
(49.08)
43.50
(50.24)
6.91
(5.15)
4.33
(2.37)
0.50
(0.50)
9.95
(4.69)
219.92
(952.16)
12593.38
(30570.92)

new

4567

1533

47.38
490.59
(18.39)
(128.23)
8.13
270.02
(12.38)
(202.53)
8.94
11.11
(2.82)
(10.25)
2.68
4.51
(1.95)
(2.44)
0.10
0.18
(0.30)
(0.38)
6.01
8.55
(2.35)
(5.21)
903.10
212.89
(1936.82)
(455.90)
30555.09
4482.54
(27116.99) (34070.86)

used/refurb damaged

mean coefficients; sd in parentheses

observations

189.90
(68.11)
start price
65.88
(81.14)
shipping price
5.00
(5.81)
duration (days)
3.54
(2.58)
returns allowed
0.45
(0.50)
# bidders
10.80
(5.26)
buyer score
192.96
(864.08)
seller score
4900.38
(25610.38)

price

new

Kindle Fire

2067

423.58
(93.94)
226.40
(180.62)
10.81
(10.10)
4.74
(2.29)
0.25
(0.43)
8.83
(5.27)
172.88
(515.87)
9464.54
(43108.39)

new

52

1123

246.36
723.57
(84.13)
(343.36)
24.85
425.99
(64.87)
(348.12)
8.52
11.03
(4.03)
(10.97)
2.17
4.70
(2.06)
(2.33)
0.15
0.15
(0.36)
(0.36)
9.38
8.72
(3.65)
(4.85)
313.00
452.72
(480.90) (2181.06)
63565.42
664.76
(62464.41) (5027.26)

used/refurb damaged

Surface

Table 1. Summary Statisics

1742

331.26
(347.12)
172.68
(262.58)
8.93
(8.03)
5.03
(2.21)
0.24
(0.43)
8.68
(4.90)
230.48
(1261.67)
10529.75
(57661.37)

347

47.12
(31.01)
13.47
(18.02)
7.81
(3.54)
4.30
(2.49)
0.17
(0.37)
6.18
(2.38)
603.45
(1454.33)
14072.46
(19485.29)

used/refurb damaged

Surface Pro

TWO-SIDED MARKET PLATFORMS

.5

1.5

Auction Duration

10

Auction Duration (days)


Kindle Fire
Surface Pro

Surface

Figure 1
Another thing to be learned from feedback scores is that buyers of damaged goods
tend to engage in more trade on eBay. This would seem to indicate that these buyers
are purchasing devices for commercial purposes (e.g., for refurbishment and subsequent
resale, or for spare parts) more so than buyers of functional goods. The distributions of
listing durations chosen by sellers (see figure 1) is fairly evenly distributed among 1 day,
3 days, 5 days, and 7 days, with very few sellers choosing to list items for longer than
a week. The average number of bidders who actively participate in a given auction is
generally between 8 and 10, with the distributions being quite similar across the three
product categories (see figure 2). Finally, figure 3 displays kernel smoothed densities of
prices for Kindle Fires, Surface, and Surface Pro tablets, respectively.
Table 2 depicts an exploration of the relation between sale prices and other relevant
variables. Three estimated regression equations are reported, each focusing on a given
product category, with sale price as the dependent variable. Item condition is a linearized
version of the condition categories, where new=1, refurbished=2, used=3, and damaged=4.
The results suggest, naturally, that the worse the condition of an item is, the less it will

TWO-SIDED MARKET PLATFORMS

.06
.04
0

.02

Density

.08

.1

Number of Bidders

10

20

30

Number of Bidders
Kindle Fire
Surface Pro

Surface

Figure 2
sell for. Seller feedback score seems to be generally correlated positively with sale price,
as one might expect. In terms of return policy (more specifically, whether returns are
accepted), the correlation is less straightforward. One can certainly see how offering a
return policy could be a costly signal of quality that would help to alleviate the lemons
problem, leading to a positive correlation with price. On the other hand, buyer feedback
for sellers may also have the effect of keeping sellers honest, and therefore may also lead
to a negative correlation between unobserved item condition and return policy. It is not
clear which effect should dominate, but in general the largest effects seem to be in the
intuitive positive direction.

10

TWO-SIDED MARKET PLATFORMS

.004
.002
0

Density

.006

.008

Price Distribution

500

1000

1500

Price
Kindle Fire
Surface Pro

Figure 3

Surface

2000

TWO-SIDED MARKET PLATFORMS

11

Table 2. Regression of Price on Item Condition, Seller Score Category, and


Return Allowed Dummy with Constant

item condition
seller score 10, < 100
seller score 100
returns allowed
constant
observations
R2

Kindle Fire

Surface

Surface Pro

price

price

price

-40.34
(0.29)
5.35
(1.22)
-4.94
(1.13)
-2.63
(0.61)
235.25
(0.86)

-35.32
(1.82)
16.17
(5.84)
6.45
(5.88)
-7.37
(4.65)
530.36
(4.38)

-25.22
(4.88)
33.27
(13.83)
24.72
(13.90)
29.77
(14.36)
904.12
(9.92)

33435
0.399

3649
0.120

1387
0.025

b coefficients; Standard errors in parentheses; - represents no data


p < 0.10, p < 0.05, p < 0.01

12

TWO-SIDED MARKET PLATFORMS

4. Exploratory Analysis of Model Assumptions.


The primary objective of this research is to produce concrete policy prescriptions on
optimal platform fee schedules for eBay, which is perhaps the most prominent example
of a platform market today. As mentioned previously, our empirical strategy will be to
combine the demand estimator proposed by Backus and Lewis (2013) with the macrolevel model of entry and exit in a dynamic auction market proposed by Bodoh-Creed
(2011). In particular, the latter will allow us to model the effect of different pricing
schemes on participation by buyers and sellers in the market in a feasible way. BodohCreed (2011) achieves tractability in a very complex model by showing how various
aspects of the environment simplify as a large market approaches a nonatomic limit.
However, given that any dataset will only include at most finite market players, the
limit approximation is just that: an approximation. Underlying this approximation are
two key assumptions stable demand over time and Poisson matching of buyers to
sellers that eliminate much strategic complexity in the limit. The assumption that
sellers are matched with bidders through an exogenous Poisson arrival rate is important
because it means no seller can affect his likely customer pool through his own strategies.
Moreover, stationarity allows for dynamic strategies to be Markovian, rather than history
dependent.
Here we turn to a preliminary empirical analysis of our data to determine the appropriateness of the assumptions relating to model tractability. As a first cut at the data, it
appears that the nonatomic property of the market is a good fit: there are over 17,000
unique sellers in our data on tablet devices over the course of roughly 10 months, with
over 35,000 total transactions and an estimated 43,000 unique buyers during that same
span. However, these figures are merely suggestive on their own.
4.1. Market Stationarity. Table 3 depicts output from a set of probit regressions to investigate the stability of market demand over time. The regressions use data from fixed
price listings (called buy-it-now sales on eBay), and the dependent variable is a binary
one indicating whether an item was sold within a 24-hour period. The first specification includes item characteristics and a full set of month controls (January is the omitted month), whereas the second one adds in some additional market-level controls for
supply-side trends, including the total number of listings on a given day, and the average
and lowest prices as well.
Note that month dummies do have a statistically significant effect on probability of
sale. However, the implied effects are economically of little consequence. For example,
moving from the first to the last day of the sample is the same as lowering price by

TWO-SIDED MARKET PLATFORMS

13

Table 3. Probability of Sale

price
seller score
days since sample start
new
refurbished
broken
returns allowed
# listings
price - lowest price per week
price - average price per week
feb
mar
apr
may
jun
jul
aug
sep
oct
nov
dec

coeff w/o supply side

coeff w/ supply side

-0.0087227
3.6286e-06
-0.0060036
0.85853
1.4546
1.4529
0.8172

-0.1053
-0.37959
-0.38671
0.12404
0.016551
-0.14915
-0.15564
-1.5412
-1.227
-0.79117
-0.55866

-0.0087922
2.249e-06
-0.0073672
1.0629
0.68772
0.27414
0.83012
-0.0012612
-4.2516e-05
0.00013471
-0.25375
-0.51074
-0.40743
0.21096
0.28781
-0.068857
-1.5121
-2.7815
-2.0993
-1.0958
-0.66087

all coefficients are significant at the 1% level with the exception of sep which is not significant

$2.73= (.006 * 365)/(.008*100), and moving from the best to the worst possible month is
the same as cutting price by about $2.02. By comparison, the 25% - 75% split in prices is
roughly $37.36. So, in conclusion, current price appears to matter more than essentially
anything else, and the market is more or less stationary.
4.2. Poisson Matching. As a first cut, we began by partitioning the data into different
subsamples that were relatively homogeneous. The most plentiful data in the sample
were on used Kindle Fires with 8GB hard drives. From this subsample we eliminated
observations where the seller feedback score was less than 100 or where the starting price
was weakly greater than $50. Within the remaining subsample, we then partitioned it
into four sets of listings lasting 1 day, 3 days, 5 days, and 7 days, respectively, and
then collected a vector of observed bidder participation {k t }tT=1 . Now, assuming that
participation follows a Poisson distribution
K f K (k) =

k
exp(),
k!

14

TWO-SIDED MARKET PLATFORMS

we estimated the Poisson arrival rate via maximum likelihood. Figures 4 and 5 depict
the results of the unconditional Poisson model, where comparisons are also made to
the Kaplan-Meier empirical CDF (which does not impose any a priori shape restrictions).
Note that for each of the four subsamples, the estimated Poisson model fits the empirical
CDF remarkably well.
The fit is least tight for the subsample of 7-day listings for which the empirical CDF
is roughly a mean-preserving spread of the Poisson model. It is worth mentioning that
there are various other covariates not being controlled for (return policy and ending
day) which might account for the increased variance relative to the estimated parametric
Poisson model. Overall though, these figures suggest a remarkably tight fit between our
data and the assumption that buyers are matched to sellers via an exogenous Poisson
process.
Tables 4-6 contain the results of a conditional Poisson model of bidder participation in
a given auction. Here, we specify a single index model = 0 + 1 x1 + . . . + L x L = X,
which leads to the conditional Poisson distribution
K | X f K |X (k | x ) =

(X)k
exp(X).
k!

Estimation was performed via Maximum Likelihood and results are represented in Tables 4-6.
Table 4. Conditional Poisson Model: Amazon Kindle Fire

constant
new
used/refurbished (omitted = damaged)
storage size
listing duration
ends on mon
ends on tue
ends on wed
ends on thu
ends on fri
ends on sat (omitted = sun)
reserve price
returns allowed dummy
seller score 10, < 100
seller score 100 (omitted = score < 10)

beta

stderr

t-stat

p-value

4.8755
7.2070
5.5463
0.0978
0.0566
0.0669
0.0819
-0.0293
-0.0199
-0.0724
0.1682
-5.4702
0.1411
0.4338
0.5968

0.0492
0.0465
0.0426
0.0006
0.0039
0.0269
0.0299
0.0377
0.0323
0.0430
0.0261
0.0124
0.0204
0.0345
0.0318

99.0181
154.8457
130.1670
160.1545
14.6542
2.4913
2.7425
-0.7781
-0.6161
-1.6840
6.4479
-440.0418
6.9122
12.5910
18.7446

0
0
0
0
0
0.0127
0.0061
0.4365
0.5378
0.0922
0.0000
0
0.0000
0
0

TWO-SIDED MARKET PLATFORMS

15

Table 5. Conditional Poisson Model: Microsoft Surface


beta
constant
new
used/refurbished (omitted = damaged)
storage size
listing duration
ends on mon
ends on tue
ends on wed
ends on thu
ends on fri
ends on sat (omitted = sun)
reserve price
returns allowed dummy
seller score 10, < 100
seller score 100 (omitted = score < 10)

9.0060
4.6341
4.0441
0.0041
0.0503
0.2171
0.3876
0.4464
0.0351
0.4760
0.2410
-0.0218
-0.0289
0.0362
0.1784

stderr

t-stat

0.4879 18.4604
0.4785 9.6852
0.4743 8.5270
0.0017 2.4949
0.0092 5.4553
0.0811 2.6752
0.1190 3.2555
0.0808 5.5230
0.1170 0.3003
0.1290 3.6909
0.0966 2.4948
0.0002 -137.1902
0.0684 -0.4219
0.1043 0.3468
0.1042 1.7118

p-value
0
0
0
0.0126
0.0000
0.0075
0.0011
0.0000
0.7640
0.0002
0.0126
0
0.6731
0.7288
0.0869

Table 6. Conditional Poisson Model: Microsoft Surface Pro


beta
constant
new
used/refurbished (omitted = damaged)
storage size
listing duration
ends on mon
ends on tue
ends on wed
ends on thu
ends on fri
ends on sat (omitted = sun)
reserve price
returns allowed dummy
seller score 10, < 100
seller score 100 (omitted = score < 10)

8.1083
2.7879
2.5087
0.0303
-0.0008
0.0958
0.1810
0.2110
0.4904
-0.0885
0.3963
-0.0113
0.2413
0.4141
0.8076

stderr

t-stat

1.9178 4.2279
1.8928 1.4729
1.8912 1.3266
0.0025 12.2791
0.0282 -0.0291
0.2537 0.3776
0.2368 0.7643
0.2341 0.9012
0.2745 1.7864
0.1987 -0.4452
0.2745 1.4440
0.0002 -57.3945
0.2392 1.0090
0.2006 2.0648
0.2144 3.7666

p-value
0.0000
0.1408
0.1847
0
0.9768
0.7057
0.4447
0.3675
0.0740
0.6562
0.1487
0
0.3130
0.0389
0.0002

4.3. Structural Methodology. The primitives for the model of optimal platform fees are
the production cost of the seller (c), the price the seller will receive net of eBays fees (p),
and the probability the seller sells the good in the present period given the price chosen

16

TWO-SIDED MARKET PLATFORMS

Duration = 1 Days; N = 2351


1

0.6

F (k)

0.8

KM ECDF
95% UB
95% LB
MLE POISSON ESTIMATE
95% UB
95% LB

0.4
0.2
0

10

12

14

16

18

20

22

Duration = 3 Days; N = 1732


1

0.6

F (k)

0.8

KM ECDF
95% UB
95% LB
MLE POISSON ESTIMATE
95% UB
95% LB

0.4
0.2
0

10

12

14

16

18

20

22

Figure 4
(( p)). The sellers problem is then
V (c) = max ( p) ( p c) + (1 ( p)) V (c)
p 0

The basic idea is that with some probability the seller sells this period and receives p c,
but if not the seller can try to sell the item next period. We may need to adjust this for
listing fees, etc., eBay charges even if no sale occurs. Let the optimal p as a function of c
be p (c).
The first order condition for this is
0 ( p) ( p c) + ( p) 0 ( p) V (c) = 0

TWO-SIDED MARKET PLATFORMS

17

Duration = 5 Days; N = 1281


1

0.6

F (k)

0.8

KM ECDF
95% UB
95% LB
MLE POISSON ESTIMATE
95% UB
95% LB

0.4

0.2

10

12

14

16

18

20

22

Duration = 7 Days; N = 1386


1

0.6

F (k)

0.8

KM ECDF
95% UB
95% LB
MLE POISSON ESTIMATE
95% UB
95% LB

0.4

0.2

10

12

14

16

18

20

22

Figure 5
where we can use the stationarity of the model to find the value function
V (c) =

( p (c)) ( p (c) c)
.
1 (1 ( p (c)))

( p) can be estimated, and we are already preparing to make an assumption about


. Since everything is differentiable, we also know that p (c) is strictly increasing in c
from Topkiss theorem techniques. These same techniques also tell us that the FOC is
sufficient for a global optimum. As for how to solve it, we plan to use a contraction
mapping algorithm.
4.4. General Empirical Framework and Method. As mentioned above, we plan to use
methodologies and frameworks proposed by Bodoh-Creed (2013) and Lewis and Backus

18

TWO-SIDED MARKET PLATFORMS

(2013) as the basis of our empirical study. Lewis and Backus propose a model of a
dynamic auction market with exogenous entry of buyers and sellers. Briefly, within
their model a bidder has privately observed demands for a set of J goods, and faces
a simultaneous set of second-price auctions for those goods, repeated over multiple
periods of time. New buyers and sellers exogenously enter and exit each period, with
buyers making decisions each period on which of the J items to compete for, and how
much to bid. Unlike in a static second-price auction, buyers bid strictly less than their
private valuations because of the option value of other items for sale in future periods. In
Zs
private valuation
fact, optimal bidding strategies in large markets are precisely oneA
minus the continuation value. Therefore, if there are a larger number of items available
in future periods, continuation values will be higher and current bids lower, with the
reverse also being true.
These continuation values can be expressed as functions of bid distributions and state
transition processes which the researcher can glean directly from data, to empirically reconstruct the Bellman equation which determines optimal product choice and bid choice
as a function of (unobserved) private valuations. Backus and Lewis then show how
monotonicity conditions naturally arising in large market settings lead to identification
of private valuations, subject to the caveat that multiple observations of bidder behavior
over time must be available when J 2. The intuition is simple: since valuations are
persistent for a given bidder over time, then in order to identify the correlation between
valuations for say Xbox and PlayStation consoles, one must be able to glean information
on both valuations for a given bidder, which requires observations of some bidders who
bid on both over time.
Subject to these data limitations, Lewis and Backus propose three alternative estimators based on their identification argument. We plan to use their second proposal, which
uses simulated generalized method of moments techniques to estimate demand by a
two-stage semiparametric process. In the first stage, the form of the Bellman, conditional
on a given private valuation equation is nonparametrically recovered from observables
as described above. Then, in the second stage, sets of private valuations are simulated
from a parametric distribution. These sets of private valuations induce a set of individual bidder time series (including bids, participation decisions, and outcomes) through
the policy functions arising from the nonparametric Bellman equations. The moments
of simulated bidder time series are then matched to the empirical moments of observed
bidder time series in order to obtain an estimate of the joint distribution of valuations for
the J goods in the market. The above analysis performed on our eBay data for the e-book

TWO-SIDED MARKET PLATFORMS

19

and tablet markets will be the first phase of our project.2 The second phase will involve
plugging estimated demand into the Bodoh-Creed framework of endogenous entry into
the platform market, in order to obtain structural estimates of counterfactual platform
firm profits under varying levels of fees. This exercise will involve an additional empirical layer of calibrating aspects of the Bodoh-Creed model to the eBay data (e.g., the
ratio of buyers and sellers, the appearance rate of new market entrants, etc.) in order to
complete the model of endogenous entry as a function of fee schedules. Post calibration,
we shall compute counterfactual estimates of buyer/seller entry, bidder behavior, and
market outcomes, under different fee schedules. It is through the combination of these
two frameworks and our extensive data set that our objective of macro-level applied
market design will be accomplished.

2Note, however, within the Kindle Fire market J = 1 can be assumed since Kindle has close to 100%

market share for e-book readers. Therefore, repeated observations of bidders across time are not necessary
for identification in this case.

20

TWO-SIDED MARKET PLATFORMS

5. Conclusion
We provide an empirical model of a decentralized, two-sided, dynamic auction market
(inspired by the eBay market platform) in which a large number of buyers and sellers
participate in simultaneous auctions each period. Our model is well tailored toward
an applied macro-level market design exercise to uncover optimal platform fee schedules and policy counterfactuals. However, the model relies on strong assumptions that
the data generating process is (i ) nonatomic, (ii ) stationary, and (iii ) involves Poisson
matching between buyers and sellers. We provide empirical evidence that each of these
assumptions are reasonable as applied to eBay data on e-book readers and tablet PCs.
We also outline our ongoing research to estimate the full structural model.

TWO-SIDED MARKET PLATFORMS

21

APPENDIX
5.1. Additional Details on Data Acquisition and Processing. We have developed a Java
- MySQL software application over the last year which automates the download and
archiving of eBay listings. The combination of Java and MySQL packages allows the implementation to be modular, which eases software development and allows for parallel
deployment on multiple processor cores. Each basic function of the application is housed
in a separate module: one each for pre-processing, downloads, analysis, optimization,
and result output. The data collection is performed in three stages: monitoring of all
items listed in a specific category on eBay, downloading of all completed listings, and
finding and downloading of all related seller information pages. The raw data we have
downloaded consist of HTML source code for eBay pages, with organizing information
for the various pages attached to each listing being housed in a central MySQL database.
As of the time of this writing, we have completed software development for a Java application that parses out nearly all transaction information from its HTML source code.
The extracted data are then compiled into a separate MySQL database that allows for
easy access of the data for statistical analyses.
5.2. Ongoing Data Processing Efforts: Buyer and Seller Identification Over Time. One
final detail of our data processing effort is ongoing. Identification of the joint distribution
of different goods (e.g., Xbox and Playstation gaming consoles) within the Backus Lewis
(2013) framework requires repeated observations of buyers and sellers on eBay over
time. Observing seller identities over time is directly possible given information available
on eBay item listing pages. Identification of bidder identities across different auctions
is more difficult since eBay started anonymizing bidder identities several years ago,
e.g., listing user henrycarr instead as h***r for example. This means that only the
Zs
current feedback
first and last characters of each username, along with the userA
score are directly observable. The problem which arises is that it is not uncommon for
multiple usernames to appear as the same anonymized character string; e.g., username

I would also show up as AIJh***r

I.

AIJharoldpalmer
A
A
This creates a barrier to answering our empirical questions concerning optimal platform market design, since data substitution patterns across different related products
will play a significant role. For example, if Xbox and PlayStation are close substitutes,
Zs
outside option relative to a given Xbox listing will not just be the
then a bidderA
rest of the Xbox market, but effectively the rest of the gaming market. As in the models
of Bodoh-Creed and Backus-Lewis, having more plentiful outside options will impact
Zs
purchasing behavior, and by extension expected profits by sellers as well.
a buyerA

22

TWO-SIDED MARKET PLATFORMS

In turn, these things will also have a fundamental impact on the trade-offs faced by a
market platform firm wishing to optimally design fee schedules. Therefore, being able
to link bidders across auctions is crucial for our proposed research.
Moreover, observations of bidder identities would open up a new dimension to the
analysis, which has been unexplored by previous research: eBay as a primary market
versus eBay as a secondary market. For example, many buyers on eBay also turn around
and become sellers as well. Since eBay creates such a convenient forum for liquidation
of assets, buyers on eBay may take this into account when making primary purchases
e.g., when I get tired of playing World of Warcraft, I can sell it and get another game...
thus impacting their outside options, the entry rate of sellers, and in turn, the platform
market design problem. In short, the ability to link agents through different transactions,
both as buyers and as buyer-sellers, will substantially empower us to use our extensive
market data to produce new insights on optimal design of a platform market.
Given the large benefits to be had from the capability of linking bidders across time, we
are continuing to develop software to indirectly infer linkages from information gleaned
from downloaded bidder pages. Each time a bidder places a bid, information on that
bidders bid history for the previous month is displayed for other users to see. This
information includes number of bids submitted in different product categories, items
purchased, and other data. Taken together, these form a relatively high-dimensional
variable that can be used as a fingerprint, and compared to ed to other bid histories in
order to find a match to previous bidders who bid on other item listings in the same
product category. The high-dimensionality of the bid history presents at the same time,
a significant challenge, but also potential for accurate matching.

TWO-SIDED MARKET PLATFORMS

23

6. References
M. Backus and G. Lewis. A Demand System for a Dynamic Auction Market with
Directed Search. Working Paper, 2012. http://ssrn.com/abstract=1543835.
A. L. Bodoh-Creed. Optimal Platform Fees for Large Dynamic Auction Markets.
Working Paper, 2013. http://www.economics.cornell.edu/ab882/eBay.pdf.
J. Boehnke. Recovery from bidding fever: Why pay more than 102% for a gift card?
Working Paper, 2013. http://home.uchicago.edu/ jboehnke/GCs.pdf.
J. Bulow and P. Klemperer. Auctions Versus Negotiations. American Economic Review, 1996, 86(1), 180-194.
M. W. Cripps and J. M. Swinkels. Efficiency of Large Double Auctions. Econometrica,
2006, 74(1), 47-92.
R. Ericson and A. Pakes. Markov-Perfect Industry Dynamics: A Framework for Empirical Work. Review of Economic Studies, 1995, 62(1), 53-82.
P. A. Haile and E. Tamer. Inference with an incomplete model of English auctions.
Journal of Political Economy, 2003, 111(1), 1-51.
B. R. Hickman. On the pricing rule in electronic auctions. International Journal of
Industrial Organization, 2010, 28(5), 423-433.
B. R. Hickman and T. P. Hubbard and Y. Saglam. Structural Econometric Methods in
Auctions: A Guide to the Literature. Journal of Econometric Methods, 2012, 1(1), 1-107.
B. R. Hickman and T. P. Hubbard and H. J. Paarsch. Investigating the economic importance of pricing-rule mis-specification in empirical models of electronic auctions. Working Paper, 2013. http://home.uchicago.edu/ hickmanbr/uploads/HHP_EA.pdf
B. R. Hickman. Pre-college human capital investment and affirmative action: A structural policy analysis of us college admissions. Working Paper, 2013.
http://home.uchicago.edu/ hickmanbr/uploads/AA_Empirical_paper.pdf
F. Kojima and P. A. Pathak. Incentives and Stability in Large Two-Sided Matching
Markets. American Economic Review, 2009, 99(3), 608-627.
R. McLean and A. Postlewaite. Informational Size and Incentive Compatibility. Econometrica, 2002, 70(6), 2421-2453.
R. Myerson. Optimal Auction Design. Mathematics of Operations Research, 1981, 6(1),
58-73.
D. Nekipelov. Empirical analysis of dynamic bidding on eBay. Working Paper, 2008.
http://faculty.haas.berkeley.edu/wolfram/InnovSem/PapersF08/
nekipelov_ebay_aug08.pdf.

24

TWO-SIDED MARKET PLATFORMS

H. J. Paarsch. Deriving an estimate of the optimal reserve price: an application to


British Columbian timber sales. Journal of Econometrics, 1997, 78(1), 333-357.
W. Pesendorfer and J. M. Swinkels. Efficiency and Information Aggregation in Auctions. American Economic Review, 2000, 90(3), 499-525.
M. Peters and S. Severinov. Internet auctions with many traders. Journal of Economic
Theory, 2006, 130(1), 220-245.
J. Roberts. Unobserved Heterogeneity and Reserve Prices in Auctions. Working paper,
2012. http://papers.ssrn.com/sol3/papers.cfm?abstract_id=1695647.
J. M. Swinkels. Efficiency of large private value auctions. Econometrica, 2001, 69(1),
37-68.

Das könnte Ihnen auch gefallen