Using Agents to model the eBay
economy
David Shepard
Process
 Objective: Determine if a computer program can create
agents to emulate the behavior of real bidders on eBay
 Method: Retrieve detailed data on real eBay auctions,
create agents that use this empirical data as a model
 Success: Simulated auctions and eBay auctions follow a
similar pattern
Process
Data Retrieval
Analytics
Modeling/Simulation
with agents
Methods
 Data retrieval
 Need to know what kind of behavior to model
 eBay licenses the data to a 3rd party Terrapeak
 Terrapeak had detailed bidding summaries but despite effortful
negotiations we couldn’t meet their price
Methods
 Data retrieval continued
 eBay possesses a comprehensive API
 However API only provides methods for accessing live auctions,
completed auctions are outsourced to Terrapeak
Methods
 Live auctions
 Created a system that would monitor and record auction activity in
real-time
Search(“ipod nano”)
120829614046
IM 1
423812614446
IM 2
456197631246
IM 3
984829412551
Listings (item id’s)
IM 4
 Could only pull bidding data, eBay couldn’t push Item Monitors
 Resulted in missed bids, inefficient
Methods
 Finally settled on screen scraping
 Beautiful Soup module for python
 Able to scrape bidding information off of actual bidding
summary web pages
Modeling
 Retrieved bidding summaries for 200 ipod nano auctions
 Filtered bidding by users (approx. 7 bidders per auction, 2 bids
per bidder
 (1-3 lowballers, 2-6 proxy bidders)
 Formulated linear and polynomial equations relating a given
user’s response to a prior bid
User: d***c:
35
y = 1.1803x - 2.9514
R² = 0.9964
Bid Response
30
25
20
15
Bids
10
Linear (Bids)
5
0
0
5
10
15
Prev. Bid
20
25
30
Modeling
 Separated bidders into lowball bidders and proxy bidders
based on bidding frequency
 Lowball bidders constantly bidding in small increments above
highest bid
 Proxy bidders make educated bids 1-2 times per auction
 Simulation: Randomly add lowballers and proxy bidders to
the auction, each type randomly chooses an equation from
their respective equation list
Results/Evaluation
 Actual bidding patterns remain fairly similar across auctions
with minimal standard deviation
Bid Responses to Previous Bids (Aggregate,
100 auctions)
200
y = 1.0065x + 2.6514
180
Bid Response
160
140
120
100
80
Linear ()
60
40
20
0
0
50
100
Previous Bid
150
200
Results/Evaluation
 Simulated auctions followed a pattern similar to real-life auctions
 Simulated average slope: 1.03511331
 Actual average slope: 1.004505283
 Tentative yes, agents can model human bidder behavior
 Not perfect, however, doesn’t perfectly model proxy bidders nor does it take
into account bid sniping
Simulated Bid Responses to Previous
Bids
120
y = 1.0238x + 2.2173
R² = 0.9714
100
80
60
Linear ()
40
20
0
0
20
40
60
80
Previous Bid
100
120
Bid Response
Bid Response
Bid Responses to Previous Bids
(Aggregate, 100 auctions)
200
180
160
140
120
100
80
60
40
20
0
y = 1.0065x + 2.6514
R² = 0.9883
Li
0
50
100
Previous Bid
150
200
Future Work
 Do bidders exercise the same behavior for other eBay
categories?
 For example are the behavior of bidders bidding on antiques
different than auctions?