Learning through Simultaneous Play: Evidence from Penny Auctions

• Date: 01 December 2016
• Published date: 01 December 2016
• DOI: http://doi.org/10.1111/jems.12174

Learning through Simultaneous Play: Evidence from

Penny Auctions

RICARDO GONC¸ALVES

Cat´

olica Porto Business School and CEGE

Universidade Cat´

olica Portuguesa

Rua Diogo Botelho, 1327

4169-005 Porto, Portugal

rgoncalves@porto.ucp.pt

MIGUEL A. FONSECA

University of Exeter Business School

Streatham Court

Rennes Drive

Exeter EX4 4PU, United Kingdom

m.a.fonseca@exeter.ac.uk

This paper contributes to the emerging empirical literature on penny auctions, a particular type

of all-pay auctions. We focus on the potential learning effects that bidders may experience over

time but also (and particularly) across auctions as a result of their auction participation. Using

detailed bid-level information, we find that, similarly to earlier literature, bidders suffer from a

sunk cost fallacy, whereby their probability of dropping out of an auction is decreasing in the

number of bids they have already placed in that auction. Although we do find that learning

through repeated participation alleviates the sunk cost fallacy, participation in simultaneous

penny auctions emerges as a much more effective learning mechanism, ultimately contributing

toward bidders earning higher individual surpluses.

1. Introduction

The penny auction was popularized by firms like Swoopo and is still used by online

auction companies, as well as by traditional retailers across the world.1In a penny

auction, bidding usually starts at zero and bidders must pay a bid cost to increase the

sale price by a small amount—typically one penny, hence the name of the auction. 2A

key attraction of this type of auction is the possibility of paying substantially less than

the retail price for an object. However, this does not necessarily mean the auctioneer has

made a loss. For instance, if each bid in a penny auction costs $1 to place, an iPad with a

retail price of $500 that is sold for $75 in a penny auction yields a revenue of $7,575 to the

Wethank a telecommunications operator (which has requested its identity not to be disclosed) for providing us

the bid-level database for the auctions analyzed in this paper.Financial support from Fundac¸˜

ao para a Ciˆ

encia

e Tecnologia (through project PEst-OE/EGE/UI0731/2011) is gratefully acknowledged. We thank Ricardo

Ribeiro and participants in a seminar at Universidade Cat´

olica Portuguesa, as well as in the 7th Meeting of the

Portuguese Economic Journal (2013, Covilh˜

a, Portugal) and in the 40th Annual Conference of the European

Association for Research in Industrial Economics (EARIE) (2013, ´

Evora, Portugal) for their helpful comments

and suggestions. We also thank the editor, a co-editor, and a referee for the valuable suggestions made.

1. Swoopo filed for bankruptcy in 2011.

2. Although penny auctions bear some resemblance to the war-of-attrition game, a type of all-pay auction,

they are not a special case of any type of all-pay auctions (Hinnosaar,2014).

C2016 Wiley Periodicals, Inc.

Journal of Economics & Management Strategy, Volume25, Number 4, Winter 2016, 1040–1059

Learning through Simultaneous Play 1041

auctioneer ($7,500 from bid costs plus $75 from the actual sale price) and a substantial

profit margin. Indeed, the recent interest in penny auctions has been driven both by

its popularity as an e-commerce mechanism, as well as by the empirically observed

high-profit margins—a clear violation of auction theory (expected revenue should equal

the good’s value) and yet another real-world example of overbidding in auctions.

The standard game theoretical analysis assumes that players arrive at a Nash

equilibrium through an introspective process, in which they form beliefs about their

opponents’ actions, and beliefs about their opponents’ beliefs about their own actions,

and so on. While real people may be able to engage in this type of mental processin simple

games with few players and with a unique Nash equilibrium, it is less reasonable to

expect this to be true in more complex games with many players and multiple equilibria,

as is the case of penny auctions. Instead, as already suggested in the penny auction

literature, which we review below,it may be that repeated experience (over time, within

a given auction or in subsequent auctions) in this type of game allows players to learn

what the optimal strategy is (or at least allows players to identify and use payoff-

enhancing strategies).

However, a dimension that has thus far been ignored is the potential contribution

that participation in simultaneous auctions has in this learning process. Indeed, “ex-

perience” may be obtained “vertically,” over time through bid submission or auction

participation, but also “horizontally,” within a time window but through bid submis-

sion or simultaneous participation in more than one auction. Theoretically (Mertens,

1992), one would expect such parallel auctions to be independent and thus not affect

bidding behavior; in reality, it may actually speed up the bidders’ learning process. To

the best of our knowledge, this is a little-explored subject in the economics literature.3In

penny auctions, bidders are learning a complex object. We conjecture that by virtue of

participating in concurrent auctions with different types of bidders, or even by bidding

in auctions at different stages, subjects can learn (i) how to play a particular optimal

strategy, or learn that (ii) there are potentially different optimal strategies faster than if

they only take part in one auction.

There is a small (experimental) literature on behavioral spillovers that is somewhat

related to these learning effects. Cason et al. (2012) consider a behavioral spillover to

exist when observed behavior in a game is different depending on whether that game is

played together with other games or in isolation and acknowledge that learning effects

can be a source of such spillovers.4

3. In neuroscience, Sigman and Dehaene (2008) show the coexistence of serial and parallel brain processes

during the performance of a cognitive task. Gombrich (2011) puts forward a tentative definition of series

learning (equivalent to our “vertical” learning) as opposed to parallel learning (equivalent to our “horizontal”

learning), borrowing from the working of electric circuits: through series learning, one learns one thing after

another to arrive at a total amount of knowledge, while through parallel learning, one learns several things

at the same time to arrive at the same amount of total knowledge. In the machine learning (and artificial

neural networks) literature, Caruana (1995) proposes and tests several mechanisms throughwhich neural nets

learning through multiple related tasks can outperform sequential learning, as it enables a more generalizable

representation of a particular feature. Wason(1960) pioneered the paradigm of rule discovery, which studies

how humans develop hypotheses from observing data from unknown data generation processes. This is also

illustrated well by Baxter (1995), who points out that engaging in multiple tasks enables learning more general

representations of concepts.

4. In particular, Cason et al. (2012) look at (two different) coordination games and find strongspillovers

when the games are played sequentially,but not when they are played simultaneously. In the same vein, Falk

et al. (2013) analyze two identical and completely independent (coordination or public good) games, played

simultaneously, and also find no evidence of behavioral spillovers. By contrast, Bednar et al. (2012) do find

spillovers when (two different) games are played simultaneously, but these games are different from Cason

et al. (2012).