the peculiar economics of online marketswebfac/eichengreen/e191_sp12/morgan_2 … · pricing 101...
TRANSCRIPT
JOHN MORGAN HAAS SCHOOL OF
BUSINESS UC BERKELEY
The Economics of Platform Competition and Online Markets
All markets are becoming online markets
Travel
Books/music/movies
Cars
Houses
Clothes…
Adventures Online
Some economic stories in the online world?
In Five Acts
Act I: No $500 Bills
Principle 1: No arbitrage
Improved Principle 1: There are no $500 bills lying on the ground
If there were, someone would pick them up.
Testing the No $500 Bill Theorem
Selling the same item on two different platforms (auction sites) should yield the same revenues
A theory of equilibrium coexistence
Ellison, Fudenberg, and Mobius (2004) propose a theory where competing online auction sites can coexist even with vastly uneven market shares
Key intuition: Two competing effects
Size effect: Platform with more users is more attractive
Market impact effect: Switching platforms increases competition
Testing Coexistence versus Tipping
Prediction 1: Expected revenue to a seller should be approximately equal across the two platforms
Prediction 2: The number of bidders attracted per seller should be approximately equal across the two platforms
…all else equal
Goal of the Experiments
Try to come as close as possible to “all else equal”
Choose markets that are relatively thick across platforms
See whether we, as sellers, have a profitable deviation
Determine why this might be the case (buyer/seller ratios)
Morgan Silver Dollars
Product: Popular collectible coins sold on both sites
Search of “Morgan Dollars (1878-1921) produces 12,559 listings on eBay and 1,209 listings on Yahoo!
Other ratios:
Antique books 2:1
Antique firearms 3:1
Beanie babies 20:1
Considerable price variation for these items
Experimental Design
Identical “batches” of Morgan silver dollars slabbed and graded.
8 coins/batch x 11 batches
Sold some batches on eBay and some on Yahoo using identical text, photos, auction characteristics
“Borrowed” reputations from coin dealers
Compare revenues and number of bidders
Dataset was generated by 88 auctions.
88 auctions
48 Yahoo! 40 eBay
32 zero 24 zero 16 positive 16 positive
16 hard 16 soft 8 hard 8 soft
site
reserve
ending rule
treatments
Add photo of coin here
Yahoo versus eBay revenue comparison.
0
20
40
60
80
100
120
Revenues
1 2 3 4 5 6 7 8
Coin
Mean Revenues by Coin
eBay
Yahoo
Yahoo versus eBay number of bidders comparison.
0
1
2
3
4
5
6
7
8
Number of
Bidders
1 2 3 4 5 6 7 8
Coin
Mean Number of Bidders by Coin
eBay
Yahoo
Regressions examining Prediction 1
revenue=β0+β1site+β2reserve+γX+ε
Site = 1 if eBay; 0 if Yahoo
Controls:
Coin dummies
Random effects
Book value
Dealer Price
Equilibrium coexistence implies β1 = 0.
Results of regressions examining Prediction 1
Site coefficient is positive and significant at the 1% level in all specifications
Baseline:
eBay premium = $13.173 (or 26.8 percent)
Baseline + Reserve Price
eBay premium = $14.90 (or 29.5 percent)
Results of regressions examining Prediction 2
Site coefficient is positive and significant at the 5% level in all specifications
Baseline:
eBay attracts 2.125 more bidders than Yahoo
Baseline + Reserve Price
eBay attracts 2.1 more bidders than Yahoo
Does the ending rule matter?
0
1020
30
4050
6070
80
90
Revenues
1 2 3 4 5 6 7 8
Coin
Mean Revenues by Coin
Soft Close
Hard Close
Does the ending rule matter?
The ending rule does not seem to matter
No significant difference in revenues
Coefficients at most 11 cents
No significant difference in number of bidders
No significant difference in the timing of bids
What’s Going On?
Market in the slow motion process of tipping to eBay
Yahoo Press Release, June 15, 2007:
“After careful consideration, we have decided to close down our Yahoo! US and Canada Auction sites to better serve our valued customers through other Yahoo! properties."
Act II: The Law of One Price
Principle 2: In markets for identical products, merchants should offer the same price
If they didn’t, the ones offering the higher price would have no sales and go out of business
Testing the Law of One Price
Price comparison site
Lots of merchants
Identical item
Manufacturer warranty
How E-Retail is different from conventional retail
Speed of price changes
Median duration of a price quote on Shopper.com = 1 day
Speed of location changes
Price comparison sites
Some Stylized Facts About Price Competition Online
Ubiquitous and persistent price dispersion
No single “low-price” firm
Market structure matters
Theoretical rationales
Some Stylized Facts About Price Competition Online
Ubiquitous and persistent price dispersion
Source: Nash-equilibrium.com, July 16, 2007
Value of Price Information
Source: Nash-equilibrium.com, July 16, 2007
Price Dispersion Online – Coefficient of Variation
Figure 1: Time Series of the Coefficient of Variation
0.00
2.00
4.00
6.00
8.00
10.00
12.00
14.00
07/01/2001
09/01/2001
11/01/2001
01/01/2002
03/01/2002
05/01/2002
07/01/2002
09/01/2002
11/01/2002
01/01/2003
03/01/2003
05/01/2003
07/01/2003
09/01/2003
11/01/2003
01/01/2004
03/01/2004
05/01/2004
07/01/2004
09/01/2004
11/01/2004
01/01/2005
03/01/2005
05/01/2005
07/01/2005
Collection Date
Mean
Co
eff
icie
nt
of
Vari
ati
on
(Perc
en
t)
Some Stylized Facts About Price Competition Online
Ubiquitous and persistent price dispersion
No single “low-price” firm
Some Stylized Facts About Price Competition Online
Ubiquitous and persistent price dispersion
No single “low-price” firm
Market structure matters
Market Structure Matters
Figure 5: Price Gap by Number of Firms
0
5
10
15
20
25
30
2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50 52 54 56 58 60
Number of Firms Listing Prices
Avera
ge P
rice G
ap
(Perc
en
t)
Source: Baye and Morgan (2004)
Theoretical Rationales
Clearinghouse models of price dispersion e.g., Varian (1980), Rosenthal (1980), Narasimhan (1988)
Tradeoff between selling to “loyals” and attracting “shoppers” combined with comparison sites Cutting prices attracts shoppers but loses margin on loyals Keeping prices high to loyals gives up on price conscious market
Leads to “hit and run” pricing strategies Fun fact: The economics mostly predates e-retail!
Information “Gatekeepers”
Baye and Morgan (2001)
The platform (information gatekeeper) plays an important role in competition Optimal pricing by “gatekeeper” induces price dispersion
Inducing too much price competition is bad for the platform
Price dispersion persists even when: All consumers are “shoppers”
The product market is “competitive”
Pricing 101
Basic pricing theory
Know your marginal cost
Know the elasticity of demand for your consumers
Use simple MR = MC condition
Complications
Many customer segments
Rivals’ pricing strategies
Dashboard for online pricing
Baye, Gatti, Kattuman, and Morgan (2007)
How elastic is firm demand online?
Key difficulty: Hard to see actual demand easier to see clicks
Most leading comparison site price per click compared to per conversion
Clicks versus purchases
Time delay in clicks v purchases
Oxford Lasers
But still, how elastic is demand online?
Answer 1: Amazingly elastic
Ellison and Ellison (2004) estimate elasticities of -25 to -40 for computer memory
Answer 2: Not that elastic
Chevalier and Goolsbee (2003) estimate Amazon’s demand elasticity (for books) to be -0.6.
Why such a difference?
Market Structure Matters
Books sold online:
Concentrated market
Heavy branding activity
Direct sales
Repeat customers
Computer memory sold online:
Fragmented market
Little branding among retailers
Comparison site sales
Sophisticated consumers
Who cares what the elasticity of demand online is anyway?
Competitiveness of online markets:
“The explosive growth of the Internet promises a new age of perfectly competitive markets. With perfect information about prices and products at their fingertips, consumers can quickly and easily find the best deals. In this brave new world, retailers' profit margins will be competed away, as they are all forced to price at cost.”
The Economist, November 20, 1999, p. 112.
Our Dataset
Data from Kelkoo’s leads log
Every time a ‘lead’ is generated, record taken of: time, cookie, merchant, product, price, location on the screen.
Results presented today
18 pda’s over period Sept 2003 – Jan 2004.
19 merchants gained over 34,000 ‘leads’ from over 18,000 ‘cookies’.
Wish list:
Larger, cleaner dataset using leads logs
Extended analysis generated for 50+ pda’s and 100+ digital cameras over period Sept. 2003 to Sept. 2004.
What’s it like to shop at Kelkoo?
When do people click?
At work
Proportion of leads by hour
0
2
4
6
8
10
1 3 5 7 9
11
13
15
17
19
21
23
Hour of Day
% o
f a
ll le
ad
s
How do people choose what firm to click on?
0
5
10
15
20
25
30
35
40
45
50
1 2 3 4 5 6 7 8 9 10
Price rank / Screen Location
% o
f all
lead
s
Price rank
Screen Location
Both price and screen location seem to matter
What determines leads?
SALES
CLICKS
Christmas
Date
Weekend
Position
on Screen
Price
# of
Competitors
Product
Popularity
Accumulated
Brand Equity
Restocking &
Return Policy
CONVERSIONS
Information on
firm’s site
Information on
rivals’ sites
Ease of UseLook & Feel
Offline
Presence
Information on Kelkoo Site (X) Information on Sellers’ Sites (Z)
Determining Leads
Preliminaries: Qijt denotes the number of leads for firm i, product j, on date t.
Since over half of all leads are zeros or ones, count data is appropriate
We use the Pseudo-maximum likelihood (PML) approach (Gourieroux, et al. 1984) and assume: E[Qijt | X] = exp[X β]
How do I estimate demand from clicks?
Let X denote the information available from the Kelkoo site Includes all competing firm’s prices
As well as some reputation information about all firms
Let Zi denote the additional information a consumer learns about retailer i from clicking through to its site
Where does demand come from:
Suppose that the expected number of leads to firm i are
E[Qi | X]
Suppose the “conversion rate” of firm i’s clicks is
Pr[Buy | X, Z]
Then firm i’s expected demand is
E[Di] = Pr[Buy | X, Z] x E[Qi | X]
Relating Leads to Demand
Proposition:
Suppose that the convert rate is independent of some firm-specific variable xi listed on the Kelkoo site, then the elasticity of demand with respect to xi can be estimated solely with clicks data.
Main application:
Suppose xi = pi, firm i’s price, then demand elasticity can be estimated from clicks alone
Demand elasticities for each product?
Suppose that demand is given by: E[Qijt | Xijt] = βj x ln[priceijt] + γj X1,ijt
Controls: Screen location
Month
Weekend
Note: Demand is assumed to be continuous---no jumps
What accounts for the elasticity differences?
-15
-13
-11
-9
-7
-5
-3
-1
1
1 2 3 4 5 6 7 8 9
Average Number of Firms Listing a Price
Ela
sti
cit
y
Coefficient Estimate Significant at the 1% Level
Coefficient Estimate Not Significant at any Conventional Level
How to explain the “elasticity paradox”
The larger the number of competing firms, the closer the degree of substitution Hence the more elastic is firm demand
Formal “toy” model N identical capacity constrained firms compete in a market
where goods are substitutes
Firm elasticity = N x Market elasticity
Estimating the elasticity of demand pooled over all PDAs
We estimate: E[Qijt | Xijt] = [β0 + (njt -1) β1] x ln[priceijt] + β2 njt + γ X1,ijt
njt denotes number of competing firms
Controls: Same as above, plus
PDA model x month
Bricks and clicks retailer
Results
Model 1 Model 2
Likelihood Specification for Clicks Poisson PML Poisson PML
Log Total Price -4.61 -3.761(8.91)** (7.45)**
Log Total Price x (Number of Sellers ? 1) -0.288(4.14)**
Number of Sellers 1.593(4.05)**
Position on Screen -0.186 -0.175(4.54)** (4.47)**
Bricks and Clicks Retailer 0.262 0.236(1.58) (1.67)
Weekend -0.242 -0.265(10.82)** (11.46)**
Product Dummies 17 17
Month Dummies 4 4
Product x Month Dummies 55 55
Robust Standard Errors Clustered by Firm Yes Yes
Observations 6151 6151
Overdispersion Test Test Statistic 2656.46 2488.77
P-Value 0 0
Table 3: Continuous Clicks Specifications
Note: Robust z statistics in parentheses. * Significant at 5%; ** Significant at 1%
How should I interpret these results?
Recall that in books, there are effectively fewer than five firms competing
In computer memory hundreds are competing
We find:
Monopoly firm price elasticity = -3.761
Ten firm price elasticity = -6.641
Additional Interpretations
Adding one firm charging the average price costs a firm about 4% of its clicks
Moving a firm down one screen location costs it about 17.5% of its clicks
Having a bricks and clicks presence raises clicks by 25%
What about the jump in demand?
Many theoretical models (Varian, 1980; Rosenthal, 1980; Narasimhan, 1984; Baye and Morgan, 2000) suggest a “jump” in demand at the lowest price
But the previous specifications mostly ignored this
This has the potential to bias the elasticity estimates
Ln(Price)
Ln(Quantity)
Best Price
‘Shoppers’
Firm i’s
Estimated
Demand Firm i’s Actual
Demand
Misspecified demand leads to estimates
that are “too elastic”
How we estimated discontinuous demand
βx'
epQ jj
Suppose there are two types of consumers: Shoppers (S), who buy at the lowest price, and Loyals (L) who do not
Assume that the number of clicks from each group is given by:
More on estimating discontinuous demand
L
sL
SLSL
epDq
ppD
epDepQQq
ere wh)1(
otherwise 0
if 1 where
min
βx
βxβx
'
''
]+p)ln 1)-(n+[(e]|[ ,12ijt1jt0
'γXX ijtjtijt DnxpQE
This yields the estimating equation:
Results Model 1
Likelihood Specification for Clicks Poisson PML
Log Total Price -2.459(9.11)**
Log Total Price x (Number of Sellers ? 1) -0.252(4.60)**
Jump from Shoppers 0.603(7.11)**
Number of Sellers 1.415(4.52)**
Position on Screen -0.175(4.37)**
Bricks and Clicks Retailer 0.321(2.41)*
Weekend -0.268(13.79)**
Product Dummies 17
Month Dummies 4
Product x Month Dummies 55
Robust Standard Errors Clustered by Firm Yes
Controls for Unobserved Firm Heterogeneity No
Observations 6151
Overdispersion Test Test Statistic 1942.27
P-Value 0
The Value of Unpredictability
Source: “A Dashboard for Online Pricing” by Baye, Gatti, Kattuman, and Morgan
Act III: Plus Shipping and Handling
Principle 3: The total price, including all add-ons, extras, etc. determines demand
Reason: All the dollars are just as green
Hidden Fees
BMI Baby
List price for flight from Nottingham to Edinburgh: Listed Price = ₤4.99
What wasn’t mentioned in the offer price: “Taxes, fees, and charges”
Checked bag fee
Credit card fee
Lots of other fees
What I actually paid: Real Price = ₤29.84
Southwest
Why Does BMIBaby Do This?
Price components reflect incremental services and costs
Allows consumers to select what they want and opt out of what they don’t want
Why Does BMIBaby Do This?
Strategically decompose and shroud prices to profit from “naïve” consumers
Other Examples of Price Decomposition and Price Shrouding
Hotels – room rate, tax, mini bar pricing
…Plus shipping and handling on TV offers
Hidden fees at banks, car rental companies, restaurants
“Energy” fees on airlines
Central Question
Does these practices work?
If so, when?
How big an effect?
Theories
Milgrom, Jovanovic Non-disclosure “unravels” with sophisticated
consumers
Gabaix and Laibson (2006) Competition for sophisticated and naïve consumers
leads to shrouding Ellison (2005)
Consumer brand and quality preferences create opportunity for “add-on” shrouded pricing
Kahneman and Tversky (1984), Thaler (1985) Separate mental accounts create incentives for differing
price decompositions
Our Approach
Run field experiments at online auction sites around the world
Vary the decomposition and shrouding
Study a natural experiment on eBay US
Rules change at eBay made shipping charges less shrouded
Price Decomposition
Online auctions offer a natural venue for examining price decomposition
The reserve price in an auction is equal to: Minimum opening bid + shipping charge
Strategic equivalence implies that only the reserve price should affect revenues in the auction
Price Decomposition (2)
Varying the shipping and minimum opening bid allows us to examine price decomposition effects in a setting where such changes are, theoretically, revenue neutral
Shrouding
Online auctions also offer a natural venue for shrouding Shipping is typically less salient than the current
auction price
disclosed: Disclose shipping in the header of the listing
Shrouded: Disclose shipping in the body of the listing.
Why Online Auctions?
Hide in the crowd:
Large number of ongoing auctions of similar items
Variation in selling practices
Natural manipulation price components and shrouding
Bidders are familiar with the rules
A transparent price discovery process
Bids convey information about willingness to pay
Field Experiments
2006 – Taiwan
36 iPods of various models
2008 – Ireland
40 iPods of various models
Describe Taiwan in detail, Ireland similar
Taiwan Experiments: Design
One “seller” who is not a newcomer
Identical auction rules except for opening bid and shipping fee
Identical descriptions for each item
Two sets of treatments with “shrouded” and “disclosed” shipping fee
Within each set, three treatments varying the opening price and shipping fee combinations
Taiwan Experiments: Objects
We chose goods where 1 shot private value assumptions are a reasonable approximation
Not “expertise” goods
Resale does not play a large role
Bidders have unit demand
Multiple identical products available
Large number of buyers and sellers
Chose 6 different iPod models
Taiwan - Treatments
Screenshot – Disclosed
Screenshot – Shrouded
Decomposition Effects Disclosed Treatment
Item DL DR
iPod nano 1G black 4,380 4,530
iPod nano 1G white 4,330 4,630
iPod nano 2G black 5,430 5,480
iPod nano 2G white 5,430 5,580
iPod shuffle 1G 3,130 3,100
iPod shuffle 512m 2,190 2,210
Opening Price High Low
Shipping Fee Low High
Shrouding No No
Decomposition Effects: Shrouded Treatment
Item SL SR
iPod nano 1G black 4,080 4,530
iPod nano 1G white 4,330 4,480
iPod nano 2G black 5,230 5,500
iPod nano 2G white 5,230 5,480
iPod shuffle 1G 3,080 3,280
iPod shuffle 512m 1,860 1,980
Opening Price High Low
Shipping Fee Low High
Shrouding Yes Yes
Tentative Conclusions
Price decompositions matter whether prices are shrouded or not
Shrouding Effects High Opening, Low Shipping
Item DL SL
iPod nano 1G black 4,380 4,080
iPod nano 1G white 4,330 4,330
iPod nano 2G black 5,430 5,230
iPod nano 2G white 5,430 5,230
iPod shuffle 1G 3,130 3,080
iPod shuffle 512m 2,190 1,860
Opening Price High High
Shipping Fee Low Low
Shrouding No Yes
Shrouding Effects Low Opening, High Shipping
Item DR SR
iPod nano 1G black 4,530 4,530
iPod nano 1G white 4,630 4,480
iPod nano 2G black 5,480 5,500
iPod nano 2G white 5,580 5,480
iPod shuffle 1G 3,100 3,280
iPod shuffle 512m 2,210 1,980
Opening Price Low Low
Shipping Fee High High
Shrouding No Yes
Shrouding Effects High Opening, High Shipping
Item DH SH
iPod nano 1G black 4,580 4,080
iPod nano 1G white 4,480 4,480
iPod nano 2G black 5,380 5,380
iPod nano 2G white 5,980 5,580
iPod shuffle 1G 3,380 3,280
iPod shuffle 512m 2,160 2,180
Opening Price High High
Shipping Fee High High
Shrouding No Yes
Ireland Experiments – Key Differences
Less product mix:
2G ipod shuffles in various colors
Less extreme shipping
25/75th percentile
Minimal opening bid
Within week treatment differences
Colors alternated treatments each week.
Ireland Experiments - Treatments
Summary Statistics
Pooled Results
Conclusion
Transferring a larger portion of the price to shipping fees increase revenue, market competition does not eliminate the framing effect
Ignoring shrouded prices seem to be more important than mental accounting
The impact of this framing effect can be removed by small institutional changes that make all price attributes transparent
When possibility of shrouding is eliminated, revenue seems to increase
Results hold across cultures and markets
Act IV: What’s in a Name?
Principle 4: Reputation signals are valuable when they’re cheaper for the good guys to acquire than for the bad guys
If not, the bad guys would have the same “reputation” as the good guys
Studying Reputation
EBay’s reputation system key to competitive advantage
What does it cost good guys versus bad guys to gain reputation?
A Market for Feedback
Advice from the eBook
“Look on eBay for items that cost next to nothing. You can
find the eBay search feature to find items which cost
anywhere from .01 to $1.00. Try this. ... Now bid on 100
items. If you want to speed things up a bit, try and find
auctions with the "Buy It Now" option. If the seller offers
PayPal as a form of payment, go right away and pay for the
item. ... If you do this with a hundred different sellers you
should be able to get your feedback score up to 100 in just a
few days.”
The Strategy of thelandseller
Create reputation through penny transactions in the market for feedback
thelandseller reputation = 598
Cost of reputation: $360
The New Strategy of thelandseller
Sell “lakefront property in Texas”
Benefit from reputation: Raise price by up to 5%
Limits to Trust?
Big ticket items are key growth driver for eBay
Trust is crucial for these items
But benefit of pretending to be a reputable seller is also great in these markets
Key challenge: Can the value of reputation be sustained in these markets?
Act V: First Mover Advantage
Principle 5: In network markets, the first mover wins even if it is worse
Famous Example:
QWERTY versus DSK
Properties of Network Markets
Platforms (Two-sided networks)
2 types of users
Matching technology
Scale improves match efficiency
Users of one type benefit from more of the other type
Users of one type are harmed by more of the same type
What are platforms?
Online auctions
Display advertising exchanges
Financial exchanges
Dating sites
Gaming consoles
Search engines
Forces Driving Consolidation
Scale and Size Effects
Sellers benefit from more buyers
Buyers benefit from more sellers
Both sides benefit from scale
Platform competition leads to monopoly
Forces Opposing Consolidation
Market Impact Effect Provides Checking Force
Sellers hurt by more sellers
Buyers hurt by more buyers
Platforms can coexist in equilibrium
Compensating price differences across platforms
Market Structure of Platforms
Market Concentration
More tipped Less tipped
What Accounts for these differences?
Market impact effects
Price differences
Differentiation
Path Dependence
A Simple Platform Game
Number of players of the player's
own type (including herself) in the
platform she joined
1 2
Number of
players of the
opposite type
in the platform
the player
joined
0 5 5
1 9 6
2 12 11
The subscription fees are, pA = 4 and pB = 2
Tipping is an Equilibrium
Number of players of the player's
own type (including herself) in the
platform she joined
1 2
Number of
players of the
opposite type
in the platform
the player
joined
0 5 5
1 9 6
2 12 11
The subscription fees are, pA = 4 and pB = 2
Coexistence is an Equilibrium
Number of players of the player's
own type (including herself) in the
platform she joined
1 2
Number of
players of the
opposite type
in the platform
the player
joined
0 5 5
1 9 6
2 12 11
The subscription fees are, pA = 4 and pB = 2
An Amended Platform Game: Reduced Market Impact
Number of players of the player's
own type (including herself) in the
platform she joined
1 2
Number of
players of the
opposite type
in the platform
the player
joined
0 5 5
1 9 8
2 12 11
The subscription fees are, pA = 4 and pB = 2
An Amended Platform Game: No Coexistence
Number of players of the player's
own type (including herself) in the
platform she joined
1 2
Number of
players of the
opposite type
in the platform
the player
joined
0 5 5
1 9 8
2 12 11
The subscription fees are, pA = 4 and pB = 2
Key Features
Coexistence turns on:
Size of market impact effects
Price differences
Differentiation of platforms
Key Questions
Does the size of the market impact effect explain market structure?
If tipping occurs, can we predict the winning platform?
What are the dynamics of platform competition?
Experimental Design
A “market” consists of 4 subjects
Subjects are assigned a “type”
Two are “squares” and two are “triangles”
Two competing platforms
Named “firm #” and “firm %”
Platform competition lasts 15 periods
Subjects simultaneously select a platform in each period
Experimental Design (2)
After 15 periods, subjects are randomly re-matched into new “markets”
Repeat platform competition game
Each 15 period block constitutes a “set”
Four sets per treatment
Feedback
Subjects know:
Payoff matrix for each platform
Access fees for each platform
Result of the previous round of the platform competition game including market outcome and realized gross and net payoffs
Treatments
Varied:
The size of the market impact effect
Platform payoff matrices
Order
Typical session:
Set 1: Non-tipped equilibrium (N)
Set 2: Only tipped equilibria (T)
Set 3 = Set 1
Set 4 = Set 2
Treatment 1: Homogeneous Platforms
Platforms have the same match technology (payoff matrices)
Differ in access fees (so there is a Pareto ranking)
Vary market impact effect to turn on and off coexistence
Homogeneous Setting: Session-level Results
Time Series of Platform Choice Throughout the Sessions
60%
70%
80%
90%
100%
110%
1 5 9 13 17 21 25 29 33 37
Period
Mark
et
Sh
are
of
the C
heap
er
Pla
tfo
rm NTNT Sessions
TNTN Sessions
Homogeneous Setting: Market-level Results
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
1 2 3 4
Other
Pareto Dominant
Homogeneous Markets
Set
Percent
of
Markets
Tipped
Platforms are Not Created Equal
Perhaps coordination is easy because platforms differ only in their access fees
Test: Create “differentiated” platforms by varying payoff matrices
Platforms vary in matching efficiency as well as access fee
Net payoff is relevant for efficiency
Treatment 3: Vertically Differentiated Platforms
Number of players of the player's own type
(including herself)
1 2
Number of
players of
the
opposite
type
0 (6, 3) (6, 3)
1 (10, 9) [(7, 6)] {(9, 8)}
2 (13, 12) (12, 11)
The subscription fees are, pA = 5 and pB = 2
Key Features
Platform B is less efficient
Tipping to B is Pareto dominant (owing to cheaper access fees)
In N treatment, 50-50 market share is also an equilibrium.
Differentiated Setting: Market-level Results
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
1 2 3 4
Other
Pareto Dominant
Differentiated
Set
Percent
of
Markets
Tipped
Potential Confound
Pareto dominant platform is also the cheaper platform
Heuristic strategy: Go to the cheaper platform
Produces same predictions
Treatment 4: Differentiated – Cheap
Change payoff matrices so that the Pareto dominant platform is also more expensive
Will subjects learn that “paying for quality” is optimal?
Differentiated-Cheap Setting: Payoff Matrices
Number of players of the player's own type
(including herself)
1 2
Number of
players of
the opposite
type
0 (4, 4) (4, 4)
1 (11, 8) [(8, 6)] {(10, 6)}
2 (13, 11) (12, 10)
The subscription fees are, pA = 3 and pB = 2
Differentiated-Cheap Setting: Session-level Results
Pareto Dominant Platform Choice in the Differentiated-
Cheap Treatment
50%
60%
70%
80%
90%
100%
110%
1 6 11 16 21 26 31 36 41 46 51 56
Period
Ma
rke
t S
ha
re o
f th
e
Mo
re E
xp
en
siv
e
Pla
tfo
rm NTNT Sessions
TNTN Sessions
Differentiated–Cheap Setting: Market-level Results
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
1 2 3 4
Cheap
Pareto Dominant
Differentiated-Cheap
Set
Percent
of
Markets
Tipped
Path Dependence
Expectational lock-in
First to market has an advantage in creating expectations
Positive feedback cycle
Durable advantage for the first-mover platform…
Even if it is inferior
First-Mover as a Business Strategy
Amazon
“Get big fast.” – Jeff Bezos
Dot-com Biz Plans
Eyeballs rather than profits as success metric
Vaporware
Microsoft “version 1.0” products
First-Mover Advantage
In first 5 periods, only one platform is active
Subjects can only choose the active platform during this time.
Starting in period 6, an “entrant” platform arrives and subjects can choose either platform
Treatment 6: Differentiated – First Mover
Replicate differentiated treatment
But make Pareto dominant platform second mover
Treatment 6: Results
Treatment 7: Differentiated Cheap – First Mover
Replicate differentiated cheap treatment
Pareto dominant platform second mover
Treatment 7: Results
Alternative Explanation
Novelty heuristic
Subjects coordinate on the “new” platform in period 6
Test: First mover is always the cheaper platform
Treatment 8: Incumbent is cheaper but Pareto Superior
Is First-Mover Advantage Worth Anything?
Compare to treatments where neither platform has a head start
Head Start: Cheaper Platform is Pareto Superior
Head Start: Cheaper Platform is Pareto Inferior
Tentative Conclusions (2)
There is no evidence for QWERTY
The (Pareto) inferior platform triumphs 0% of the time
There is no evidence for expectational lock-in
Mild evidence for first-mover disadvantage
Surplus is what counts
100% of markets converge to the platform offering the higher surplus
Horizontal Differentiation
We consider a market where agents have different preferences for the platforms
Two platforms with different gross payoff matrices
Access fees are different for different agents
A pair of square and triangle players have the same set of access fees and another pair of square and triangle players have another set of access fees
Heterogeneous Agents: Market-level Results for PD
Tipping Under PD
0.0%
20.0%
40.0%
60.0%
80.0%
100.0%
1 2 3 4
Set
Pe
rce
nt
of
Mar
kets
Tip
pe
d
Platform #
Platform %
Heterogeneous Agents: Market-level Results for ND
Tipping Under ND
0.0%
20.0%
40.0%
60.0%
80.0%
100.0%
1 2 3 4
Set
Pe
rce
nt
of
Mar
kets
Tip
pe
d
Platform #
Platform %
Heterogeneous Agents: Market-level Results
Efficiently Coexisting Platforms
0.0%
20.0%
40.0%
60.0%
80.0%
100.0%
1 2 3 4
Set
Pe
rce
nt
of
Mar
kets
Co
exi
stin
g
Under PD Under ND
Conclusions
When platforms are undifferentiated or mainly vertically differentiated:
Markets mainly tip to Pareto dominant platform
Theoretical coexistence is behaviorally irrelevant
First mover advantage is worthless
When platforms are mainly horizontally differentiated
Coexistence is the norm
Epilogue
The internet offers the world’s biggest and best laboratory
Ask yourself why
Do it yourself
Be open to new possibilities
You can never have enough data