Case Study: Innovation at Uber: The

Launch of Express POOL

Prof. Sean Zhou

Decisions, Operations and Technology

CUHK Business School

Learning Objectives
• Learn the product development/innovation process inside a
platform–based firm
• Identify how to assess the benefits of innovation, in the
aggregate and separately for the different user groups in a multi-
sided platform
• Familiarize different ways of collecting information and data
• Understand the roles of experimentation and data analytics in the
decision-makings in platform-based firms

2
Uber
• Founded in 2009 in Silicon Valley, U.S., a first mover in the fast
growing ride sharing industry, 77% market share in the U.S.
• On demand service that connects idling independent drivers and
cars with waiting passengers for a fee
• Main products: Uber Ride/Drive, Uber Eat, Uber Freight
• Global competitors: Lyft, DidiChuXing, Ola (India), Grab
(Southeast Asia), Go-jek (Indonesia)
• Provide luxury service at the beginning, introduce UberX in 2012,
cheaper rides on average (than taxi)
• In 2022, monthly active platform consumers are 131 million, 2.1
billion trips per month, annual revenue is about 31.9 billion

3
What is Uber’s value proposition?

4
Business Model: Value Creation

Many many
Many many riders Ridesharing Platform
drivers

Value Proposition to Value Proposition to Both Riders and

Value Proposition to Drivers
Riders Drivers
• Trust
• Good money (good prices
• Reliable (can follow car on app)
• Quick and constant business)
• Hassle-free payment
• Convenient • Flexible hours (I work when I
• Simple (same app everywhere)
• Cheap ride want)
• Safety (can trust drivers/riders)
• Drivers know where • Guaranteed payment
• Transparency (of
to go • Tell me where to go
prices/locations...)
• LOTS OF DRIVERS (information)
• Social component (talk to nice
• LOTS OF RIDERS
people)
Operations
• Search and matching (algorithm)
• Pricing (set prices, payment infrastructure)
• Trust (ex-ante, during, ex-post: screening / reviews / dispute resolution)
• [ride-sharing specific] GPS guidance to pick-up and destination
5
Uber Platform and Products
• Ride/Drive: Express and Uber POOL, UberX, UberXL, and
UberSelect and UberBLACK, UberSUV, and UberLUX
• Uber keeps 25% of the gross fare each ride
• Its driver/partner median earning is $15-$30 per hour
• Dynamic pricing, use ``surged” price when demand outpaces
supply, e.g., rush hours, raining/snowing days
• Organization structure: rider, driver, and marketplace

6
Uber App: Rider View

7
Uber App: Driver View

8
What is Express? Why is Uber launching Express?

9
Express POOL Project
• Still a carpool product, but not door-to-door, riders wait up to 2
mins (matching time, not driver enroute time to pick up riders),
walk a short distance to/from their pickup/drop-off points
• Why?
• POOL is losing money
• Reduce number of detours (POOL matches other riders on the way), increase the
efficiency of matching (waiting, from one-to-many, to many-to-many), increase
the seat utilization
• Reduce costs while improving profitability

• Pilot launch in Boston and San Francisco in Nov. 2017

10
Express POOL User Interface

11
What is the product innovation process in Uber?

12
Internal 6-city launch + wait time
Tiger Team V0 Tiger Team V1 Trip parties SF/Boston Pilot
Conversation/ideation experiment

6-City Launch + Wait Time

Internal Conversations Tiger Team V0 Tiger Team V1 Trip Parties SF/Boston Pilot
Experiment
Input:
• Data from past
POOL requests to
simulate trips as-if
Input: the new algorithm Input:
• JIT algorithm was in place Input: • Actual requests
• Conversations • Demand elasticities • Actual requests compared to
Input:
Input: with product from other Uber and market similar cities
• Conversations
• Strategic managers products outcomes without Express
with product
priorities Output: • Additional Output: POOL (Synthetic
managers
• Existing similar • New matching willingness-to-pay • Debugged control
Output:
products algorithm (from from conjoint algorithm experiment)
• New matching
Output: 1- to-many to surveys • Initial results on Output:
algorithm
• Add waiting and many-to- many • Team analysis of actual • Platform-wide
with flexible
walking matching) with particular trips customer results on actual
waiting
fixed waiting Output: experience and customer
• Flexible “pick-up • Simulated efficiency experience and
points efficiency gains efficiency
• Starting waiting
and walking
parameters (wait
for 2 mins)
• Initial pricing
13
Express POOL (Cont.)
• Well received in San Francisco, lukewarm in Boston (weather?)
• Feb. 2018, synthetic control launched in 12 cities: 6 treatment and 6
control
• 5-week freeze on experimental changes
• Switchback experiment in Boston to test waiting time of two vs. five
minutes
• Three weeks before the end of the freeze at 12 cities, the result from
Boston measuring the cost and benefit of increasing wait time from
2 to 5 mins came in
• Cancellations increase but the cost of Uber per ride decreases

• Decision: whether to keep the launched experiment unmodified or

overrule the 5-week freeze period and increase the wait time from 2
to 5 mins in the 6 treatment cities?

14
Innovation at Uber
• Uber made continuous, incremental improvement of its core
products, relying heavily on data analytics
• Substantial investment in data science, 60 out of 200 in its
marketplace team
• Each product team: product manager, data scientist, engineer,
product operation manager, designer, and marketing team
• Three ways of collecting data: Survey, simulation, experiment
• Three main types of experiments
• User level A/B: individual user is an observation unit, e.g., design change of an app
• Switchbacks: evaluate the effect of product tweak on some outcome variable of
interest, 160 mins time interval in a given city is an observation unit
• Synthetic control: create control and treatment cities to evaluate the effect of product
tweak on a set of outcome variables of interest, entire city is an observation unit

15
What are the pros and cons of each type of
experiments run by Uber?

16
Characteristic Technical Term User A/B Switch Synthetic
backs Controls
1. Ability to reduce contamination between Treatment and control interference /
treatment & control groups spillover effects /
2. Ability to measure effects on market
equilibrium (e.g. average prices and total
matches)

3. Number of observations Statistical power

4. Ability to detect small effects with precision

5. Experiment duration

6. Number of experiments that can be run

concurrently

7. Ability to perfectly randomize

8. Economic costs (potential losses)

17
Why did Uber launch the switchback experiment in
Boston if they were already testing Express with a
synthetic control experiment in 12 cities?

18
Consider the supplementary data set, what is the
effect of extending wait time from 2 mins to 5 mins
on the total number of shared rides completed,
number of cancellations, the proportion of shared
rides that were matched, and driver payout per trip?

19
Hypothesis Generation

Effect of moving from 2 to 5-minute waiting on outcomes of interest.

Customer Experience Metrics Efficiency Metrics

-Cancellations. Substitute towards: -Cost per trip
• POOL -Earning per trip
• UberX -Match rate
• Lyft -Double match rate
• Other means of transportation -Seat utilization during trips
-Total Shared trips:
• Express POOL trips
• POOL trips
-Ratings
-Complaints

20
Data Analysis

T test
Avg 5 minutes Avg 2 minutes Avg 2 minutes - Avg 5 minutes 95% Cl lower 95% Cl upper
t p-value
bound bound
Ridesharing
3,880 3,967 -87 -0.85 0.397 -289 116
trips

POOL trips 1,461 1,355 106 2.33 0.021 16 195

Express POOL
2,419 2,612 -193 -2.20 0.030 -365 -19
trips
Rider
190 165 25 2.76 0.007 7 43
cancellations

Cost per trip $7.00 $7.36 -$0.36 -3.46 0.001 -$0.59 -$0.16

Match rate 0.62 0.66 -0.03 -2.52 0.013 -0.06 -0.01

Double match
0.35 0.32 0.03 2.48 0.014 0.01 0.05
rate

21
What would you recommend that Stock do? Should
he increase the wait time from 2 to 5 mins in the six
treatment cities for the remaining 3 weeks?

22
• For
• Extending to 5 mins to make at least $1.6 million
• More data on the effect of 5 mins wait time

• Against
• If extending to 5 mins in 6 cities, the experiment results suggest that riders do not
like Express
• Is this due to the extended wait time or to the fact that riders are asked to
wait and walk?
• Should Uber abandon Express altogether or revert to a shorter wait time?
• This worst case scenario has significantly more costs than waiting three more
weeks
• Patience is key for obtaining definitive data on hypothesis

23
Case Epilogue
• Add the five minutes wait time option to riders in Boston and San
Francisco, make no changes to the app in the six treatment cities
• Since launching Express, Uber integrated minimal waiting across
its product offerings given the benefits, all Uber products adopted
the many-to-many matching algorithm developed for Express
• Incorporate Express into POOL product and drop the explicit
distinction between Express and POOL
• Under POOL, customers can choose whether they were wiling to walk and wait
for price discounts
• Explicit wait times were dropped, giving Uber flexibility

• The Express development had far reaching implications for Uber

24
25
Platform/Marketplace/Sharing Economy
• Platforms: help match many fragmented sellers and buyers of
goods or services while providing related services
• Develop and launch new product/service frequently with fast
iterations/improvements
• Operations and decision makings such as new product/service
launching, search and matching, advertising, flexible or auction-
based pricing, etc., are supported by data analytics,
experimentation, machine learning

26
Examples
• Retailing: Amazon, Ebay, Taobao
• Ride hailing: Uber, Lyft, DidiChuXing, Grab
• Food/grocery delivery: Deliveroo, UberEats, DoorDash
• Vacation rental: Airbnb
• Video sharing: Youtube, Tiktok, Kwaishou
• Logistics service: Cainiao
• …

27
Revenue Models
• Commissions from matching supply and demand
• Charges of other value-added services, e.g., IT,
logistics/fulfillment/delivery, promotion, customer traffic
management
• Advertising

28
Operational Challenges
• Multi-sidedness: Managing both supply and demand, e.g., no
supply then no demand, and no demand then no supply,
sometimes their interests are not aligned
• Pricing, matching, assortment decisions need to be made in
almost real-time
• Online-offline integration (e.g., logistics, customer services) and
user experiences
• Customer/seller experience is critical as switching costs are often
low; so continuous improvement is needed

29