Coordinating Supply and Demand on an On-demand Service...
Transcript of Coordinating Supply and Demand on an On-demand Service...
Coordinating Supply and Demand on an On-demand Service Platformwith Impatient Customers
1
Speaker: Jiaru Bai, UC Irvine, The Paul Merage School of Business
Co-authors: Rick So, UC Irvine, Chris Tang, UCLA , Xiqun Chen, Zhejiang University, Hai Wang, Singapore Management University
Content
• Definition
• Research Questions
• Literature Review
• Model Setup
• Analytical Results
• Numerical illustration: Didi Data
• Summary
2
On-demand Service Platform
3
On-demand Service Platform
• (1) Customers desire quick service
• (2) Many of the platforms use Independent providers
• (3) Use of technology
4
Independent Drivers
Customers Uber
On-demand Service Platform
5
Transportation Delivery
Home Services
Food & BeverageDining & Drinks
Health & Beauty
Some Operating Challenges
6
Quick delivery
• Customers are getting increasingly impatient Choices and competition
Mobile apps
• High degree of variability in both supply and demand
Intricate relationship between endogenous supply and demand
• Set wage and price rates to affect supply and demand
Wage w
Platform
Supply k
Price p
Demand l
CustomersIndependent
Providers
Maximize profit
Maximize earningsMaximize utility
On-demand Service Platform
7
Waiting time
Utilization
Research Questions
1. How to model demand and supply in equilibrium?• Customers are time sensitive
• Service providers are earnings sensitive
2. How should an on-demand service platform set its price p and wage w?
a) When payout ratio = wage/price = w/p is fixed? (e.g., 80%)
b) When payout ratio = w/p is dynamic?
c) What is the benefit of “dynamic payout ratio”?
8
Literature Review
• Sharing economy• Benjaafar et al. (2015), Fraiberger and Sundarajan (2015), and Jiang
and Tian (2015), Li et al. (2015), …
• On-demand service platforms
• Kokalitcheva (2015), Wirtz and Tang (2016), and Shoot (2015), Chen and Sheldon (2015), Moreno and Terwiesch (2014), Terwiesch (2014), Allon et al. (2012), Taylor (2016), …
• Dynamic Pricing
• Riquelme et al. (2015) and Cachon et al. (2015), Hu and Zhou (2016), Gurvich et al. (2015), …
• Two-sided markets in industrial economics
• Rochet and Tirole (2003, 2006), Anderson (2006), …
• Service Pricing with delay costs in operations management
• Naor (1969), Armony and Haviv (2003), Afeche and Mendelson (2004), Zhou et al. (2014), …
9
Modeling Framework
• Customer demand l depends on price p, and waiting time Wq
Independent Drivers
Customers Uber/Didi
Price p
Waiting time Wq
Demand l Supply k
10
Customer Demand Consumer utility: 𝑼 𝒗 = 𝒗 − 𝒑 𝒅 − 𝒄𝑾𝒒
• 𝑣 : Value per service unit with distribution 𝑈[0,1] (parameter)
• 𝑝 : Price per service unit (decision variable)
• 𝑑 : Amount of service units per request (parameter − assumed constant)
• 𝑐 : Unit waiting time cost (parameter)
• 𝑊𝑞: Waiting time (endogenously determined)
Consumer will request if 𝑼 𝒗 = 𝒗 − 𝒑 𝒅 − 𝒄𝑾𝒒 ≥ 𝟎
Equilibrium price: 𝒑 = 𝟏 − 𝒔 −𝒄
𝒅𝑾𝒒
• ҧ𝜆 : Maximum (potential) customer demand rate (parameter)
• l : Realized customer demand rate; 𝜆 ≤ ҧ𝜆 (endogenously determined)
• 𝑠 : service level =𝜆
ഥ𝜆= 𝑃𝑟𝑜𝑏 𝑣 > 𝑝 +
𝑐
𝑑𝑊𝑞 = 1 − 𝑝 −
𝑐
𝑑𝑊𝑞 11
Modeling Framework
• Supply of independent service providers k depends on earnings, which depend on wage w, utilization r
Independent Drivers
Customers Uber/Didi
Wage w
Utilization r
Demand l Supply k
12
Supply of Service Providers Provider’s earning rate:
𝑬 = 𝒘𝝁 𝝆 = (𝒘𝝁)𝝀𝒅
𝒌𝝁= 𝒘
𝝀𝒅
𝒌
• 𝑤 : Wage per service unit (decision variable)
• 𝜇 ∶ Average service speed of providers (parameter)
• k : Number of participating providers (endogenously determined)
• r : Utilization =𝜆𝑑
𝑘𝜇
Provider will participate if 𝑬 = 𝒘𝝀𝒅
𝒌≥ 𝒓
• 𝑟 : Reservation wage per unit time with distribution 𝑈[0,1] (parameter)
Equilibrium wage: 𝒘 = 𝜷𝒌
𝝀𝒅=
𝒌𝟐
𝑲𝝀𝒅
• K : Maximum number of service providers; 𝑘 ≤ 𝐾 (parameter)
• b : Participation ratio =𝑘
𝐾= 𝑃𝑟𝑜𝑏 𝑟 ≤ 𝑤
𝜆𝑑
𝑘= 𝑤
𝜆𝑑
𝑘 13
Modeling Framework
• Platform: How to set price p and wage w?
Independent Drivers
Customers Uber/Didi
Price p Wage w
Demand l Supply k
14
Platform’s Decision Problem• Profit function
𝝅 = 𝝀 𝒑 − 𝒘 𝒅 = 𝝀 𝟏 − 𝒔 −𝒄
𝒅𝑾𝒒 −
𝒌𝟐
𝝀𝒅𝑲𝒅 = 𝝅(𝒌, 𝒔)
Average profit per request Equilibrium price
rate 𝑝Equilibrium wage rate 𝑤
15
Maximize 𝝅 𝒌, 𝒔
Decision variables: 𝑘𝜖𝜆𝑑
𝜇, 𝐾 , 𝑠𝜖[0,1]
• One-to-one correspondence between (p, w) and (k, s)
Modeling Framework• Waiting time Wq and utilization r both depend on supply k and
demand l
Independent Drivers
Customers Uber/Didi
Price p
Waiting time Wq
Demand l
Utilization r=𝜆𝑑
𝑘𝜇
16
Wage w
Supply k
Waiting time 𝑊𝑞
Use M/M/k queueing model
Exact formula too complicated
𝑊𝑞 =1
1 +𝑘! (1 − 𝜌)
𝑘𝑘𝜌𝑘 σ0𝑘−1 𝑘𝑖𝜌𝑖
𝑖!
𝜌
𝜆 1 − 𝜌
Numerical results:
𝑊𝑞 =𝜌 2(𝑘+1)
𝜆 1 − 𝜌𝑤ℎ𝑒𝑟𝑒 𝜌 =
𝜆𝑑
𝑘𝜇
Exact when 𝑘 = 1
Very good estimate for 𝑘 > 1; See Sakasegawa (1977)
Analytical results:
𝑊𝑞 =𝜌 2(𝑛+1)
𝜆 1 − 𝜌𝑤ℎ𝑒𝑟𝑒 𝜌 =
𝜆𝑑
𝑘𝜇𝑎𝑛𝑑 𝑎𝑛𝑦 𝑓𝑖𝑥𝑒𝑑 𝑛
Provide analytical support for our numerical results
17
Waiting time 𝑊𝑞
18
Use M/M/k queueing model
Exact formula too complicated
𝑊𝑞 =1
1 +𝑘! (1 − 𝜌)
𝑘𝑘𝜌𝑘 σ0𝑘−1 𝑘𝑖𝜌𝑖
𝑖!
𝜌
𝜆 1 − 𝜌
Numerical results:
𝑊𝑞 =𝜌 2(𝑘+1)
𝜆 1 − 𝜌𝑤ℎ𝑒𝑟𝑒 𝜌 =
𝜆𝑑
𝑘𝜇
Exact when 𝑘 = 1
Very good estimate for 𝑘 > 1; See Sakasegawa (1977)
Analytical results:
𝑊𝑞 =𝜌 2(𝑛+1)
𝜆 1 − 𝜌𝑤ℎ𝑒𝑟𝑒 𝜌 =
𝜆𝑑
𝑘𝜇𝑎𝑛𝑑 𝑎𝑛𝑦 𝑓𝑖𝑥𝑒𝑑 𝑛
Provide analytical support for our numerical results
Waiting time 𝑊𝑞
19
Use M/M/k queueing model
Exact formula too complicated
𝑊𝑞 =1
1 +𝑘! (1 − 𝜌)
𝑘𝑘𝜌𝑘 σ0𝑘−1 𝑘𝑖𝜌𝑖
𝑖!
𝜌
𝜆 1 − 𝜌
Numerical results:
𝑊𝑞 =𝜌 2(𝑘+1)
𝜆 1 − 𝜌𝑤ℎ𝑒𝑟𝑒 𝜌 =
𝜆𝑑
𝑘𝜇
Exact when 𝑘 = 1
Very good estimate for 𝑘 > 1; See Sakasegawa (1977)
Analytical results:
𝑊𝑞 =𝜌 2(𝑛+1)
𝜆 1 − 𝜌𝑤ℎ𝑒𝑟𝑒 𝜌 =
𝜆𝑑
𝑘𝜇𝑎𝑛𝑑 𝑎𝑛𝑦 𝑓𝑖𝑥𝑒𝑑 𝑛
Provide analytical support for our numerical results
Models and Results
20
1. Base model with a fixed payout ratio
•𝑤
𝑝= 𝛼, 0 < 𝛼 < 1
2. General model with a dynamic payout ratio
• Free w and p
Base Model: Fixed Payout Ratio
21
• Under additional constraint: 𝛼 =𝑤
𝑝, 0 < 𝛼 < 1 (fixed
payout ratio)
• Analytical Result:
• Both the optimal wage rate 𝑤∗ and the optimal price rate 𝑝∗ increase in the maximum demand rate ҧ𝜆 and average service unit 𝑑.
General Model: Dynamic Payout Ratio
Price p* Wage w* Payout Ratiow*/p*
Profit𝝅∗
Max. # providersK
Not monotone ↓ ↓ ↑
Service rateμ
Not monotone ↓ ↓ ↑
Unit waiting costc
Max. demand rateത𝛌
Avg. units requestedd
22
23
Price p* Wage w* Payout Ratiow*/p*
Profit𝝅∗
Max. # providersK
Not monotone ↓ ↓ ↑
Service rateμ
Not monotone ↓ ↓ ↑
Unit waiting costc
Not monotone ↑ ↑ ↓
Max. demand rateത𝛌
Avg. units requestedd
General Model: Dynamic Payout Ratio
24
General Model: Dynamic Payout Ratio
Price p* Wage w* Payout Ratiow*/p*
Profit𝝅∗
Max. # providersK
Not monotone ↓ ↓ ↑
Service rateμ
Not monotone ↓ ↓ ↑
Unit waiting costc
Not monotone ↑ ↑ ↓
Max. demand rateത𝛌
↑ ↑ ↑ ↑
Avg. units requestedd
↑ ↑ ↑ ↑
Extension: Total Welfare
• Total welfare function:
Π 𝑘, 𝑠 = (1 − γ)π 𝑘, 𝑠 + γ(𝐶𝑠 + 𝑃𝑠)
• 𝛾 = 0, basic model (platform profit)
• 𝛾 = 1/2, equal weights on profit and consumer/provider welfare
• ℎ𝑖𝑔ℎ𝑒𝑟 𝛾 = higher weight on consumer/provider welfare25
Platform’s profit
Equitable payoff parameter,0 ≤ γ ≤ 1
Consumer and Provider surplus
Total welfare
Extension: Total Welfare
Main Results:1) For any γ ≤ 2/3, the results in the basic model continue to hold.
2) When the “equitable payoff” γ increases (higher weight on consumer/provider welfare)
The optimal wage rate 𝑤∗ increases
But the optimal price rate 𝑝∗ is not necessarily monotonic.
Optimal payout ratio (w*/p*) increases
Platform profit p* decreases
Social welfare (Cs+Ps) increases26
Didi Company
• Founded in June 2012
• China equivalent of Uber
• The largest on-demand ride-hailing service platform operating in over 360 Chinese cities
27
Didi Data• Hangzhou, capital city of Zhejiang province with over 7 million people
• Sep. 7-13 and Nov. 1-30 in 2015
Peak hour (9am, 7pm):• l=2000• m=19 km/hourNon-peak hour (11pm):• l=1000• m=26 km/hour
28Non-peak HoursPeak Hours
Didi Data• Hangzhou, capital city of Zhejiang province with over 7 million people
• Sep. 7-13 and Nov. 1-30 in 2015
Average travel distance fairly constant during peak and non-peak hours:• d ~ 6-7 km
• Price rate is higher during peak hours
and lower during non-peak hours
• Price rate and demand has a
correlation coefficient of 0.8129
Focus on Express/Private services
Maximum number of available drivers 𝑲≈390
The fixed payout ratio α ≈ 80%
Reservation wage 𝒓 ~ RMB 30 - 40
Value per km 𝒗 ~ RMB 2 – 4
Fixed d = 6km
Two scenarios:Peak hour: ത𝝀 = 200/ℎ𝑟, 𝝁 = 19 km/hour
Non-peak hour: ത𝝀 = 100/ℎ𝑟, 𝝁 = 26 km/hour
30
Model Illustration Based on Didi data
General Model: Optimal price and wage rates
Peak Hour Scenario Non-peak Hour Scenario
c : unit waiting cost
p* : optimal price rate
w* : optimal wage rate
k* : optimal realized number of drivers
31
General Model: Optimal Payout Ratio
Optimal payout ratio a* increases as unit waiting cost c increases
32
Optimal Profit: Fixed vs. Dynamic Payout Ratios
Substantial profit increases when the optimal payout ratio is significantly
different from the fixed ratio!
Optimal a*=0.6
Optimal a*=0.8
33
Summary• A modeling framework for optimal price and wage decisions for an on-
demand service platforms
• Price- and time-sensitive customers
• Independent (wage-sensitive) service providers
• Queueing model to incorporate customer waiting cost
• Some key findings
• Optimal price p* and wage w* are increasing in the max potential demand ҧ𝜆
• Optimal payout ratio w*/p* is also increasing in the max potential demand ҧ𝜆
• Using Didi data, we illustrate that the firm can earn significantly more by using a dynamic payout ratio
• Limitations/Future research
• Dynamic pricing
• Platform competition34
Thank you!
Questions?
35