This document discusses predictive analytics for vehicle price prediction delivered continuously at AutoScout24. It describes AutoScout24's use of a random forest model for price prediction and their approach to automatically generating Java code from the R-based model to deploy it as a high-performance web application via a continuous delivery pipeline. Key lessons learned include forming cross-functional data science and engineering teams, setting up early usage reporting to improve the product, and addressing challenges of generating large amounts of Java code like optimizing garbage collection.

1. Predictive Analytics World London| Oct 12, 2016 | Arif Wider & Christian Deger
Predictive Analytics for Vehicle Price Prediction
Delivered Continuously at AutoScout24

2. Dr. Arif Wider
Senior Consultant/Developer
awider@thoughtworks.com
@arifwider

3. Christian Deger
Chief Architect
cdeger@autoscout24.com
@cdeger

4. PL
S
RUS
UA
RO
CZ
D
NL
B
F
A
HR
I
E
BG
TR
18countries
2.4m+cars & motos
10m+users per
month

5. The task: A consumer-facing data product
5Predictive Analytics World London 2016 Predictive Analytics Delivered Continuously – A. Wider & C. Deger

6. The task: A consumer-facing data product
6Predictive Analytics World London 2016 Predictive Analytics Delivered Continuously – A. Wider & C. Deger

7. The task: A consumer-facing data product
7Predictive Analytics World London 2016 Predictive Analytics Delivered Continuously – A. Wider & C. Deger

8. The prediction model: Random forest
8
Volkswagen GolfCar listings of
last two years
Predictive Analytics World London 2016 Predictive Analytics Delivered Continuously – A. Wider & C. Deger

9. How to turn an R-based prediction model
into a high-performance web application?
9
?
Predictive Analytics World London 2016 Predictive Analytics Delivered Continuously – A. Wider & C. Deger

10. Traditional approach: Manually translate
model to an efficient implementation
10Predictive Analytics World London 2016 Predictive Analytics Delivered Continuously – A. Wider & C. Deger

11. Traditional approach: Manually translate
model to an efficient implementation again
11Predictive Analytics World London 2016 Predictive Analytics Delivered Continuously – A. Wider & C. Deger
:-/

12. Our approach: Automatically generate
implementation & deliver continuously
12
 Continuous Delivery!
Predictive Analytics World London 2016 Predictive Analytics Delivered Continuously – A. Wider & C. Deger

13. Application code in
one repository per
service.
CI
Deployment package
as artifact.
CD
Deliver package to
servers
Predictive Analytics World London 2016 Predictive Analytics Delivered Continuously – A. Wider & C. Deger
Typical continuous delivery pipeline

14. Continuous delivery pipelines
14
Prediction Model Pipeline
Web Application Pipeline
Predictive Analytics World London 2016 Predictive Analytics Delivered Continuously – A. Wider & C. Deger

15. The price for CD: Extensive model validation
15

16. Lessons learned
16
 Form cross-functional teams of
data scientists & software engineers!
 Set up usage reporting early to improve your
data product in a data-driven way.
 Generating gigabytes of Java code
is a challenge for the JVM
 Use the G1 garbage collector
 Do extensive warm-ups
Turn off Tiered Compilation
Predictive Analytics World London 2016 Predictive Analytics Delivered Continuously – A. Wider & C. Deger

17. Conclusions
17
 Continuous Delivery allows us to bring prediction
model changes live very quickly.
 Only extensive automated end-to-end tests provide
confidence to deploy to production automatically.
 Java code generation allows for very low response
times and excellent scalability for high loads but
requires plenty of memory.
Predictive Analytics World London 2016 Predictive Analytics Delivered Continuously – A. Wider & C. Deger

18. Thank you!
Questions?
– Arif Wider & Christian Deger –