Slide 1
Slide 1 text
Alexei Cherenkov
How a boring accounting topic
became the exciting data science / AI
project
Slide 2
Slide 2 text
© Koninklijke Philips N.V.
Preface: Warranty and business
2
• Warranty? Of course, if my product is broken, I can repair it for free!
• Warranty is the obligation of a manufacturer (or vendor) to repair items
sold to a client during limited time since passing of the ownership (it is the
term of sales contract)
• Company must make a financial provision in its ledger to be able fulfilling
the warranty, this is the law in almost all countries
Slide 3
Slide 3 text
© Koninklijke Philips N.V.
Some Philips products must visit repair workshop despite of
lockdown…
―I need to bring you to the
workshop, my little Ironny!
―No, mom, I don’t want to go!...
―We should! Cause it’s still
warranty period. Philips pays. 57
million Euro provisioned, don’t
spoil this budget, let them spend
it exactly!
At the end, each little Iron must be counted…
Slide 4
Slide 4 text
© Koninklijke Philips N.V.
The story begins…
4
•CoNQ (Cost of Non-Quality) for B2C business unit
•~50M / year
•Estimated as linear trend with known expected corrections
•So, what’s the alternative?
•Bottom Up:
–Total Cash-Out forecast to be estimated as the aggregation of
forecasts for each product in each country.
•Kaplan-Meier estimator
–Wallace R. Blischke, M. Rezaul Karim, D. N. Prabhakar Murthy.
Warranty Data Collection and Analysis (Springer-Verlag London
Limited 2011)
Slide 5
Slide 5 text
© Koninklijke Philips N.V.
Kaplan and Meier present the estimator
• The black curve (Fig. above) shows the share of products surviving during
a time (horizontal axis, weeks) passed after purchase. Line is ended at
106th week (length of warranty). In this example, approximately 96% of
sold items survive over warranty period.
• Red curve is Weibull distribution function extrapolation over empirical
stepwise distribution.
5
All products of a category and
country look different
Slide 6
Slide 6 text
© Koninklijke Philips N.V.
Is our data ready for Kaplan-Meier?
6
• How the claim process work?
• Sell-in vs. Sell-out
• Some products are not included in data
–Spare parts
–Bundles
There is always some space for data science…
Slide 7
Slide 7 text
© Koninklijke Philips N.V.
Predict
Train
14-Sep-22
Flying in a helicopter
7
Sell-In
Projected
claims
Sell-In forecast
Sell-Out
forecast per
product,
2014-now
Cash Out
forecast
Sell-Out
(GfK)
Transfor
-mation
rule
K-M
Survival
model
Claims
Sell-In Actual
Avg
Cash
Out per
product
QI
Bundles
Country aggregation rules
Sell-Out 2014-2023
Slide 8
Slide 8 text
© Koninklijke Philips N.V.
Data is the new oil, but oil quality differs a lot…
• Actual Sell In from 5 different databases (legacy+current+very new)
• Sales plan (till end of the year, requires artificial 24 prolongation based
on strategic plan)
• Warranty Claims (Invoices) – World, China, USA. Note: 3 warranty
business models
• Product hierarchy (master file)
• Bundle hierarchy (master file)
• Quality improvement schedule
• All databases have different country granularity (grouping)
8
Huge volume
of data Not enough
data
Slide 9
Slide 9 text
© Koninklijke Philips N.V.
Better than perfect
• Poor products are passing improvements at factory
• This drives to the appearance on the market of product
modifications / generations
14-Sep-22 2019, Philips confidential
9
Slide 10
Slide 10 text
© Koninklijke Philips N.V.
Time granularity
• Sell-in and sell-out is monthly data
• GCS delivers exact claim date, with strong weekly pattern
• One of the requirements: early forecast of bad behavior of NPIs, so
the model must be built on weekly level
• COVID-19 impact
10