order fulfillment forecasting at john deere: how r facilitates creativity and flexibility
DESCRIPTION
Statistical analysis has been known to be invaluable to any manufactory’s quality assurance for decades. Recently the value of valid statistical analysis has also been demonstrated to radically improve the ability of a company’s ability to weather extreme peaks and valley in customer demand. John Deere has been able to adjust to commodity spikes and housing downturns much better than its competitors have. This is in part due to the implementation of statistical analysis and the use of R software in the order fulfillment function of John Deere.TRANSCRIPT
Valid Statistical Analysis at John Deere and Use of the R Programming Language
Derek Hoffman
Nov-8-2012
A bit about your speaker…
• BS in Statistics andMaterial Science@ Winona State
University• Masters in Statistics
@ Iowa StateUniversity
• 5 Years @ John Deere
Forecasting Group in 2012
• Improvements due to the science of forecasting• Explosion in value and statistician hiring• Increase in problem solving flexibility due to use of R• Huge company saving with dropping flop forecasting software
• Revenue of roughly 35 billion, 8.7% profit
• Has been a Fortune 500 company for the last 56 years, roughly 94th in rank.
• Employs about 50,000 people world wide –roughly 5,000 of them in the Moline headquarters.
Deere & Company – 3 parts
• Agriculture ~70%
• Turf~15%
• Construction~15%
Why does Deere hire forecasters?
• Availability needs to match demand OR you lose market share
• Inventory needs to stay low OR you pay lots in taxes and storage costs
• New factories need to be built at the right size and time OR you made a multi million dollar mistake.
• Work force needs to be hired/cut depending on production plans OR you lose tons training and severance.
My group’s reach at John Deere
CEO, Presidents, Financials
Factory Shifts and
Production
Flexibility of Inventory
Next Month
New Markets, 10 Years Out
Forecasts
My group’s reach at John Deere
CEO, Presidents, Financials
Factory Shifts and
Production
Flexibility of Inventory
Next Month
New Markets, 10 Years Out
Forecasts
Why do statisticians love R?
• Common statistical methods are available as packages (advantage over C++)
• Large support group of users worldwide• Credibility due to submission standards and
university usage.• Often the program of choice during education• Easy to send results to another person (even
if just text files for data and code)
Why does Deere love R?
• The cost is right• Open source – no black box mysteries, no
propriety lock downs• Easy to share across the business• Relatively easy to learn• Often works better or faster than microsoft
products for data and analysis• Infinitely customizable to your problem and
your products – vertical integration
Case Studies at John Deere
• Short Term Demand Forecasting• Crop Forecasting• Long Term Demand Forecasting• Parts Decision Tree (APO)• Order Line Up• Data Coordinator
Short Term Demand Forecasting
Composite Forecast
Estimate Group
Forecast
Factory Forecast
Marketing Forecast
Potential Good:•Multiple view points•Buy-in from all players•Disciplined in forecast creation
Potential Bad:•Group-think•Pressures other than accuracy•Poor information digestion
Bad Forecasting Philosophies
News, Experience
Experience + Feelings on that Day + Outside
pressures
“Forecasts” and directives and
goals
Executive OverrideNews,
Experience, Last YR’s #’s
Math Comparisons, Finical Forecasting,
Experience, Outside forecasts
Forecasts
Gut Feel / Art
History
?
Forecasts (NO estimates of
accuracy, NO interpretation)
Blackbox Forecasts
Forecasting Philosophies
Historical Data(known because is in the
past or current)
Data + Math/Statistics
as calculated by a trained statistician
Forecasts andMEANINGFUL
plus/minus intervals
(flexibility and bad forecast detection)
Statistical ModelsAssumptions
(user generated assumptions about the
future)
Data + Math/Statistics
as calculated by a trained statistician
Forecasts and Analysis of
Forecast Error Contributions by
Assumptions
Assumption ModelsData, Assumptions,
News, ???, Outside Forecasts
Data + Economics + ???
as created by a trained economist
Forecasts, Outside
Forecasts, Current Economic
News
Economic Models
Use of Data-Driven Analysis
Analysis done in my group using R and company data.
Case Studies at John Deere
• Short Term Demand Forecasting• Crop Forecasting• Long Term Demand Forecasting• Parts Decision Tree (APO)• Order Line Up• Data Coordinator
Crop Yields Forecasting
Relative Land Area and Use
Circle = Total Land
Acres in Major World Crops
Circle = Total Crop Land
Crop Yields Forecasting
Crop Yields Forecasting
History
1 Year OUT
2nd Year OUT
3rd Year OUT
The whole time, calculating the valid forecast error and influences.
A large computational task, heavily using programs written in R.
Changes in Crop Splits
Corn Yields
Case Studies at John Deere
• Short Term Demand Forecasting• Crop Forecasting• Long Term Demand Forecasting• Parts Decision Tree (APO)• Order Line Up• Data Coordinator
The Wrong way – Growth f(t)
• The problem really is that we are looking at a correlation with time, not a causation. Also we will always be extrapolating (because the future value of time is outside the our historical data set).
What are Likely Causes?
• Crop Yields• Planted Acres• Crop Prices• Population• Gross Domestic Product• Farm Size• Government• Mechanization Level of Farming• Crop Choices (Corn damages combines faster than
wheat.)
Example of Calculations
The whole time, calculating the valid forecast error and influences.
A large computational task, heavily using programs written in R.
Case Studies at John Deere
• Short Term Demand Forecasting• Crop Forecasting• Long Term Demand Forecasting• Parts Decision Tree (APO)• Order Line Up• Data Coordinator
Parts Forecasting
• Tons of parts, need direction how to best forecast with SAP.
Parts Forecasting – Trilingual?
Case Studies at John Deere
• Short Term Demand Forecasting• Crop Forecasting• Long Term Demand Forecasting• Parts Decision Tree (APO)• Order Line Up• Data Coordinator
Order Scheduling
Order Scheduling
Restraint on Feature A: At most 2 per 4 in a row.
We’re OK!
Order Scheduling
Restraint on Feature A: At most 2 per 4 in a row.
We’re OK!
Order Scheduling
Restraint on Feature B: At most 1 per 3 in a row.
We’re OK!
Order Scheduling
Restraint on Feature A: At most 1 per 3 in a row.
We’re got a problem!
Have to move Matt or Shawn’s tractor to another spot and recheck it all!
Harvester Lineup – Random Guess
Harvester Lineup – Program Results
Order Scheduling – Time
Order Scheduling = $$$
• Old Process– Done manually by
hand– Weekly– Duration: 8 Hours– Not necessarily perfect
• Derek’s Process– Automates the process– Duration: 1.5-2 hours– Human time:15 mins
– Saves about 8 hours per week
– Saves ~$12K per year, per product implementation
Case Studies at John Deere
• Short Term Demand Forecasting• Crop Forecasting• Long Term Demand Forecasting• Parts Decision Tree (APO)• Order Line Up• Data Coordinator
Data Coordinator Uses
DB2
DB2
SQL
Oracle
Multiples Data
sources and Data types
Multiple ODBC
Connections
Single R source Code
DB2
Export Channels
Scheduled Tasks
Batch File execution
A forecast of “Analytics”
• A short history of “cool topics”
• The future of forecasters
• The coming data flood and analytics boom
increase in scalpels ≠ increase in surgeons
The cool word of the year – Dot-com
The cool word of the year - Radiation
The cool word of the year – Big Data
How can we grow responsibly as data scientists and statisticians?
Signs you are in the hype
• Everyone claims it will change the world• It’s taught in business schools• Features on covers of general magazines• TONS of snake-oil salesmen• Legitimate ease in access to the new thing
Cautionary tale:
• Thousands spent on a weather “forecast”
• Ridiculous accuracy measures
• Business users don’t know the short falls till it’s too late
• A need for educated gate keepers to weed bad analysis from good.
• More people are needed to practice forecasting as a profession – or the whole industry will suffer.
• More data, more ease, more computing needed, with greater need for responsible use.
Growing Need of Forecasting Professionals
Statistics and R at John Deere
• John Deere is among the best in large manufactures in implementing good forecasting methods to demand planning
• There are still huge areas to grow – no where near the data usage of companies like Amazon or Wal-Mart
• The challenge is to increase usage and access while maintaining a good internal and external reputation