Most of analytics modeling work today focuses on the production of single-purpose "artisanal" models for predictions. This approach to analytics is fragile with respect to model consistency, reorganization, and resource availability. This talk will argue that instead the focus of analytics modeling should be toward the production of analytics interchangeable parts, which can be combined in creative ways to produce a wide variety of analytics results. This "nuts and bolts" approach allows analytics groups to produce results in an agile way where the time between ask and answer is determined by the right combination of analytics, rather than the modeling.
2. § Method Diversity
§ Limited Penetration
§ More model maintenance
2
Analytics Maturity Stages
DOW RESTRICTED
3. Once group is
successful, demand
can be insatiable
Still have to contend
with model lifecycle
3
Analytics Demand Becomes Overwhelming
DOW RESTRICTED
Project-Based Analytics does not scale
4. § Artisanal manufacturing
•Customer wants a chair
•Artisan builds a chair
§ Division of labor
•Customer wants a chair
•Chair has components which are
Assembled together and ready to go
4
Industrial Revolution
DOW RESTRICTED
5. § Type of model driven by ask
•Churn model asked for
ØData scientist build a model to predict churn
• Prototype approach to Enterprise Analytics
5
Artisanal Analytics Models
DOW RESTRICTED
6. § With experience, senior data
scientists can determine
common predictive elements
between project silos
§ Model leveraging as core
element of analytics strategy
• Backwards—what past problem
looks like this?
• Forwards—model design so that
they can be used over and over
again
6
Industrial Revolution for Analytics
DOW RESTRICTED
7. § Fundamental Models
• Built outside of projects
to be widely leveraged
§ Connector Models
• Simple models built for
need
7
Components for Analytics Assembly
DOW RESTRICTED
8. Difference in Approach for Forecasting
DOW RESTRICTED 8
Transactional Data
Modeled Transactions Modeled Results
Aggregated Data
Modeling
Aggregation
9. § Fundamental models
•Order likelihood
•Transactional demand
•Transactional unit raw material cost
§ Connector models
•Customer demand à production needs
•Customer demand à raw material needs
9
Dow’s Experience—Model-building
DOW RESTRICTED
10. •Combinations of these models drive
–Near-term demand and earnings forecasting
–Revenue optimization
–Customer metrics for churn and sales intervention
10
Dow’s Experience—Predictive Results
DOW RESTRICTED
11. § Components Approach to Analytics Model
•Experience with prior models
•Built to make predictions scale
•Model Leverageability
§ Modeling business at a fundamental scale
§ New role: Systems Engineer for Analytics
11
Requirements for New Approach
DOW RESTRICTED
12. 12
New Measure of Analytics Maturity
DOW RESTRICTED
Prototypes
Models at Scale
Analytics at Scale