Edison: Predictive Analytics for Sales Order Fulfillment
Hasso Plattner Institute
Enterprise Platform and Integration Concepts
Research Group of Prof. Dr. h.c. Hasso Plattner
Sales Order Fulfillment ProcessIssues Interrupting the Process
Chart 2
During sales order processing several issues may occur:
An issue usually halts the process and causes the delivery to be delayed.
Incomplete data
Delivery blocks
Credit blocks
Billing blocks
Delivery issues
Unconfirmed quantities
Shipping blocks
Incomplete data
Credit blocks
Shipping issues
Accounting issues
Invoicing issues
Purchasing issues
Manufact. issues
In Order Entry In Supply In Delivery In Invoice Process Step
Possible Issues
Chart 3Solving issues after they have occurred is often too late
Sales Order Fulfillment ProcessStatus Quo
Number of current issues by process step
SAP’s Sales Order Fulfillment Monitor (SOFM):
■ SOFM allows for detection and categorization of current issues
■ Employees then try to resolve current issues as fast as possible
■ Learn which fields and conditions caused issues by looking at
data of millions of historical sales orders
■ Predict the expected delay whenever a new sales order
enters the system
■ Display a list of the most pressing expected issues
■ Propose the most effective action to mitigate an issue in
advance
■ Resolving issues even before occurring saves time and
money and makes customers happyChart 4
Sales Order Fulfillment ProcessOur Vision
Predictive Analytics for Order FulfillmentPrototype
Chart 5
Most effective action to mitigate issues in advance
Process overview with expected delays by step
Expected issues of the currently selected process step and the data used to calculate risk percentages
Predictive Analytics for Order FulfillmentRisk Indicators for Expected Issues
Chart 6
Indicating how many similar sales orders experienced the issue in the past
Sales order fields increasing the risk of an issue
Sales order fields decreasing the risk of an issue
Decision Trees■ Binary classification, e.g.,
delivery block (SO02) or no delivery block
■ New orders are classified based
on historical data
■ Find splits which lead to
homogenous groups
Machine Learning TechniquesPredicting Issues based on Sales Order Information
Chart 7
...
Machine Learning TechniquesPropose Most Effective Action
Chart 8
SØ
S4
S3
S2
S1
S5
Partially Observable Markov Decision Process (POMDP)
States: Issues or issue combinations, e.g.,
SØ: no issueS1: {delivery block, billing block}S2: {shipping block}...
Machine Learning TechniquesPropose Most Effective Action
Chart 9
SØ
S4
S3
S2
S1
S5
13%
67%
5%
3%
2%10% States:
Issues or issue combinations, e.g.,
S1: {delivery block, billing block}...
Initial belief state: Decision trees for every state determine contribution to belief state
Partially Observable Markov Decision Process (POMDP)
Machine Learning TechniquesPropose Most Effective Action
Chart 10
SØ
S4
S3
S2
S1
S5
A1
Actions and transitional probabilities:■ Actions change a sales order's state
with a certain probability (and can affect multiple states)
■ Performing an action influences the belief state
A1
A2
A1
A2
A2
A2
A1
A1
Partially Observable Markov Decision Process (POMDP)
Machine Learning TechniquesPropose Most Effective Action
Chart 11
SØ
S4
S3
S2
S1
S5
Rewards:■ Actions have costs and benefits
■ The reward for performing an action in a state helps to find the best action
■ Costs could be delays caused by issues, contractual penalties, attorney’s fees, ...
A2:r2
A2:rØ
A2:r3
A1:rØ
A2:r3’
A1:r1
A1:r2
A1:r3
A1:rØ’
Partially Observable Markov Decision Process (POMDP)
Machine Learning TechniquesPropose Most Effective Action
Chart 12
SØ
S4
S3
S2
S1
S5
POMDP:■ Finds best long term action (not
necessarily obvious action)
■ Considers an action's consequences on all states
■ It is therefore superior to decision rules (e.g. mitigate current most severe issue)
A2:r2
A2:rØ
A2:r3
A1:rØ
A2:r3’
A1:r1
A1:rØ’
A1:r2
A1:r3
13%
3%
2%
Ai:rØ
...
Partially Observable Markov Decision Process (POMDP)
Predictive Analytics for Order FulfillmentAction Proposal to Avoid Issues
Chart 13
Predictive Analytics for Order FulfillmentAction Proposal to Avoid Issues
Chart 14
Enterprise Platform and Integration Concepts
Hasso-Plattner-Institut für Softwaresystemtechnik GmbHAugust-Bebel-Str. 8814482 Potsdam, Germany
Contacts:
Johannes Huegle, [email protected] Jan Kossmann, [email protected]. Matthias Uflacker, [email protected]