D1 Solutions AG a Netcetera Company Cost Reduction in Bill-Insert Campaigns With Predictive Analytics Stamatis Stefanakos Predictive Analytics World, October 20-21, 2009, Washington DC
2 Outline Who we are Sunrise Communications: predictive analytics & CRM The challenge The solution (and an alternative) Summary & lessons learned
3 D1 Solutions in the Netcetera Group Netcetera Group Locations: Zurich Bern Vaduz Skopje Dubai Founded: 1996 Employees: 200+ Netcetera Metaversum D1 Solutions Brain-Group System Integration & Software Development CRM Solutions Business Intelligence Financial Advisory Solutions
4 Expertise and Customers Core competencies Customers of D1 Solutions (Selection) Expertise Data Warehousing Reporting/MIS Predictive Analytics Requirements Management Project Methodology Technology
5 Outline Who we are Sunrise Communications AG: Predictive Analytics & CRM The challenge The solution (and an alternative) Summary & lessons learned
6 Sunrise Communications The company: Number 2 telecom operator in Switzerland Mobile & wireline 1.7M mobile customers 730K landline customers 360K internet customers
7 The data landscape in a typical telecom operator ODS Billing Network External Sources DWH Reporting Data Mining CRM Campaigns
8 Predictive analytics at Sunrise Data analysis Customer profiling Campaign analysis Modeling Prepaid churn Postpaid churn Landline churn Payment risk
9 CRM activities Focus Retention Cross-selling Up-selling Channels Mailing campaigns Bill inserts Campaigns are sent to the customers together with the monthly invoices Opt-out from printed invoices is possible SMS campaigns Email campaigns Call campaigns
10 Campaign selection Socio-demographics Customer Base Campaign selection NAB Segment Churn risk Call behavior Language
11 Outline Who we are Sunrise Communications AG: Predictive Analytics & CRM The challenge The solution (and an alternative) Summary & lessons learned
12 Monthly billing August September October Telephone usage Invoice sent
13 Bill inserts August September October Telephone usage Invoice for September is sent along with some offer Invoice sent
14 Threshold-based billing August September October Telephone usage ß Previous month s usage is low and no invoice is sent Invoice for August & September is sent Invoice sent
15 Threshold-based billing & bill inserts August September October Telephone usage ß Previous month s usage is low and no invoice is sent Invoice for August & September is sent along with some offer Invoice sent
16 The problem with bill inserts & threshold-based billing August September October?? Telephone usage Selection of customers & printing of bill inserts has to be done in the beginning of September The bill insert is sent together with the invoice Invoice & bill insert preparation
17 An example September October CRM creates the campaign selection: 861K customers are eligible for the campaign. At this moment the September revenues are not known? Telephone usage 861K Bill inserts printed 250K bill inserts have to be thrown away. This costs ~4rp / insert = 10 000CHF. 613K Bill inserts sent
18 The business problem Volume Estimation for Bill-Insert Campaigns At the time the bill-insert campaign preparation begins, the volume of the campaign can not be calculated accurately. This is causing unnecessary printing costs. CRM needs a method to estimate campaign volumes.
19 Outline Who we are Sunrise Communications AG: Predictive Analytics & CRM The challenge The solution (and an alternative) Summary & lessons learned
A potential solution 20
21 Campaign selection Socio-demographics Customer Base Campaign selection NAB Segment Churn risk Call behavior Language
22 A potential solution (cont d) The rate with which invoices are sent depends on the customers that are selected for a particular campaign. For such an approach to work we would need a different model for each campaign. This is not feasible: o Too many campaigns. o Some campaigns are done just once and no historical data is available. Campaign A Campaign B
23 Our solution using predictive analytics Design considerations Deliverable A predictive model generating each month a prediction for each customer whether next month s invoice will be sent or not. Evaluation Costs The model should over-predict the amount of invoices that will be actually sent by a small safety margin. (Punish false negatives.) Development time should be kept to a minimum to justify the business case. Deployment Complex data dependencies should be avoided to ensure on-time availability of the model.
24 The predictive model A predictive model generates a prediction each month for each customer. The prediction is whether the invoice will be sent or not. SQL SEGMENTATION_ DM.BIL0.. C 5.0 Type TARGET Table The prediction is made using historical revenue & invoice data only. Filter SQL dwh_pdv1_ nabappl.bil.. We use a C5.0 decision tree algorithm. The model is implemented in SPSS Clementine. output.txt
25 An example September October CRM creates the campaign selection: 861K customers are eligible for the campaign. With predictive analytics we estimate that only 653K customers will get an invoice next month. 653K Bill inserts printed Only 613K invoices are sent. 40K bill inserts have to be thrown away. This costs ~4rp / insert = 1 600CHF. 613K Bill inserts sent
26 Details of the model The model decides whether a customer will receive an invoice or not o o based on the revenues of the last months based on if an invoice was sent the last months. No seasonality is built into the model. No other data is used. The estimates of the model are used as-is. However, a larger safety margin in the prediction can be introduced if needed by the campaign manager.
27 Deployment of the model Revenue & invoice data Scoring with predictive model Campaign selection tool Scores in DWH Selection in DWH Campaign Volume Estimation
28 Data availability Month N-10 Month N-3 Month N-2 Month N-1 Month N Scoring will be done beginning of month N-1 Bill insert to be sent end of month N Revenue and invoice data used Revenue and invoice data used in the model in the model For the timely scoring of the customers, delivery of revenue & invoice data for month N-2 was critical. In our test runs the prediction could not be done due to DWH loading problems. The performance of the model met our goals even if we did not use data from month N-2.
29 Evaluation of the model Review of our design principles Deliverable Costs A predictive model generating each month a prediction for each customer whether next month s invoice will be sent or not. Evaluation The model should over-predict the amount of invoices that will be actually sent by a small safety margin. (Punish false negatives.) The model enables us to reduce the costs of campaigns by 25% to 30%. Development time should be kept to a minimum to justify the business case. Deployment Complex data dependencies should be avoided to ensure on-time availability of the model.
30 Outline Who we are Sunrise Communications AG: Predictive Analytics & CRM The challenge The solution (and an alternative) Summary & lessons learned
31 Summary & lessons learned Summary Predictive model to estimate the volumes of bill-insert mailing campaigns. New application of predictive analytics (PA) within an organization with an established PA infrastructure. Why we like this project Compact case with easily quantifiable results exhibiting the power and flexibility of PA. Ice-breaker within organization for further use of PA also outside the typical telecom data mining projects. Versatile use of PA as a tool/framework. Lessons learned Even with small investments PA delivers tangible results. Value of fast prototyping. Implementation & deployment of predictive analytics can be fast and with minimum development & maintenance costs justifying its use even in smaller business problems.
D1 Solutions AG a Netcetera Company Contact Dr. Stamatis Stefanakos stamatis.stefanakos@d1solutions.ch phone: +41 44 435 10 03 mobile: + 41 78 634 09 24 D1 Solutions AG Zypressenstrasse 71 CH-8040 Zürich