Marketing return On Investment Modelinig
- 1. Developing a ROMI Analysis Introduction and Discussion
- 2. Agenda Overall Approach GMAX ™ Modeling Benefits Case examples Discussion Data availability
- 3. Methodology Overview Build database Normalize data from various sources Build parallel models to look at different variables and variable combinations Refine models (focus on DMAs with enough advertising data to make confident conclusions) Generate and test hypotheses with models Find themes that emerge from the models Translate mathematical results into actionable business recommendations Drill down to gain better understanding of relationships
- 4. Why is the OG ROMI approach different? Proprietary tools and data platforms permit examination of more variables Typically 3 orders of magnitude more Identifies complex relationships and patterns in the data Interactions Curvilinear functions (multi-order polynomials) Models the real-world environment “ X in combination with Y”
- 5. Methodology The approach/methodology used a combination of analytical techniques GMAX – Genetic Approach - 10s of thousands of model combinations - Determines important variables Regression - Seeks to understand and calibrate individual variable influence
- 6. Classic Regression $10 $9 $5 $4 $7 $8 $3 $2 $1 $6 R R R N R N N N R
- 7. Genetic Programming $10 $9 $5 $4 $7 $8 $3 $2 $1 $6 R R R N R N N N N R
- 8. GMAX ™ Benefits Cost-effective examination of more variables Including non-linear relationships … and interactions between variables Helps avoid “errors of omission” (filtering potentially useful data based on heuristics or prior experience of the modeler)
- 9. Case example HP: Small Business Target
- 10. Project Objective Develop model and understanding of relationships between marketing expenditures and sales Total Sales $ Client Controlled Attitudinal Outcomes Sales Outcomes
- 11. Print Costs While print costs appear in the GMAX model, the relationship is not clearly seen in graphical analysis of print costs by themselves
- 12. Marketing Communications Variable Tree Share of voice, print, online, and direct mail all have an affect on sales Note how Print has an impact by itself AND in combination with Direct Mail Sales Shipments Prod B Share of voice Prod A Share of voice Print Out of pocket Direct Mail Print Out of pocket Online costs
- 13. Catalog Circulation Higher volumes of catalogs correspond with higher sales volumes Diminishing returns at approximately 160,000 pieces Suggests that higher cost catalogs (i.e. CPM) produce results
- 14. Other variables examined: Supply Chain An example from one set of modeling … Variable Relative Frequency AvFqcy Variable 1.83000 PRTTOOP Print out of pocket costs 1.07000 WTPcompAP Wtd average pricing vs Comp A 0.75000 Category SOV Share of voice 0.53500 BRANDSOV Share of voice 0.31000 CATTOOP Catalog out of pocket costs 0.15500 PRTCIRC Print circulation 0.08500 SOVSOM Share of voice / Share of market ratio 0.06500 WTD compaPR Wtd average proposed pricing vs. CompA 0.06500 SERSPEND Total spending on XXXX 0.04500 BACKDOL $ value of backlog 0.04500 CATQTY Catalog circulation (quantity) 0.03000 SPCompBPR Proposed price vs CompB 0.02500 DIRMAIL Direct mail circulation 0.01000 SPRICE Price vs CompS 0.00500 DMCOST Direct mail total out of pocket costs 0.00500 BACKQTY Backlog quantity (units)
- 15. ROMI Model This analysis yields a moderately complicated, but understandable and interpretable model Total Sales $ = 240,000,000 + 1.984 * ((Print Out of Pocket $/Print Circulation) * (Catalog Out of Pocket $) * (Weighted Pricing vs. compA/Weighted Proposed Pricing vs compA)) + $335 * (Direct Mail Cost)
- 16. Sales per $ For most marketing expenditures it was possible to calculate an estimated of return (SMB sales) per dollar spent
- 17. Case example U.S. Navy: Recruiting
- 18. Lead Contract Lag Although the lag between a lead and a contract varies, almost 60% of contracts are signed within 4 months of generating a lead Cume 58.2%
- 19. Positive Impact on Leads Total Spend General Mkt Spend TV Spend Radio Spend Internet Spend Internet CPL Each of these variables contribute to lead generation at > 85% CL
- 20. Influence on Leads per $ Spend Of the variables that have a positive correlation with leads, Internet (specifically CPL), Direct Marketing, Radio and TV have the highest rate of return per an additional $100K of spending
- 21. Positive Impact on Contracts Total Spend TV Spend Media Events Direct Marketing Internet CPL Each of these variables contribute to contracts at > 85% CL
- 22. Contracts Total media spend Print Media Events War Handling Media Events Media Events work by themselves, but also act as catalysts to Print and Direct Marketing
- 23. Contracts: Influence per $ Spend Of the variables that have positive correlation with number of contracts, Internet (specifically CPL and Search), Direct Marketing and Media Events have the highest rate of return per additional $100K spent
- 24. Implication: Media Influences LEADS CONTRACTS Overall Spending General Market Spend Radio TV Internet CPL Direct Marketing Media Events
- 25. Discussion Data requirements – As many variables as you are able to identify OG supplements with secondary data sources and GIS database information Time periods Ideally 2-3 years of back-data depending on category, and if it is monthly/weekly Data format It doesn’t matter …we do the Extraction, Transformation and data Loading (ETL) work
- 26. In Summary… Marketing Analytics … not just “market research” Primary data + secondary data Integrate and synthesize data from throughout the company – not just research data but also sales, marketing plan investments, etc. Using proven tools and templates… We frame data and convert to actionable information
- 27. Intersecting marketing, science and technology™
Editor's Notes
- #2: 1
- #6: The first step in the model building process is collecting all of the various data inputs. At HP, we have a ton of data – but it’s not always easy to get your arms around it. We spent several weeks identifying the data that we wanted to include in the modeling process, finding the owner of the data, and getting data feeds to OG so that they could work it into a form that would be useful in a large-scale modeling process. Once the data was gathered, a data mining tool called GMAX was used to identify the important variables and uncover some unique relationships between the variables … these “driver variables” were then exported to more traditional stat analysis packages to refine and calibrate the end models.
- #7: A quick explanation of how GMAX works – Most statistical analysis utilizes regression as the principal way to identify positive (above the line) and negative (below the line) data variables. This is generally displayed as a sloped line through a data plot.
- #8: The Optimization Group utilizes a data mining and modeling process called “Genetic Programming” (GP). This is software that literally programs itself, based on the data it is testing. OG’s proprietary GP software product is called “GMAX”.