Predictive modelling is the name given to a family of statistical model types whose goal is to predict the likely, expected or future value of an important marketing-related variable such as response rates, product purchase rates, subscription rates, donation rates, attrition rates, default rates, and many others like these. The key application is that once the attributes of the buyers (etc.) are modeled then marketers are able to identify those that are most likely to buy in the next campaign - whether this involves using an existing customer database, on the ground (at their residences) or on mailing lists. There is in this field of modelling an important distinction between those types of models that leverage existing customer data – typically from databases - to build the models (e.g. loyalty models, cross-sell and up-sell models) and those that are designed for new customer acquisition. The latter cannot typically assume the existence of variables relating to the behaviour of existing customers to use in targeting marketing campaigns. At the low end of this class of models are what are known as “propensity models” which typically use aggregate attributes in the predictive process.

One of the most basic yet powerful simple projects that EA can do for you is to add hundreds of variables to your databases to be used in your modelling efforts. This is done by using special geodemographically-driven “data overlays” and related techniques for assigning both behavioural and attitudinal variables to persons and households.

The quantitative researchers at EA have a vast experience in building many different types of predictive models for many types of clients over a 25 year period. We would be delighted to sit down with you and discuss your specific modelling needs. There is no such thing as a standard approach to modelling these types of variables. Every firm's needs are different. EA analysts are familiar with all available data sets in Canada and the U.S. that can be used in predictive modelling and have worked with hundreds of proprietary client databases. We have many “tricks of the trade” in leveraging data and integrating data, using data overlays, building supplementary explanatory variables, matching datasets for practical modelling purposes.

The modelling methods that we use most in predictive modelling are: multiple regression, principal components, discriminate function models, logit models, multinomial logit models, CHAID, CART, “supervised” neural nets of several types (multi layer perceptrons and radial basis function models). We tend to use neural models rarely because of their “black box” nature. In addition, we have developed a number of mixed multivariate techniques, and algorithms for these, which we use when these are required.

Statistical models are used to estimate total business volumes by product or category for small areas. These models are customized and complex and utilize customer data.

Untapped Potential = Predicted $ - Actual $

For best results, untapped potential is calculated at small areas and then assigned to or summarized for locations, regions, sales territories, trade areas, etc

How it is used:

• Set realistic sales objectives
• Allocate marketing dollars strategically
• Customize personnel by market potential
• Align channels of distribution
• Identify new target groups/areas

Retail site evaluation models are statistical/mathematical models which are designed to estimate the business volumes that are expected to be attracted to a retail store at particular site – both existing sites and potential new ones. These models are also called sales projection models or just “site models”. A typical finding would be an estimation that a store of a certain format, size and other attributes, after being open for two years, should attract for example, $4.9 million of sales annually, or between $4.6 million and $5.2 million with a certain confidence.

There are many types of models from cost effective market screening to high-end market optimization. The main model types include:

• Analogue models
  • Generate sales predictions by identifying best matches to existing store network
  • Evaluate entire store network in seconds
  • Intuitive and easily understandable results
  • Delivery in easy-to-use software or as sets of reports
• Site Screening models
  • Discover areas where high business volumes can exist
  • Rapidly evaluate thousands of sites/locations
  • Consider local trade area statistics, competition and accessibility
  • Outcome represented as a surface map showcasing highest potential areas
• Regression-based models
  • Sophisticated method for predicting store sales
  • Assess by store type, context, regionality or any combination
  • Take into account hundreds of variables to produce sales projections
  • Delivery in easy-to-use software or as sets of reports
• Spatial Interaction Models
  • Elaborate modelling approach producing accurate predictions of sales or transactions
  • Deal thoroughly with cannibalization and competition
  • Market or region specific calibration and refinement
  • Delivery in easy-to-use software or as sets of reports
• Network Optimization
  • Estimate the best set of locations in a market to maximize sales or market share
  • Spatial Interaction Models are calibrated and used within network optimization models
  • Used for opening, closing and relocating locations to determine network-wide impacts
  • Delivery in easy-to-use software

The better and more reliable models are based on a lot of good data and rigorous statistical analysis. The most complete types of models, including those using multiple regression methods, use predictive variables from these 5 headings:

1. store area socio-economics and demographics - to capture the amount of demand
2. store and site attributes (e.g. hours open and square feet)
3. situational attributes (e.g. nearby traffic generators)
4. effective competition (e.g. amount of competition within a certain distance)
5. in-store management quality type variables

Retail site evaluation methods also apply to banking and other financial institutions, real estate offices and a wide range of purveyors of consumer services.

We offer the widest possible range of services relating to site models and deliver them in a variety of ways, from excel spreadsheets and data reports to software systems including web-based and those with intuitive maps.

The most sophisticated site evaluation models are spatial interaction models (called SIMs, a sophisticated modern version of the old “gravity models”) which are designed to predict the flows of dollars, transactions and customers from small demand area origins to store or bank branch destinations. The store system is modeled as a network.

Examples of Questions Addressed:

• What will the sales be at a new branch of format “Type 2” opened at a particular location?
• What is the impact on sales in a market if we open 4 new branches and close 2?
• What happens if my competitor opens a new super store across the street?
• Will a particular new store opening cannibalize our existing sales?
• Which branch formats should we use?

Spatial Interaction Models allocate potential from each neighbourhood (DA or BG) to each store or branch in a network based on:

• Demand for banking product(s) in the neighbourhood
• Branch attractiveness
• Drive-time or distance between neighbourhood and branch
• Location, number and strength of competitors
• Number of local detractors and attractors
• Socio-economic behaviours

These models can take into account banking close to home and close to work.

Network optimization models are a special type of site model. Whereas regular site models are used to estimate the expected sales at a single location - or one at a time - to find a good site or best site, network optimization models are designed to estimate the best set of multiple locations in a network to maximize some criterion such as system wide sales or market share. Good network optimization models are really Operations Research models using mathematical programming. Historically they have been called “location-allocation models”. However, most location-allocation models were not designed for retailing in which customers patronize stores that are not just the closest ones. Realistic network optimization models need regular site evaluation models as part of the tool set in order to estimate the business volumes at each location in the network - considering all other retailers as competitors. The best network optimization models must use the best site evaluation models inside them. This means that spatial interaction models must be calibrated and used inside network optimization models.

Examples of Questions Addressed:

• If I want to open 10 new branches in an urban market, where should they be?
• Which are the best 4 branches to close or consolidate?
• Which 15 are the best to keep open with minimal loss of market share?
• What is the optimal number of branches in this market?

Outputs from these models often include projections – such as business volumes by product mix in five years based on the use of a set of assumptions in what is called a “scenario”.

EA has many algorithms and programs to solve retail and banking network optimization problems. We can also use various components of these to construct new customized solutions for clients with special requirements.