Analytics White Paper

Choice Modeling for Sales Forecasting

The term “Choice Modeling” refers to a family of tradeoff and conjoint techniques. The goal of choice modeling is optimization of one or more variables, given a limited number of options and constraints. Choice modeling is a proven and reliable sales forecasting technique.

The adoption of choice modeling took wings in the 1980s and reached even greater heights after 2000 when Daniel McFadden, Ph.D., an American professor, was awarded the Nobel Prize in Economics for his pioneering work in the application of choice modeling to economic decision making. Since then, choice modeling's growing popularity has also been fueled by software advances, more powerful computers, and improvements in 2D and 3D animation technology.

Choice modeling provides a powerful set of equations and tools to help improve sales forecasting for existing brands and for new products. Not only does choice modeling improve the accuracy of sales forecasting in most instances, choice modeling also helps reveal the interrelationships among key marketing variables so that brand managers know which marketing levers to pull to achieve success in the marketplace. It’s this diagnostic and prescriptive power that makes choice modeling so valuable.

Choice Modeling

Consumers make brand “choices” when they go to the grocery store, the drugstore, or the car dealership. Making brand choices is easy for consumers; it’s something they do almost every day. Choice modeling techniques make it possible to imitate this shopping and decision-making process, with all of the important variables carefully controlled by a rigorous experimental design, so that a brand’s market share and sales revenue can be accurately predicted, given a set of marketing variables and values.

An example might make this easier to understand. Let’s suppose that a new brand of peanut butter is about ready to enter the market, and the manufacturer wishes to optimize the new product’s chances of success. The product formulation is ready, but there are unresolved marketing issues. Under consideration are four package designs, three pricing levels, six vitamin additives, and five nutritional claims for the package. It's immediately obvious that the unresolved issues add up to 360 unique possibilities (4 x 3 x 6 x 5 = 360). This is where choice modeling comes to the rescue. By choosing a subset of all these possibilities following an experimental design, choice modeling permits the relative marketing outcomes for all 360 possible combinations of variables to be accurately estimated.

Here’s how it works. Once the experimental design is created, a sample of target consumers is surveyed (sample sizes typically range from 500 to 1,000 respondents). Each survey participant sees and responds to a number of shopping “scenarios.” Think of a scenario as one shopping trip. Each participant sees a typical retail shelf display, with all of the major peanut butter brands shown. The survey participants (typical consumers of brands in the category) are asked to look at a shelf display and indicate how many jars of each peanut butter brand they are likely to buy in the next 30 days. That completes one scenario. Then, the online “shelf set” systematically changes—package designs change, prices change, nutritional claims change, and vitamin additives change. The respondent is then asked to shop the peanut butter category again and choose what she would buy in the next 30 days, given the new shelf set and new variables. This completes the second scenario, and the process continues. Each respondent typically completes 6 to 10 scenarios (i.e., shopping trips), and in each scenario the marketing variables change following the experimental design.

Since all of the marketing variables are carefully manipulated following this experimental design, the predicted market shares (and sales forecasts) for the new peanut butter can be calculated for all 360 combinations of marketing variables. The equations that predict sales of the new peanut butter are then used to build a forecasting simulator, so that a research analyst or brand manager can play "what if" games by changing one or more marketing variables, to see the effects on market share and the sales forecast.

Existing Brands Sales Forecasting

While the discussion thus far has focused on an imaginary new product, choice modeling is equally important as an optimizing and forecasting technique for existing brands. Brand managers have many hypotheses about marketing variables and marketing actions that might improve their brands’ market shares, but the possibilities are too many (often hundreds or even thousands of combinations) to test by traditional research methods. The most important of these hypothesized marketing variables and actions can be incorporated into one or more choice modeling experiments.

The results of the choice modeling experiment (and the equations thusly derived) are used to build a predictive simulator. The simulator is similar to a video game, in that the brand manager can change variables and the values assigned to variables to see the net effects on her brand’s market share and competitors’ market shares. Market shares are easily translated into dollar sales forecasts, when combined with total category sales volume. The simulator gives the brand manager a powerful tool to optimize his brand’s marketing, and a powerful tool to guide reactions to competitors’ actions.

New Product Sales Forecasting

Now, back to the imaginary new product (a new brand of peanut butter). How might choice modeling be used to forecast its sales? The design of the choice modeling experiment would involve identifying the most important brands and the most important variables in the category and the range of those variables (based on good qualitative research, if possible). The choice modeling experiment itself would involve surveying hundreds of category users (i.e., peanut butter users). Each survey participant would see a number of “scenarios” and choose the brand or brands of peanut butter to buy. The results of this large experiment would be translated into predictive equations and those equations would be fused together into a predictive simulator. Now, the simulator is ready for optimizing the new product’s marketing variables to improve its chances of success, but the simulator is not yet ready for forecasting actual in-market sales.

At this point, the choice model’s predictions assume 100% awareness of all brands, 100% distribution levels for all brands, and product parity (i.e., that the new product is as good as competitive brands in terms of quality and performance relative to price). Also, the predictive simulator’s sales forecast is not the traditional trial-repeat purchase forecast for the first 12-months of the new product’s introduction (starting from the point a minimum distribution level is achieved). Rather, the forecast is a “going rate” forecast after the introductory period (i.e., one or two years after market introduction). Typically, the simulator forecasts “retail depletions” at retail prices (i.e., inventories in the factory, in transit, and in warehouses are not included in the sales forecast). To translate the “retail prices” sales forecast into a manufacturer’s sales forecast requires discounting the retail sales forecast by retailers’ average markup.

The predictive simulator must be adjusted for awareness of each brand in the model and the distribution level for each brand (retail stores plus online stores). If market share history, brand awareness trends, and distribution level trends are available, a best practice is to calibrate the predictive simulator to the actual historical data to improve accuracy.

Product parity was mentioned as an assumption in the predictive simulator. Since one of the most important variables in forecasting is the quality or performance of the new product itself, another best practice is to submit the new product to in-home usage testing to make sure the new product is as good (and ideally better) than competitive products in consumer acceptance and preference. The importance of objective, third-party consumer testing of the product itself cannot be overstated.

Consumable Goods Versus Durable Goods

The hypothetical new peanut butter falls into the category of consumable goods, and choice modeling is an ideal way to optimize the marketing of these types of consumer products. Durable goods (i.e., products lasting 3 or more years, such as lawnmowers, cars, appliances, etc.) are equally amenable to choice modeling. Typically, a product clinic would be conducted first, to familiarize target consumers with the test product and major competitive products. For example, several hundred consumers might be recruited to visit a mock showroom to evaluate a number of new cars. After examining and evaluating the new cars, a choice modeling experiment could be the last step. Durable goods are often more complex than consumable goods, so the car clinic in this example educates and familiarizes consumers with the features of the new cars, so that consumers know enough to participate in choice modeling exercises.

Choice modeling is an ideal technique for optimizing and forecasting sales of durable goods. If you think about all of the variables linked to a new car (electronics, exterior design, mechanical features, interior design, etc.), the number of possible combinations is easily in the thousands. A best practice is to limit the number of variables to those deemed most important, so that the choice modeling process is not overwhelmed by complexity. Two or three choice modeling experiments might be necessary to cover all of the important variables. Otherwise, the application of choice modeling to durable goods is very similar to its application to consumable goods.

Reduced Risk

The marketing of an existing product, or the introduction of a new product, are inherently risky. No optimization method or forecasting technique can guarantee success 100% of the time. Choice modeling, however, can greatly reduce the risks associated with marketing and new product development. The simulator-based knowledge and flexibility allows brand managers to confidently optimize marketing decisions and quickly adapt to changes in market conditions and competitive actions.

Jerry W. Thomas

Chief Executive Officer

Jerry founded Decision Analyst in September 1978. The firm has grown over the years and is now one of the largest privately held, employee-owned research agencies in North America. The firm prides itself on mastery of advanced analytics, predictive modeling, and choice modeling to optimize marketing decisions, and is deeply involved in the development of leading-edge analytic software. Jerry plays a key role in the development of Decision Analyst’s proprietary research services and related mathematical models.

Jerry graduated from the University of Texas at Arlington, earned his MBA at the University of Texas at Austin, and studied graduate economics at SMU.

• 
•