Consumer Electronics Industry Takes A Conjoint Approach To Forecasting

Jonathan Weiner

This paper presents one method of estimating unit demand for a new product, in a relatively new product category, using conjoint modeling based on existing unit demand data for competing products. In addition, this paper provides actual unit sales data for the new product as a measure of the accuracy of the tool.

Introduction

Conjoint analysis has traditionally been used for determining the ideal product configuration for new and existing products. Additionally, conjoint analysis has been used as a basis for estimating demand.

Unit demand can be estimated using a variety of techniques. These include such techniques as the Delphi method; and trend analysis and moving averages, based on past sales data with seasonal adjustments or comparisons with other products in a similar category.

This paper outlines an approach that combines conjoining-developed new product configurations and pricing strategies with the existing unit sales data of competing products.

Background

A new consumer electronics category was nearly one year old, with only two major products competing for market share.

Brand A was a basic product produced by a company with a great deal of money to spend on advertising, promotion and distribution. Their product was priced at $89. Brand B was an advanced product produced by a company with few dollars to spend on advertising, promotion and distribution. Their product was priced at $179.

At the time of the study, Brand C was about to introduce their new product. Brand C commissioned the followed study:

• To determine the appropriate product configuration and price for
• To develop an estimate of first-year unit demand

Design

A full profile conjoint study was designed to test the impact of three product features. These were brand, price and a peripheral handling feature.

• The brands tested were:
• Brand C
• Brand A
• Brand B
• Brand D
• The prices tested were:
• $69
• $119
• $159
• $199
• The bundling feature tested was:
• One peripheral included
• No peripheral included

Using an orthogonal fractional-factorial design, a total of 16 product combinations were created for evaluation by respondents.

Sidebar Discussion on Factorial Designs

A simple design is the full factorial design, which consists of all possible combination of the levels for each factor. For this study, the full factorial design would consist of 32 product combinations (4 brands x 4 prices x 2 bundling features). In a full factorial design, all main effects (brand, price and peripheral bundling) and their interactions are estimable and are uncorrelated (independent from one another). However, 32 cards would have been too many for a respondent to consistently evaluate.

For this reason, an orthogonal fractional-factorial design was selected. When a model is fit with an orthogonal design, each utility estimate is independent from all of the other estimates. The result is the ability to estimate main effects with a smaller design, but with the accuracy of a full factorial design. In this study, one-half of the number of cards could be used to provide uncorrelated estimates of brand, price and peripheral bundling.

Data Collection

Prior to evaluating the 16 product combinations, sample products were shown so that respondents were able to grasp the look and feel of each brand's products. Visual portrayal of the products in any study is very desirable, but in this study it was especially important. In this category, it was particularly important because both the Brand B and Brand C products were higher-priced, full featured products, while the Brand A product was a lower-priced product with fewer features.

After having reviewed the actual products, respondents were asked to evaluate and rank-order the 16 product combinations in terms of the overall purchase interest. These 16 combinations were presented to respondents in black and white block lettering on 5" x 8" poster board cards.

One-hundred-and-twenty interviews were collected across five cities throughout the U.S. One-on-one interviews were conducted with respondents being recruited using a mail-intercept methodology.

Aside from the product combination sort, respondents were also asked a battery of brand imagery and demographic questions.

Analysis

Utility Estimation

Utility estimates were developed for each level of each attribute using an Ordinary Least Squares (OLS) approach. The typical form of this model is:

(1.1) ^Y = a + ß1 + ... + ßn Xn + e

where ^y = the predicted purchase interest measure;
ßn = the regression coefficient or utility weight for feature n;
Xn = is the dummy coded feature variable; and e is the error term

The average utilities for each level of each feature are shown on Table 1. The price variable was estimated using a linear model (as opposed to a part-worth model) so that only a vector coefficient was estimated.

The data suggest that the full featured products (Brand B and Brand C) were the most preferred and that any product should be packaged with peripheral features. The price coefficient, while not appearing large is very important. When multiplied by $69, the lowest price, the coefficient becomes -3.657. When multiplied by $199, the highest price, the coefficient becomes -10.547. Obviously, the higher the price, the greater the negative impact on the product. In addition, each respondent is also assigned an intercept term.

Market Simulations

Market simulation models are then conducted by first computing product utilities to determine which product is the most preferred of all those simulated. The product utility is simply the sum of all the utilities that define a product. For instance, Table 2 illustrates the product utilities for the two products currently in the market.

These product utilities are then used in a market simulator. Market simulations were conducted using a form of the Bradley-Terry-Luce (BTL) probability model. This model assigns a probability of purchase to each product being simulated for each respondent.

The probability of purchasing Product 1, P(1), is computer by:

(1.2) P(1) = U(1) / U(i)

where U(1) is equal to the product utility for Product 1; and U(i) is the sum of all the product utilities in the simulation.

For the two product market described above, the associated product shares are shown below.

Table 3
Two Product Market
Estimated Product Shares

U(i)		P(i)
Product 1: Brand B	1.587	23.6
Product 2: Brand A	5.141	73.4
U(i)	6.728

Estimation of Unit Demand

Assumptions

Several assumptions were made in order to develop a unit demand estimate. These included:

• A static market: unit demand estimates to be developed were to be projections assuming that the total number of units sold in the category would be equal to the number of units sold the prior year.
• The total number of units sold the prior year was estimated to be 3.650 million units.
• Brand A, the market leader, accounted for approximately 3.500 million units (96 percent); Brand B accounted for approximately .150 million units (4 percent).
• Brand A, Brand B and Brand C represent the entire market.
• These three products were used in the simulations. All products included a peripheral and prices were varied to measure the impact of price on unit demand.

Method of Estimating Unit Demand

Two estimates for market demand were developed a low-end, conservative estimate and a high-end estimate. The low-end estimate represents a product receiving very little advertising and promotional support, and with generally low awareness. The high-end estimate represents a product which received a great deal of advertising and promotional support, and with generally high awareness.

Since Brand A was the market leader, weighing Brand C's data up with a weighing factor based on Brand A's performance would provide an indication of how well Brand C might perform if comparable levels of awareness, advertising and promotion dollars, and distribution power were in place.

Conversely, weighing Brand C's data down with a weighing factor based on Brand B's performance would provide an indication of how Brand C might perform if lower levels of awareness and promotional dollars were spent on its introduction.

By weighting Brand C's data back to real world figures, the ensuing share of preference estimates would approximate actual market share data and as a result be more projectable.

• Formula 2.1 illustrates the computation for the low-end unit demand, and is based on the existing unit share data of Brand B's product.

(2.1) U(s) = P(s) * C(a) * 3,650,000;

Where U(s) is the predicted unit demand for Brand C;
P(s) is the predicted share of preference for Brand C;
C(a) is the calibration to Brand B's performance; and
3.65 million is the total market size.

• Formula 2.2 illustrates the computation for C(a), the calibration to Brand B performance.

(2.2) C(a) = A(a)/P(a);

Where A(a) is the actual market share for the Brand B; and
P(a) is the predicted market share for the Brand B.

• The high-end estimate represents a product developed by a company that has spent a great deal on advertising and promotion, and has a high level of awareness.
• The formula for the high-end unit demand is based on the existing unit share data of the Brand A product.

(2.3) U(s) = P(s) * C(n) * 3,650,000;

Where U(s) is the unit demand for Brand C;
P(s) is the predicted share of preference for Brand C;
C(n) is the calibration to Brand A performance; and
3.65 million is the total market size.

• Formula 2.4 illustrates the computation for C(n), the calibration to Brand A performance.
  

(2.4) C(n) = A(n)/P(n);

Where A(n) is the actual market share for the Brand A; and
P(n) is the predicted market share for the Brand A.

Findings

Calibrations

The first simulations were used to develop the calibrations C(a) and C(n). These first simulations predicted that Brand A should account for approximately 76 percent of current market volume; and that Brand B should account for approximately 24 percent of current market volume. The real would data these models were actually 96 percent and 4 percent, respectively.

Therefore, the calibration, C(n), to estimate unit demand for a high share Brand C product was 96/97; the calibration, C(a), to estimate unit demand for a low share Brand C product was then 4/24.

Projections

Including Brand C (peripheral included) in the following simulations then produced estimates for a three-product market.

Table 4 illustrates the calibrated market share estimates for Brand C at five price points and at two levels of support (estimates have been rounded to the nearest tenth).

Brand C's low-end share at $149.99 is computed by multiplying the predicted share by the low-end calibration (15.2*4/24=2.5). The high-end share at $149.99 is computed by multiplying the predicted share by the high-end calibration (15.2*96/76=19.2).

Table 4
Weighted Market Share Estimates

	Low-End	High-End
$149.99	2.5%	19.3%
$139.99	3.8%	27.3%
$129.99	5.6%	40.1%
$119.99	6.5%	46.4%
$109.99	6.8%	48.4%

Table 5 illustrates the low-end and high-end demand estimates developed at the five Brand C price points (estimates have been rounded on the nearest 1,000).

Table 5
Weighted Brand C Demand Estimates

	Low-End	Mid-Point	High-End
$149.99	97,000	396,500	696,000
$139.99	140,000	569,000	998,000
$129.99	204,000	834,000	1,464,000
$119.99	236,000	964,000	1,692,000
$109.99	246,000	1,006,000	1,766,000

The range between the low-end and high-end demand estimates are wide and provide an estimate of projected sales if Brand C were to spend the same dollar amounts as the low-end or high-end brands on advertising and promotion.

Actual Market Results

Brand C was introduced at the $149.99 price point with very strong advertising and promotional support. Actual first-year unit sales for Brand C were between 540,000 and 660,000 units. Table 6 illustrates the projected (using the high end estimate) and actual sales data for Brand C in its first year.

Table 6
Actual and Estimated Demand

Brand	Price	High End Demand Est.	Actual Demand
Brand C	$149.99	696,000	540,000-660,000

Conclusions

Conjoint analysis is one tool that provides the flexibility altering features and estimating the impact on unit demand. Unit demand forecasts should be further grounded in reality by basing the demand estimates on existing known unit demand for other products in the same category.

The method described in this paper asks the researcher to make certain assumptions about the market specifically that there will be no growth in the next year. For this research, this was a realistic assumption. However, in general, static markets are the exception and not the rule. As a result, if demand estimates are to be accurate, the market size data must be based on realistic assumptions.

This type of modeling allows marketers to explore a range of product, pricing and promotional support alternatives providing a realistic guide for strategic market planning.

Additional Readings

Green, Paul E.; Donald S. Tull and Gerald Album (1988), Research For Marketing Decisions (Fifth Edition): Prentice Hall.

Kuhlfield, Warren F.; Mark Garratt and Randall D. Tobias (1993), "Nonorthogonal Experimental Design Theory with Marketing Research Applications." Paper presented to the AMA Advanced Research Techniques Forum, June 13-16, 1993, Monterey, Calif. Jonathan Weiner, Sr. Vice President-Research Methods & Analysis

The author thanks Dick McCullough, Randy Batsill and three anonymous Marketing Research reviewers for their helpful comments.

MACRO CONSULTING
Intelligent Decisions