Marketing productivity is down, and marketing resistance is up. It’s tough to disagree with the statements that Craig Wood, group president of marketing research firm/consultancy Yankelovich, made in the beginning of his luncheon keynote speech at the Kansas City Direct Marketing Days on March 7.

To improve response amid this resistance, he continued, marketers continue to send more messages. That’s why whereas in the 1970s the average consumer saw 300-500 ad messages a day, today he is bombarded with 3,000-5,000. The result: even more consumer resistance.

Wood gave several reasons for this resistance, all of which ties in to what Wood said are the five trends driving direct marketing:

1. “iPriority,” which he defined as “doing what’s right for you.” It’s not a repeat of the “me generation” thinking of the 1980s, however. “It’s not all about me. It’s more about being true to yourself and putting your needs first.” This sentiment can be seen in the rise in people treating themselves to visits to day spas, for instance.
2. Willful disobedience, or “being naughty once in a while saying it’s okay to color outside the lines once in a while,” Wood explained. Kids wearing pajamas to school and “What happens in Vegas stays in Vegas” are examples.
3. The upside of obscurity. “It’s finally okay to be a little bit different,” Wood said. He used the success of the documentary “March of the Penguins” as an example: “When it first came out, seeing it was an obscure thing to do.” But obscure in a fashionable way, as evidenced by the word of mouth that eventually led to it taking in more than $80 million in box office.
4. Age nullification. “It’s about being chronologically irrelevant.” That’s why baby boomers on average see “old age” as starting at age 75 roughly the average life expectancy of Americans.
5. Looping. “It’s about understanding what’s going on behind the scenes,” Wood said. “In some ways it’s about taking the need to know to a new level.” Clearly this ties in to consumers’ lack of trust in advertising and businesses. The more they feel they know, the more trusting they’re likely to be.

To make these trends work for you, Wood said, you need to put attitudinal information into your database alongside the demographic and behaviorial data. As an example of “targeting with attitude,” he referred to a midsize cataloger/retailer that understood its customers demographically older, middle-income women but not attitudinally. After conducting customer research, however, it found that 75% of its customers fell into one of two attitudinal groups. Group one consisted of cynical, nostalgic homebodies who favored the tried-and-true; group two was made up of forward-thinking, style-conscious buyers who were price- and feature-driven.

The company refined its marketing to these two group, going so far as to tailor product descriptions for each. For instance, when it crafted copy for a double boiler that emphasized the past (“just like Mom used to use”), it obtained a 48% lift for that one item among the group. In fact, overall both groups saw a double-digit lift in response, Wood said.

Neural network models have been around since the 1940s, but most marketers still find them a bit of a mystery.

“I’ve seen the term thrown around quite a bit, especially by analysts, to impress or intimidate people,” says Norbert Schumacher, senior statistical consultant with Colloquy, a provider of loyalty marketing services. “They know their managers may be impressed, thinking it’s being used as a brain model.”

A linear model is characterized by a simple linear relationship between the “target variable” (the thing that you want to predict) and the “feature variables” (the variables that you use as predictors). For example, if you had 10,000 customers, you may notice that generally speaking for every 10 years of age of the customer, the profitability of that customer goes up by $3. A 20-year-old customer might be predicted to be worth $1 per year, a 30-year-old would be predicted to be worth $4 per year, and a 40-year-old customer would be predicted to be worth $7 per year. Here the target variable is “profit” and the “feature variable” is “age.”

By contrast, a neural network is a non-linear prediction. So a 20-year-old might be worth $1 in profit per year, a 30-year-old $4 in profit per year and a 40-year-old worth $5 per year. Since the profit prediction does not go up by a constant amount, and might even start to go down again for 60-year-old customers, you have a non-linear model, and that’s what neural network models are designed for.

A neural network model can predict profitability based on customer age even though the relationship between the two variables is non-linear.

Neural network models, Schumacher says, are most useful when the target variable has a high useful to irreverent ratio, or when interpretation is not the goal. But when the data contain outliers or when data are missing, they lose their effectiveness.

In a Webinar held Feb. 22, Schumacher pointed out four challenges in neural network modeling:

• Neural networks can “overfit” the data, or call out random errors within the data. “For example, our prediction for 20-year-olds might be a profitability of $1 per year and our prediction of 30-year-olds might be $4 per year, our data set might contain customers who are 21.5 and 21.75 and 21.8 years old,” Schumacher said. “Clearly, any sensible prediction of profitability of the customers within the 21 to 22 age range should be similar to one another. The differences in the data between these three example customers will be due mostly to random error. An “overfit” model is a model where we used the ability of neural network models to capture the non-linear relationship of age and profit, for example, but then it went on to model some of the inevitable ‘noise’ within the data.”
• Neural networks are impossible to interpret. “Linear models produce easily interpretable results like ‘older customers are more profitable than younger customers,'” Schmacher said. “A neural network model, by contrast, does not. But you can still use a neural network model to make customer predictions. So while, in our age/profit example, we cannot in general say that younger customers are less profitable than older customers – because the 60-year-old customers were relatively unprofitable – we can still use a neural network model to produce a prediction for any new 60-year-old customer that we might see.”
• Neural networks do not handle missing data well.
• Neural networks are more difficult to build (an experienced analyst is needed).

Kelly Hlavinka, director of Colloquy, and Schumacher also outlined five marketing plans in which some sort of model should be considered:

1. Customer acquisition efforts
   If pure acquisition is the goal, then a response model of any kind (logistic regression or neural net model) is a good fit. The categories of data that need to be considered are demographic overlays, direct mail responsive scores (including presence on multiple list sources), physical attributes such as distance to nearest retail location, and financial scores from credit bureaus.
2. Sales campaigns (for instance, catalog mailings)
   This, they said, is harder to shoehorn into a common neural network. Pure response to a catalog doesn’t consider the consumer’s level of spend. Pure regression neural network models account for level of spend but do not indicate responsiveness. The solution is to build a two-stage model – the first to predict the probability that a consumer will respond to a catalog; the second, using data from the first-stage responders, to predict the level of spend.
3. Upsell campaigns
   Hlavinka believes that this is one of the best reasons to use a neural network, as you’re combining qualitative information such as customer acquisition rates with responsiveness, and creating more than two categories.
4. Customer retention efforts
   Customer lifetime values are positive and highly skewed (like the price of a home), therefore ordinary regression models (linear or nonlinear) are not appropriate. The recommended way to model this is via “survival analysis,” which is a specific type of generalized linear model. An alternative way to model customer attrition is to take a group of customers, see which customers attrited in some period of time, and try to predict the attrition/non-attrition event via the predictor variables given.
5. Increasing brand affinity through communication relevance
   This would be a reason for a neural network model called a Self-Organizing Feature Map (SOM), in which multiple customer attributes can be run through to create an algorithm. Where they fall in the SOM can help cluster your customer list into a smaller group of silos to determine if they should be informed about a specific campaign.

“The key point is that you have got to keep focused on the whole reason to use any model in the first place, and that is to increase your marketing efficiency,” Hlavinka says. “If you don’t lose sight of exactly what you’re trying to do, you are on the right path for a healthy debate for the right fit for your data. The analyst’s work can help with the marketing partnership and helps drop any preconceived notions marketers may have when they come up with their plan.”