Applied Data Innovation Starts Where the Model Stops

By By Garry Rosenfeldt, Senior Principal Analytics, Insights and Strategy, Allant

People sometimes talk about predictive models as if they are supposed to be perfect. They are not.

A model is only as good as the information you have at the moment you build it. It can tell you who is more likely to respond and who is less likely to respond based on the variables you know or are able to derive That is valuable. That is exactly what it should do. But it does not mean the model has captured every factor that influences real human behavior. No one’s modeling approach is perfect, and it never will be. The point of innovation is to look at alternate data sources in different ways to make your predictions better.

That idea sits at the center of how I think about applied data innovation.

Key takeaways:

1. Predictive models are a starting point, not a final answer.
   A model can identify who is more or less likely to respond based on known variables, but it will never capture every factor that influences behavior. The article makes the case that model limitations are not failures; they are opportunities to keep learning.
2. New signals can reveal overlooked demand.
   By adding localized transaction data, including category spend patterns in specific ZIP codes, Allant was able to identify prospects in lower model deciles who lived in areas with high relevant spending activity. That added context helped uncover a population that otherwise may have been undervalued.
3. Applied data innovation improves audience strategy by building on what already works.
   The article’s central point is that innovation does not mean abandoning a strong model. It means making it sharper by identifying what is missing, testing additional data sources, and using those learnings to expand the high-value prospect pool in a smarter way.

In one recent specialty retail program, we were looking at a familiar modeling challenge. We had a prospect universe around store locations. We scored that universe based on known characteristics such as distance to store, income, net worth, category interest, and broad range of other relevant variables across multiple data sources. From there, we ranked people into deciles, with the highest-propensity prospects at the top and the lowest at the bottom. As expected, the people with the strongest scores responded at higher rates. That is how a model should behave.

But the real question was not whether the model worked. The real question was whether there was still more to learn.

One of the concerns raised in the discussion was whether it might look bad if people in lower deciles could still perform well. I feel strongly that this is the wrong way to think about it. Lower-decile responders do not mean the model failed. They mean there are influences on behavior that were not fully captured in the original scoring logic. That is not an indictment of the model. That is an opening for innovation.

So we asked a different question: What if the next useful signal was not just about the individual, but about where that person lives and what is happening around them?

To test that, we used transaction data aggregated at a local geographic level. We looked at the number of credit cards in a ZIP code were being used and started to build a richer picture of category behavior. We examined spend at sporting goods merchants, country clubs, competitive brands, golf-related merchants, and other signals tied to the category. That let us move beyond static individual traits and start looking at the commercial environment surrounding a prospect.

From there, we ranked ZIP codes based on how much spend was happening in the category. Then we took those high-intensity geographies and layered them back against the existing model. Specifically, we looked for people in lower deciles who lived in places where there was unusually high spending in the kinds of products and merchants that mattered to the retailer. That was the population we targeted. And that population responded really well.

The thinking behind it was straightforward. If a person lives in an area where the category is highly active, where neighbors and nearby consumers are already spending in relevant ways, that context may be telling you something important. We describe it as a kind of “keep up with the Joneses” effect. The original model did not specifically account for where someone lived in that way, or what the surrounding spending environment might imply. Once we added that information, we improved our ability to predict a responder.

That is what applied data innovation looks like in practice.

It is not innovation for the sake of novelty. It is not throwing random new data into a workflow and hoping something happens. It is starting with a strong model, understanding its limits, identifying what may be missing, and testing whether an additional data source adds real decision-making value. In this case, the added context gave us a new way to find opportunity in a part of the audience that otherwise would have been undervalued.

What I like most about this example is that it shows how innovation expands the prospect pool in a smart way. The value of this work was that it increased the available high-value prospect pool the retailer could go after. That is a good outcome all around.

It also reinforces something I believe strongly: better audience strategy does not come from pretending older approaches were worthless. It comes from building on what already works and making it sharper. A model gives you a disciplined foundation. Applied innovation makes that foundation stronger by introducing new signals, new attributes, and new ways of understanding demand.

For marketers, that matters because growth opportunities are often hiding just outside the logic of the current model. And for analysts, it is a reminder that the work is never really finished. The next breakthrough often comes from asking, very simply: what are we not seeing yet?