Modeling Turns the Risky Unknown Into the Hopeful Potential

Let’s assume you must choose between two hypothetical options: a near-certain answer that is 95% accurate in two to three months; or a predicted estimate that is only 80% accurate in one to two weeks. In most common situations (i.e. not life-or-death), you would likely pick the latter. Why? Because the rapidly changing world doesn’t wait for certainty. It is better to try something based on assumptions and predictions, fail fast if inevitable, and promptly move on. Now, the real joke is most predictions don’t come close to 80% accuracy anyway. There are many reasons for it, but the most prominent one is insufficient data. Even seasoned data scientists with state-of-the-art toolsets would have a hard time overcoming data shortages. Predictive analytics is science, not magic. Filling in gaps The irony is that one of the primary purposes of predictive modeling is to fill in the gaps resulting from missing data. If we had every piece of information about the present and could see into the future, we would never need any modeling. Even commercial data compilers cannot possibly obtain all the information about every individual and household in the country. Do you really think they know everyone’s income and age, just because their data dictionaries are marked as 100% filled? It is an open secret that big chunks of data are filled with estimated values. Yet we use them because such estimates are far better than complete unknowns. Besides, decision making is more about following directional guidance than absolute truth. (Unless, of course, it is a matter of life and death.) But inaccurate sales predictions or inadequate offers to target individuals hardly qualify as such. We need modeling even in the days of big and ubiquitous data. Let’s mark "filling in data gaps” as the No. 1 reason why we build models. Increasing targeting efficiency One of the first applications of predictive modeling in marketing was for individual level targeting. Even in the 1960s—practically prehistoric in terms of computing powers—prominent publishers employed modeling techniques to cut down on marketing costs. Simply, wouldn’t it be nice to know who is likely to respond to offers before marketers commit to send expensive print catalogs to everyone in the database? Out of a million customers, who should receive the offers? By predicting likely responders in the planning stages, marketers were able to cut down the mailing size by 70% to 80%, and still obtained the desired number of conversions. That led to huge cost savings, which in turn fueled development of analytics teams in marketing organizations. Computers used in those ancient times were ridiculously large and expensive, and even more ridiculously underpowered and slow in comparison to basic modern PCs. But mathematics and statistics didn’t change much over time. The algorithms behind machine learning were developed a long time ago; it is just that computing power finally caught up with the idea relatively recently. In any case, "increasing targeting accuracy" was another main reason why modeling techniques have been employed by marketers. Summarizing large data sets for effective personalization You may ask why we need to build models in the digital age when contact cost is significantly lower than in the time of direct marketing. That may be partially true, but there are hidden costs of inaccurate personalization (or lack of targeting). Acquisition cost is very much real; therefore, customer opt-out is real loss. More importantly, personalization isn’t just about knowing who to talk to. We must also figure out what to offer, how often, and through what channel. For effective personalization, we must summarize large amounts of data into manageable segments, personas and affinity model scores, all through modeling. So, let’s add “converting complex data into simple answers” and “staying relevant with target customers” to the list. Finding hidden patterns There are side benefits, too. When done right, models provide consistent results, while rule-based targeting is often hit or miss. We can even estimate risks when going deeper into lower model groups to expand contact universes. With man-made rules, you just don’t know when the gas will run out. My favorite one is that models often reveal hidden patterns in data. When encountering hundreds of data variables, human minds often focus on familiar and intuitive ones such as age and income. But most human behaviors cannot be explained that simply, and mathematics and machines don’t care about what we are accustomed to. They just find the best predictors for the assignment, often exposing interesting interactions. When data become more compact and consistently predictable through modeling—by humans and machines alike—marketing automation become that much more feasible. Modeling in an Imperfect World We will never achieve complete knowledge about everything, and prediction will never be 100% accurate. But analytics is the pursuit of the least wrong answer given an incomplete knowledge base. And you will definitely benefit from such not-necessarily-perfect answers when making decisions. Aware of it or not, most of us are already using information provided by predictive engines on a daily basis. Then why not embrace modeling fully, and be more purposeful users? Marketers Should Use Modeling to Compress Large Data into Small Answers