Predictive analytics models are cross-disciplinary marketing tools that require participation and contributions from a broad range of disciplines. As such, you’re going to have to bring a lot of different types of people together with diverse personalities, and work styles may clash. I touched on the specific job titles that are necessary for predictive analytics models in my last post, but thought it would be a worthwhile exercise to identify what will likely be the two opposing factions in any disagreement amongst the predictive analytics team.
The first group that is required for predictive analytics models are the real-world marketers - the people who know how to create a marketing plan and measure results. The second group will consist of hardcore statisticians. These two groups of people are often at complete opposite ends of the spectrum in terms of temperament and the ways in which they work, but marketers and statisticians must have an environment that allows them to fully express their opinions while being open to the different perspectives of their colleagues. The path of least resistance to team harmony is to focus everyone on the art and science of their predictive analytics models. Root these two different camps in the real world so they focus on changes in economics and global trends rather than conflicting opinions. This often requires strong leadership from senior management and creation of an “esprit de corps” that celebrates and respects diverse opinions and skill sets. With everyone pulling together to make sure the team’s hypothesis holds up and the numbers dazzle management, predictive analytics models can be an effective tool that lead to tremendous results.
For years, marketers have used predictive modeling to forecast sales and determine relative return among competing marketing investments. Only in the last few years have those “competing” investments included online and web spending. These are important marketing expenditures, but have not readily been included in predictive models. We’re going to demonstrate the value these new data sources, such as web data, can add to your predictions. Through an examination of a specific scenario, I’ll demonstrate how the inclusion of web data can help you deliver insights beyond a simple forecast-- insights that can help shape your planning process.
Now let’s get into our scenario, and the nuts and bolts of predictive modeling. Generally, predictive modeling utilizes some form of regression modeling. Depending on what you are trying to predict, it may take a linear or nonlinear form. Let’s suppose you are working with a client, ‘Company U,’ which sells a number of entry-level electronic devices. Currently 80% of its sales are through brick and mortar locations and 20% of sales occur online. A brand manager at Company U has reached out to your team to develop an offline sales forecast for DVD players. She provides you with historic sales, pricing and media information and asks you to develop a monthly 12-month DVD sales forecast.
The first step in the predictive modeling process is always determining what data you need and what data you want:
Data You Need:
* Sales - Ensure that the sales data provided is discrete enough for your needs (often weekly DMA level or finer)
* Media - Media data at the finest level available. We can always aggregate it if need be
* Pricing - Pricing information specific to the level at which you are modeling (brand / model)
* Industry - Industry sales levels or macro economic factors that may impact the baseline level of sales for your brand
Data You Want:
* Website Data - Interactive data can provide great measures of what to expect for your brand in the future
* External Interest Data - Measures of overall interest or brand perception can help shape the outlook for your brand
* Economic Data - The overall state of the economy has an undeniable impact on most brands sales
Once you have the appropriate data in place, you can begin building your predictive model and it all starts with the basic form:
Sales = Industry / Econ + Media + Pricing + External Interest + Website Data
While this may seem overly simplistic, the devil is in the details. In reality, the format of a predictive model is simple because it’s based on the influences you can measure and the levers you can control. The complexity lies in how those levers interact with sales – and with each other.
We’ll begin by looking at media and a few tools that help us better understand how each channel relates to sales. For instance, a TV ad’s airing does not often immediately have an impact on sales. People see an ad that may incite them to purchase, yet may wait a few weeks before buying the product. That is why the relationship between media and sales is not always simple or straightforward.
The first way to counteract oversimplification is to test for a lagged effect. Consider a TV spot that airs today. It may have an effect this week or next week or three weeks from now. We can take this media lag into consideration by a certain number of weeks in order to determine the strongest possible relationship.
The second way to address the non-immediate effects of media is through ‘decaying’ media. Consider that same ad that airs today. While it may have its largest impact this week, the spot still retains some of its impact next week, some impact in week three and even some impact in the following weeks. In order to capture this relationship, analysts ‘decay’ the impact of media by using the concept of half-lives. By using a half-life of 3 weeks, we say that an ad retains 50% of its original impact three weeks later.
If your brand airs 120 TRPs of television media this week, how might that media weight react over time? Given a three-week half-life, the impact of that media would be:
* Week one = 120
* Week two = 95
* Week three = 76
* Week four = 60
* Week five = 48
* Week six = 38
* Week seven = 30
Using the concept of decayed media helps explain how media influences consumers and provides a smoother data set that shows the cumulative impact of media weights. When people think of regression they often think of the simple linear relationship, y = b*x which yields a straight line. While this might be sufficient in some occasions, it’s not always appropriate.
Our second tool for dealing with media impacts is the concept of diminishing returns. By replacing the standard regression equations with a more dynamic equation like the generalized logistic regression, y = k / (1 + exp(-b)). In this treatment, we develop both a ceiling and a growth rate represented by k and move away from the straight line to any number of new possibilities. Using this format allows us the flexibility to account for the fact that you can throw as much money as you want into media and it won’t bring you additional value after a certain point. This concept is often referred to as a saturation point or the point of diminished return.
If we go back to Company U’s media plan, we may see the concept of diminishing returns in print advertising. Print can be an important part of a media plan but if we were to consider increasing a standard print budget tenfold, we would most likely reach the point of diminishing return. When we examine the idea of increasing our print budget by this wild amount, we can see why we need a model that accounts for this type of relationship. Since company U is trying to sell DVD players a tenfold increase of their print budget would likely move them beyond electronics interest magazines, general interest magazines and into specialty magazines about topics ranging from cooking to hiking or else they would increase their ads within their current magazine base from one ad per issue to five ads per issue. Do we really think that an ad in ‘Birder Magazine’ is as impactful as an ad in ‘Electronics Monthly’ or that the fifth ad in ‘People’ is just as impactful as the first ad? No, we don’t. And while I know that this is oversimplifying things – we would account for ads per magazine and use TRPS to ensure the ads are appropriate for our target – it is merely meant to point out the fact that you need to move past a straight line mentality to really forecast expected sales levels.
Now that we’ve discussed a couple of tools that we can use for our inputs into the model, let’s shift to the data you want in your model. A lot of people take industry/macro economic factors, pricing and media spend as a given in a predictive model but then they stop there. While those factors are important, and they definitely influence your predictions, they are lumbering relationships. Now you may be asking yourself, ‘Lumbering relationships?! What the heck does that mean?’ It means that large scale predictions are good when examining media and macro factors. You can use these variables to forecast at the annual national level. However, if you try to use these variables to predict at the weekly regional level, you often run into problems. These ‘lumbering’ predictors have long term overall relationships with sales, but short-term variations are common.
With the influx of web data, we now have more ‘reflective relationships’ that can be included in our predictive models. Rather than relying only on passive econometric trends and company instituted factors like media spend and pricing, we are able to include what consumers are actually doing. This provides much tighter predictions that are able to adapt to changing levels of consumer behavior. Let’s look at our DVD player sales and how our marketing strategy might change by including web data in the predictive model.
A predictive model without web data may very well forecast sales, particularly at a national level. However, although the forecasts of national sales may be accurate, attributing the DVD player sales directly to marketing investments will be misleading without including web data. When missing from the model, the impacts of consumer web activity are hidden by other media spending, or even other external factors such as industry or economic trends. When included, “reflective” data such as display ad clicks, paid search and brand website activity (i.e. “click here for more retailers selling DVD Brand X in your area”) will help attribute sales to both higher level marketing media (traditional offline) and lower-level individual media online.
Without inclusion of the lower-level web data, marketers would overvalue the higher level marketing and undervalue its contribution to a sale created by online media and web activities. We may “feel” that some marketing dollars should support online, but we wouldn’t know what the right mix is and what a website click is worth to us in sales. When measuring the web data in our predictive models, the same tools should be used to fit the model with the data, namely, lagged effects, decay rates or non-linear diminishing returns.
Combining offline media data with web data (online ad activity and brand website data) and properly accounting for the nature of their relationship to sales will allow marketers to attribute sales across all marketing expenditures, and therefore, measure a truer return on marketing investment.
The inclusion of web data into econometric models has greatly increased their predictive power. We have seen this for a variety of products from durable goods to consumer package goods. The inclusion of web data, as a measure of overall demand, has helped push the predictive power of these models to within 5% of actual performance and in many cases to within less than 3%.
It’s a brave new world and inclusion of this type of data combined with the right team players and temperament can yield highly accurate models that quantify the complex relationships between marketing activities, sales and the economy.
Steve Kerho is the SVP, Analytics, Media and Marketing Optimization at Organic (www.organic.com).