So how many customers can you upsell in your next campaign? How many cancellations can you expect? How can you convert content recommendations into sales? What lifetime value can you expect for each new customer you acquire?
For each one of these questions, there is a model to give you an answer. Even their names give away their functions: “Predictive Response+LTV for Targeted Customer Acquisition”, “Prioritization via Response+LTV+Variety/Rotation for Email Channel”, or the really exciting “Predictive Churn to pre-empt cancellations”. But how many models do you need? And how do you manage the stack of models? The concept sounds neat. The execution can be messy.
Start with a good view
“We’re awash in data, but we are thirsty for knowledge,” said Jeff Tomlin, CMO at Vendasta. Marketers can succumb to “paralysis by analysis”, or fail to ask the right questions. So start from the top: understand the high-level things you need to know that drives the business, Tomlin said.
Companies will gather data and form hypotheses, and “cross it over all the stakeholders in the organization and have a forecast they are trying to hit.” said Jason Katz, founder and principal of Growth Marketing Advisors LLC. That is the basic strategy companies start with before crafting models that address data gaps, he explained. After all factors are considered, a firm could craft 12 models to identify an opportunity. “But you don’t need to run all 12. If you get one or two (right), it’s a big win.” Katz said.
Consultants have the expertise, but clients know their business. “The clients can help to add business sense and insights that make the model more accurate.” said Yohai Sabag, Chief Data Scientist at Optimove. “But, we have our best practice in choosing the right model, so usually we guide them which model to pick.”
Good data = good input
“The first step in building a model is to prepare and understand the data. Both sides are responsible for the quality of the input.” Sabag continued. Yet even this obvious foundation is sometimes ignored by clients.
“Almost always, the data is terrible,” Katz said. Larger, more mature companies will have some data sets in good order, he pointed out. That picture changes with small companies. “Grow now, test to see if it works, cleanup afterwards,” Katz said. “Marketing operations and data hygiene are afterthoughts.”
Lack of data uniformity is often the problem. A client might pursue a global enterprise account, only to see it effort compromised by chasing after the wrong people in the wrong territories for lack of a unified account view, Katz said. Or a database can become useless for lack of taxonomy and correct nomenclature, which is a complicated way of saying that users inputted the data in the wrong order under the wrong labels. It’s hard to build personalization when “there is no uniformity. You can’t unlock the power of your own data,” he said.
“Getting people to understand data, how to use data and thinking about data is a challenge.” added Tomlin.
Nice model
Now ponder what can go right: the client knows his business and his data is good. Time to start building models. Now what?
Vendasta takes a platform-centric approach towards model-building by offering a turnkey solution. The basic metric is efficiency of sales, and the focus points are demand generation, product metrics, sales metrics, retention, expansive sales and scale, Tomlin explained. These particular metrics emerged after surveying hundreds of clients to find out what their greatest challengers were in achieving growth.
By using the platform, the client can take a “two step view”, first looking at the macro picture, then looking at the details. The goal is to know what is the cost of customer acquisition compared to the lifetime value of the customer, Tomlin said.
“I’ve never seen a nice stack of models,” Katz added. Typically, large companies will model by channel, or they can use the same model across channels. Models begin to differ when companies need to figure out how to cross-sell or up-sell. “You get a probably lifetime value for each path. Highest LTV wins.” Katz said.
Companies may increase the number of models they use if their resources are limited, Katz continued. Say a company is doing an e-mail campaign and a unit only has a limited number of periods when it can access the e-mail list. That constraint forces the unit to model the best approach, then go with it, he explained. Or take an ad campaign. Ad dollars will be limited. “A marketing mix model gives you the channel mix where to put the dollars.” he said.
Still, no matter how many models there are in the stack, they have to work together as a system. “Each model can use the conclusions (insights) of the previous model in the line as an input. There’s no one way to array them.” Sabag said. “Even if there’s more than one model in the background, there should be a mechanism that aggregates their results into one figure, to provide a single answer.” Sabag continued. “Sometimes, in the order to get higher accuracy we (Data Scientists) should use an ensemble of few models, but all the results should be gathered into one answer.”
“There is no sweet spot here, the tradeoff is accuracy vs simplicity.” Sabag noted.
