Chris Penn
Co-founder, Trust Insights & PodCamp
Marketers face increasingly complex challenges in today’s data-rich landscape. What’s likely to happen? Which prospects are likely to convert? Which customers are likely to be good or bad customers?
To answer these questions and challenges, we turn to the art and science of predictive analytics.
Predictive analytics is an umbrella term that defines a set of tools, methodologies, and techniques that enable businesses to analyze the past performance to estimate future outcomes.
There are two broad, primary predictions created in predictive analytics: classification and time-based regression. Let’s break these down.
Predictive Analytics: Classification
Classification is a type of prediction where we attempt to ascertain the probability of something falling into a specific category. For example, suppose you’ve got a series of prospects in your marketing automation software. Your goal is to hand off as many high-quality marketing qualified leads (MQL) as possible to Sales, while avoiding handing off time-wasting, low-quality MQLs.
In the classification type of predictive analytics, machine learning software would examine all the characteristics of prospects that you’ve already determined are good or bad, and then build a model – a piece of software, essentially – that can classify future prospects.
For example, suppose you had a record like this:
Name: Justin Levy
Company: Demandbase
Industry: Computer Software Company Size: 500-999 employees
Annual Revenue: 100M
City: San Francisco
State: CA
Postal Code: 94107
Number of Contacts: 7
Number of Touchpoints: 54
MQL Status: Qualified
A machine learning model would examine every record in our database and determine whether or not any of the fields, alone or in combination, had a stronger mathematical relationship to our outcome – MQL Status of Qualified – than others. Based on that information, we’d then be able to screen future prospects and make a predictive judgement about the likelihood that they’ll also be qualified. Predictive analytics software will also let us know what data points don’t matter; it may turn out, for example, that annual revenue is a poor predictor of whether a lead is qualified for not.
The best predictive analytics software also learns; as each new prospect is qualified or not, the machine learning software refines and changes its model to reflect its new knowledge. Over time, classification- based predictive analytics tends to become more accurate as long as the underlying data is sound.
Classification-based predictive analytics is typically built into more sophisticated CRM and marketing automation software packages as an enhancement to lead scoring; your vendor may already offer it.
You’ll see classification-based predictive analytics in use with Demandbase One‘s content recommendation system as well – when a visitor arrives on site, Demandbase One will make content recommendations to them based on existing audience data and the propensity that a visitor will engage with a certain piece of content. By recommending the best possible content, we nudge our prospects closer to conversion.
Predictive Analytics: Forecasting
The second type of predictive analytics is forecasting, where we attempt to ascertain when something is likely to happen based on previous known information. Time-series forecasting is one of the most powerful applications of predictive analytics; done well, it changes our marketing from being reactive in nature to being proactive.
Time-series forecasting requires our data to have two general attributes: seasonality and cyclicality. Seasonal data is data that tends to fluctuate across seasons. In B2B marketing, we know certain times of the year tend to be less or more active than others. In the Northern Hemisphere, the period between the end of November and the beginning of January tends to be one of the slowest for getting deals done; many people are out of the office for the winter holidays. The same is true from the end of May through the beginning of September. This seasonality lends an aspect of predictability to our data.
Cyclicality is similar to seasonality, in that there’s a recurring pattern, but it’s not tied to the calendar year. For example, B2B marketing tends to be more active on weekdays than weekends; B2C marketing may be the reverse. Why? Consumer behaviors tend to favor shopping on the weekend, while B2B behaviors tend to favor occurrence during the work week. Cyclicality can also occur over multiple years; for example, certain political organizations see large swings every four years in the United States, or every five years in the United Kingdom.
Data that has neither seasonality nor cyclicality is impossible to forecast accurately; you cannot forecast something that has never happened. For example, no technology correctly predicted the start of a global pandemic in late 2019. No Marketing technology correctly predicted the ascent of TikTok or the rise and fall of Clubhouse – none of these events happened before, and thus could not be forecast.
However, much of the data Marketers work with on a regular basis does have strong cyclical and seasonal components to it, making it viable for time-series forecasting.
One of my personal favorite examples is when people search for the term “Outlook Out of Office”. Behaviorally, what is a person doing when they Google for that term? They’re about to go out of the office, presumably on holiday or some other form of unavailability, and they need to remember how to turn that particular feature on in Microsoft Outlook. Thus, it’s a proxy for people being unavailable to us as Marketers. If we examine this term in Google Trends, we see:
What’s clear is there’s both seasonality and cyclicality to the existing data. There are clear spikes every winter, just before the holidays, and a rise and fall every summer – even worldwide, even including the Southern Hemisphere for when seasons are opposed to the Northern Hemisphere. Thus, this is an excellent candidate for using predictive analytics.
Over two dozen different algorithms exist for doing time-series forecasting; today, most software packages use an ensemble of algorithms, a combination, to do forecasting more accurately.
Above, we see several different algorithms being evaluated for the greatest accuracy. In the example, the software chose an algorithm called Prophet, a free, open-source forecasting library published by Facebook’s Data Science team. Once evaluated, we end up with a final forecast model:
And when we label that forecast model, we see the weeks when people will be MOST out of the office:
Logically, we shouldn’t be launching major campaigns any time our forecast is above the red horizontal line – that’s when out of office searches are the highest. Conversely, we should be launching major campaigns any time the forecast is below the green horizontal line – that’s when people are least likely to be out of the office.
Why Do We Care About Predictive Analytics?
The discipline of predictive analytics is vitally important for planning. If we have a sense of what’s likely to happen, or we can score entities like contacts, leads, and opportunities by their propensity to do business with us, we accomplish two major goals.
First, we reduce resource usage. In the case of things like lead scoring, our sales force spends less time on poor quality leads and opportunities, freeing them up to spend more time on good opportunities (thus increasing our likelihood to close deals) or spend time on more opportunities (thus increasing our pipeline).
Second, we reduce reactivity. Marketing tends to be a reactive discipline; something happens and we react to it. Google changes its algorithm and we react. LinkedIn rolls out a new feature and we react. A customer fills out a lead form on our site and we react. That’s fine and important, but being purely reactive means we never have the opportunity to plan ahead, never have a chance to apply strategy and thought to what we’re doing. Predictive analytics, particularly forecasting, gives us insights into what’s likely to happen so that we can plan our time and resource expenditures better. We do more with less, and we increase our own job satisfaction with less panic and less scrambling to get things done in a timely fashion.
As we expand the role of Marketing technology throughout the customer journey, predictive analytics becomes even more important. When companies pivot from account-based marketing (ABM) to account-based experience (ABX), predictive analytics goes with the territory. Instead of simply doing lead scoring, we use the same scoring algorithms and models to predict customer churn, upsell opportunities, or candidate customers for customer advocacy and evangelism.
Time-series forecasting makes its way into account-based experience by forecasting customer satisfaction levels based on data like NPS scores, as well as forecasting when upsell opportunities are more probable.
How Do We Get Started with Predictive Analytics?
For most marketers, predictive analytics will make their way in some fashion into the tools and software you already use, from platforms like Demandbase One to even free applications like Google Analytics. Most modern Marketing automation and CRM software already has some degree of predictive analytics built in around lead scoring, even if it’s not driven by machine learning or artificial intelligence.
For more advanced applications, such as the time-series forecasting we demonstrated above, many companies have or are building data science capabilities that leverage popular tools like Tableau Software, Alteryx, IBM Watson Studio, and many others to offer predictive analytics capabilities. As an example, IBM Watson Studio has automated time-series forecasting built into its AutoAI software, so the user doesn’t need to write a single line of code to obtain a forecast and an accompanying machine learning model that can be put into production.
Predictive Analytics as Part of Your Marketing Technology Mix
If deploying predictive analytics isn’t already on your marketing technology roadmap, be sure that it is in some fashion. Make sure the marketing automation and customer experience software you’re using has predictive analytics either in the feature-set or in the product roadmap so you can spend less time reacting and more time building the winning Marketing that you know your organization is capable of.
Chris Penn
Co-founder, Trust Insights & PodCamp