There’s No Crystal Ball for Business
Is a new mindset key to predicting the future?
Predicting the future is hard. I am fairly certain of this because I’ve never won the Powerball, and neither have you.
Still, as business leaders, we need to understand how the market and our customers are changing, in order to make effective decisions. The most common way to gain market and behavior insights is through data, but most people use simple models of historical data to predict future behaviors. Whether we’re in “typical” market times or ones of significant turbulence like now, this simply isn’t enough. You need to think creatively about your data and your predictions to select intelligent models so that your predictions are as useful as possible.
There are a number of techniques and tools you can use to create useful projections for your business, and I’ll cover some examples of common predictions and how you can approach them. Specifically, we’ll cover:
- How to project metrics like revenue when your data is very noisy
- How to predict customer actions like churn
Predicting revenue
Let me start with a caveat. Predicting the future [1] is not an exact science. It will never be 100 percent accurate. Even so, forecasting provides valuable guidance and continues to improve the longer you use it. Applying a better approach to modeling can greatly improve overall results for any business.
Prediction modeling is difficult because real-world data is complex. For example, given a daily revenue data (where each point is a day), try to predict revenue for each day the following week (see figure 1).
Like most real-world data there are patterns and trends hidden in this data, making extrapolation difficult.
Basic techniques like linear regression would lay out a general trendline, but not a prediction for every day next week. We could use advanced techniques like double exponential smoothing but they are difficult to implement.
Luckily, there is an easier way to model our prediction for the following week.
Instead of trying to understand the data as a whole, the repeating cycle means we can focus on each day of the week independently. For example, using just Mondays shows us that the Monday trend and pattern is actually a straight line (see figure 2).
By creating a separate trend line for each day of the week, we can build a model for revenue every day next week with a fairly high degree of accuracy (see figure 3).
If you’re like most businesses, you don’t need to predict revenue on a day by day basis. That’s fine. This model works with weekly, monthly or quarterly data that follows common cycles.
But, you might have a business with a non-typical cycle or a seasonal business like tourism. That’s fine too. If your business data doesn’t have clear cycles or trends, you can fall back to general forecasting with trend lines or more advanced modeling techniques like the ones mentioned above.
Predicting Sales
Understanding the sales pipeline is one of the most common challenges businesses face. In order to predict revenue, it’s important to have visibility into the sales funnel. But how do we predict the volume of leads coming in through the top of the funnel? One way to do this is by implementing lead scoring.
Lead scoring is a process that ranks leads, ordering them from “hottest” to “coldest.” This allows you to allocate resources to “hot” leads while maintaining different levels of contact with cooler leads as you move them through the sales funnel.
Using lead scoring to predict future revenue starts with data collection. Relevant data can include demographic information about the prospective company or person, insight into purchase data and behavior metrics.
For example, useful company-level demographic data may include a company’s industry or size. Personal demographic information could be a person’s title and role at a company, their budget or authority level. Behavioral metrics could include the number of times a person has opened an email from you or attended a webinar.
Using this data, you then allocate points to each lead. You can even give negative points for “cold” behaviors, like unsubscribing from a mailing list.
While the traditional scoring mechanism is helpful, it does not provide you a prediction as to whether the lead will convert or not. However, there are a number of machine learning techniques that can help you. One of the more basic approaches is something called a logistic regression. A logistic regression is similar to a linear regression except that the variable you are trying to predict is a binary outcome, will the customer convert or not in this case, as opposed to a continuous value.
No matter the technique, all of them will rely on high-quality historical data that is used to train your predictive model — the more of it the better! As you go score leads and process outcomes, you’ll want to continually re-evaluate which attributes you are collecting data, and how you weight your scoring, to see if you can better predict your outcomes in the future.
So, if we don’t want to lose customers we should make them all happy, right?
Sure, but that is harder than it sounds. No product makes everyone equally happy, and there are always new competitors waiting in the wings trying to steal our customers away. It would be great if there was some way to know which customers were most at risk so we could focus on making them happier.
Luckily there is, and it’s called Churn Prediction.
Churn Prediction works much like Lead Scoring, but instead of estimating the likelihood a lead converts into a customer, it estimates the likelihood that a customer will churn. The same sort of company / personal demographic data (e.g., company size) are appropriate here. However, the behavioral data will need to be updated to also include information about how the customer has engaged with your product.
● Product usage: How many times has the customer logged in? What is the cadence of their interaction with your product? How many users are there using your product at the customer? Has the customer changed their level of service (upgraded or downgraded)? What is the tenure of your relationship with the customer?
● Support: How many support tickets have been filed by the customer? What was the severity of each ticket? What was the tone of the tickets? How quickly were the tickets resolved?
● Social: Has the customer ever mentioned you in social media? Has the customer referred you to any other companies? Are you allowed to use the customer’s logo on your website? Was the customer willing to produce a case study for you to publicize?
With all of these attributes defined and data collected, you’ll be able to score all of your customers and have a good sense as to which ones are likely going to stay customers and which are going to churn.
As with everything in predictions, there is no crystal ball that will perfectly tell you what will happen in the future. The goal is to find the most reliable leading indicators of your business that will give you enough time to adjust your strategic plan before it is too late.
[1] A. Wieckowski, Predicting the Future (2018), HarvardBusinessReview
