Forecasting and predictive modeling for marketing analytics

Working on Intuit’s Forecasting and Optimization team in Marketing Analytics, I’m well versed on forecasting and predictive modeling. Today, I’ll explain what that they are and how we use them to delight our customers. I’ll also discuss the role artificial intelligence (AI) and machine learning (ML) play in developing the models.

Predictive modeling vs. forecasting

At its core, predictive modeling is about building a mathematical abstraction or representation of reality, a model, that has inputs (something you know) and outputs (something you want to know).

Predictive modeling is only possible through an abundance of data. Our team relies on data flowing from a wide variety of internal and external partners to establish a cohesive view of the nebulous world of inputs and outputs involved in media spend and fiscal year planning. On the input side, we work with numerous data scientists on the marketing analytics team across the web and online acquisition groups, strategic partners in marketing, and external advertising and agency partners. On the output side, we work with members of finance and marketing to help optimize their spending decisions.

To do our job, we use AI and ML, which are broad terms that cover a large selection of tools and applications we use for predictive modeling. Traditionally, AI is the ability of a machine to perform a task in a smart way. ML is a subset of AI algorithms that leverage data for problems such as classification, clustering, inference, and prediction. Let’s consider an analogy to explain how we use AI and ML.

We want to take a trip to the beach, and want an AI to solve this task: Get us to the beach. The beach is our model output — we arrive at or near the desired beach location. There are a number of vehicles – predictive models – we could use to get to the beach. We want to choose our vehicle based on the design criteria for your trip. How many friends are we taking, what supplies are we taking, and how fast do we want to go? In the modeling world, we may be more concerned with how accurately we arrive at the beach (predictive accuracy), than we are with the individual roads or understanding the inner workings of the car engine we took to get there (model inference).

We have a bunch of options for our vehicles. We might first want to borrow our friend’s new Tesla (for example, deep learning and neural networks) that has lots of shiny features. We might get wonderful predictive accuracy and arrive exactly at the beach, but our understanding of the electric engine might be obscured compared to an internal combustion engine. Or, perhaps the self driving feature led us to not recall which roads connected to get us to the beach. In other words, we might not understand how we got there or how the features contributed to our end result.

Instead, we could take our sibling’s Subaru WRX (ML, random forest). Or, we could take our parent’s classic 1969 VW Bus (regression). With these latter options, we may have more power to understand how the individual pieces of our trip contributed to the end result (model inference) at perhaps a loss of predictive accuracy compared to the Tesla. Finally, all the cars require some kind of fuel (clean data to run well). Without a clean fuel supply, you won’t be accomplishing your task.

Your choice of car for your trip can radically change what supplies you can take, how far you’ll get on your fuel, how fast you’ll go, and other factors. Hence, your choice of AI/ML implementation is intrinsic to your predictive modeling problem, assumptions, and business use cases.

Forecasting is a very specific type of predictive modeling. Predictive modeling is conceptually challenging, since there are one-to-many inputs, and one or more outputs. With forecasting, you’ve added an additional layer of complexity, in the dimension of time, which complicates data collection and adds model complexities like autocorrelation – data is correlated to itself through time.

Frequently, you want to predict numerous timepoints into the future, which means estimation of uncertainty around the components of your time series, such as seasonality (morning activity vs. night activity, peak season vs. non-peak season, weekdays vs. weekends) and trends (price changes and worldwide pandemics).

Responsible predictive modeling and forecasting

Even as we help our customers – and ourselves – with fiscal decisions, we are fully aware that forecasting is a complex balance of artistry, science, and intuition. It’s important to understand how to generate a prediction and determine what inputs are needed, and it is equally important to sufficiently assess the uncertainty around the prediction and evaluate the results within that context. In essence, viewing predictions as a range of numbers as opposed to a single number. 

If a business is wrong about their forecast and subsequent fiscal year planning efforts, it can significantly impact their return on investment in marketing and operational costs, which presents challenges and difficulties for shareholders and employees. If forecasting work is oversimplified or the uncertainty is insufficiently understood, then the data can be incorrectly actioned upon.

The data scientist needs to understand the power and limitations of their models, and effectively communicate the good and bad of the outputs to leaders to drive the best decisions for the company. Responsible use of models means reporting what went wrong as much as what went right.

The benefits of predictive modeling and forecasting for marketing analytics

Here at Intuit, the application of our predictive modeling and forecasting through AI/ML to delight our customers is shaped by our company values and philosophy. We have always been a customer-obsessed company. That is, we steep ourselves in deep customer empathy.

Our efforts are reflected in the connecting of customers to the right products through ML-driven personalization. It’s also reflected in the delivering of algorithmic-assisted insights for their unique business or personal financial situations, highlighting changes in their financial health, and enabling them to easily make sound fiscal decisions.

Personally, I’ve used Intuit’s Mint personal finance management service to maximize my financial health and well being. Mint suggested a new credit card based on my shopping habits. Although it had an annual fee, I enjoyed a far better return on my spending. This is exactly the type of experience I want from algorithms extrapolating on my data: Help me make better decisions based on my unique situation.

Choosing an ideal credit card is a perfect example of a classification, machine learning problem.

How Intuit envisions the future with AI and ML

The future is always unknown, even as we work hard to gather data and extrapolate outcomes through predictions, forecasting, and marketing analytics via AI and ML technology. However, there are those who have concerns about AI and ML taking jobs. If that’s you, rest easy. Why? Because computers are just not that smart.

For example, I think we can all agree that smart bulbs and smart TVs aren’t doing much “thinking.”

Computers (or algorithms driven by AI/ML) are better than humans at solving very specific types of problems, but they’re terrible at solving other more complex problems. These algorithms will increasingly be used as tools to solve components of a complex problem, handling the parts of a problem that we simply aren’t as good at solving. They will help us classify and predict in meaningful ways. But, they won’t replace our human intuition and implicit decisiveness (at least, not for a while).

This is why Intuit envisions a future that melds the best capabilities of machines and humans to deliver personalized customer experiences, all on one secure platform. And that future is AI-powered and expert-driven.

To learn more about AI and the future, check out how artificial intelligence is redefining apps, and Intuit’s Shir Meir Lador’s lessons learned leading AI teams.