What is Predictive Analytics?

Scott King:

Welcome, everyone, to the latest episode of the Union Podcast. I’m Scott King, joined by Chris Kraus. Today, we’re focusing on predictive analytics. Here’s a description I found online that I think defines it well: predictive analytics is the process of using data to forecast future outcomes. It involves data analysis, machine learning, AI, and statistical models to identify patterns that could indicate future behavior. These insights help organizations forecast trends and make informed decisions. Now, Chris, how does predictive analytics intersect with AI and machine learning?

Chris Kraus:

In a previous podcast, we discussed that AI, in the broader context, includes machine learning as one of its categories, alongside natural language processing and natural language generation. Machine learning, in essence, is mathematical equations used for calculations. For instance, I used to use SAS software to input data, perform calculations, and then analyze the results. We had tools like linear regression and multivariate analysis. Machine learning has expanded upon these traditional tools, introducing new concepts like neural networks. We now have a wider range of mathematical algorithms available than before. Additionally, we’ve moved towards unstructured learning where we can input a variety of information and let the machine learning model tell us about it. These new tools, powered by neural networks, add to our existing repertoire of statistics and analytics. All these techniques are forms of machine learning based on mathematical algorithms.

Scott King:

So, it’s all to assist us in making better decisions. Now, considering the vast amounts of data that big enterprises produce, how is this data prepared for use in predictive analytics? What kind of data is typically used? Can you elaborate on how we would input this into a model?

Chris Kraus:

Often, our customers have a specific goal in mind, such as predicting cash flow or sales. For instance, a B2C business might want to predict sales for the upcoming Christmas season based on historical data. Insurance companies might need to predict potential loss based on hurricane season, factoring in the number of predicted hurricanes and the cost to insure and rebuild houses. They know their desired outputs and have a clear idea of the inputs. Except for cases involving neural networks, which are used for finding interesting insights, people typically understand the contributing factors to sales or other key performance indicators in their organizations. When considering automation, it points us naturally to the necessary data. For example, in a call center, we might want to predict the value of a customer based on their profile, including the number of devices they have and how often they purchase extra products. We might want to determine whether to offer them a discount based on their value. We would use data from the CRM and sales systems for these predictions. We already have access to this data; we just need to learn how to process it.

Scott King:

I’m currently in a situation with T-Mobile where they want me to change my auto pay from a credit card to a debit card to maintain my auto pay discount. It appears they’re looking to save on processing fees. Someone over there has likely created a model predicting their potential savings. Speaking of models and data, how does one build these predictive models?

Chris Kraus:

What has really changed is who creates these models. Traditionally, this work was done manually by math or statistics majors who applied specific algorithms and processed the data. Now, we have very effective machine learning models for both prediction and categorization tasks. There are numerous excellent models available. At Krista, we argue that a data scientist isn’t needed to understand the nuances of these various models. We use machine learning to build machine learning models. We might take 80% of the data to train a model, then use the remaining 20% to test it. We try different models, just using CPU cycles at that point, and then report on which model produces the best score.

Scott King:

Who would know that? You mentioned using machine learning to build machine learning models. What’s the intended audience for that? I can’t imagine many people looking for such a solution.

Chris Kraus:

Many people aren’t aware it’s available. They’re comfortable with tools like Excel’s pivot tables. Excel power users who use pivot tables and perform calculations are the same people who can use these machine learning models because we’ve lowered the technical barrier. The data that you use in your pivot tables is the same data you’d give to a machine-learning model. If something isn’t statistically important, it won’t contribute significantly to the model, and you don’t have to do that work. Business people often don’t realize that they can input their data from their business processes directly into the model and let it determine what’s important.

Every time you make a prediction, it represents a decision point. Rather than having to email someone or hold a meeting, we can automate these decision points and direct the automation down different paths to complete tasks and arrive at outcomes. Machine learning applied to processes can make confident decisions, either automatically continuing the process or notifying someone to confirm the decision. Let’s consider a vacation request in a warehouse. The automation can check how much leave you have, look into the forecasting engine to predict how many packages will ship that week, and consult time sheets to see how many people are on shift. Normally, a manager would make a decision based on this information. But now, we can give this data to a machine-learning model and let it decide.

Scott King:

They probably don’t know software could do that.

Chris Kraus:

Exactly. And once a decision is made, the steps that follow, such as entering the vacation into a system like Workday or marking it on a calendar, can all be automated.

Scott King:

Can you share a case where predictive analytics have improved decision making and invoked the next step?

Chris Kraus:

Sure. In finance, for example, there’s an interesting use case. Accounting might tell you that we’ve sent out 100 invoices totaling $500,000 or $5 million. But they don’t know when these will be paid. To predict cash flow, you need to anticipate the behavior of the people paying the invoices. The timing may vary based on the value of the invoice. You wouldn’t be able to do this manually for 300 vendors. Machine learning can adjust this and predict for each vendor and invoice price when you can expect to be paid based on past behavior. This allows finance to predict cash flow based on outstanding invoices. When will they be paid? What’s the cash flow for next Friday or two weeks out?

Scott King:

So it’s probably someone in accounting saying, “Oh yeah, that’s my friend Jeff. He always pays late and sends me a cute little note every time he does.”

Chris Kraus:

How does a human remember the payment behaviors of 300 or 2000 vendors? When it comes to multinational companies with thousands of vendors, machine learning models can handle the volume that exceeds a human’s ability to keep track.

Scott King:

Especially considering the turnover in any organization, using analytics would benefit everyone. So, what’s the future of predictive analytics?

Chris Kraus:

People will realize the power of making decisions faster with the help of predictive analytics. They can then ask, what’s the next step in the process to automate? For instance, if in finance we predict that we’ll receive a certain amount of payments in the next few weeks, can we use this information to nudge or entice people to pay early? We can then automate some actions based on these predictions.

Scott King:

Everyone’s looking into AI and better decision-making because we have tons of data. So thanks, Chris, and thanks to everyone for listening about predictive analytics. Be sure and subscribe to the union on YouTube or any of your favorite podcast players. And until next time.