In Part 1 of this blog post, we discussed how real-time analytics are used to analyze past performance. Prediction involves taking the insight you get by looking at historical data and working out what's going to happen over time. Instead of just looking at whether or not something that's currently happening is normal, we can take a peek hours or even days into the future.
There are different approaches to prediction, but they all involve machine learning, predictive modeling, or predictive analytics. Even though there are many different algorithms, they all look at the history of your system to detect patterns and correlate different types of data together. Let's refer back to our example from Part 1: if we know the pattern of transaction rates and decline rates during the course of the day, we can begin to predict the future based on what we already know about the past.
You first need to train these systems for a while by letting them watch your data. From there, these algorithms can work out the likely system performance based on the history.
Predicting ATM Cash Levels
To see how prediction can work in a real-world scenario, let's look at the cash that's sitting in an ATM. When will the cash be depleted? It really depends on how much cash was put into the machine in the first place. Banks are engaged in a cycle of trying to reduce the cost of cash sitting at the ATM. It costs money to refill the ATM, but it also costs money to have cash sitting in the ATM.
There's a fine balance between how much cash that's left in the ATM and how often it's refilled. The worst possible result would be to run out of cash and stop providing service to customers. Depending on the ATM's location, the rate of cash leaving the machine can vary quite a bit. The goal is to optimize the level of cash placed in the ATM. For large banks that have upwards of 2,000 ATMs, the cost of cash can actually make a big difference. That's why optimizing replenishment schedules is an area that we're investigating closely.
Preventing Future Problems
Let's take our example about normal ATM behavior a bit further. If you know the normal transaction rate at a particular time of day, a higher than normal rate isn't worthy of an alert in and of itself. However, if it's during the early morning and we know that peak time is coming in two hours, it's growing at a higher than normal rate and requires you to take action. We can raise an alert using a prediction model instead of just looking at what's happening now. If we can accurately predict that that the ATM will be depleted, we'll be able to prevent a potential problem from occurring.
How far back do we need to go with historical data to get meaningful results? The minimum time depends on the prediction scenario, but more data is always better. Additionally, the longer that you train the system, the better predictions it will make.
I'll also be discussing the importance of real-time analytics in an upcoming webinar, on April 27, you can register here. If the timezone isn't appropriate, by registering you will be updated on when the on-demand version is released, I hope you can join us.