What's the differences between predictive analytics and basic reporting? Predictive analytics provides insight about what will happen in the future. Basic reporting only looks at past performance. Why is this difference difficult to grasp? It's partly because transitioning to predictive analytics requires change. And most people don't embrace change.
Take an example from the energy industry. What happens when one company changes from historical reporting on electric meter fails to new reports that integrate predictive analytics? (Don't stop reading if you're not in the electricity industry. The same principles apply for pump failure in oil and gas production as well as many similar problems in other industries.)
When reviewing historical failure reports, it's natural to look for similar readings or characteristics that occurred prior to the failure, group other meters into similar looking clusters and monitor them in an attempt to detect failures faster in the future. This is a rules-based approach. The problem? Identifying one fact, like low voltage readings (X=low voltage readings) one week prior (Y=one week) to failure, ignores all the other interactions that may play an important role in meter failure.
Without predictive analytics, this means you will get a report based on a query of X + Y, and maybe another group that meets the different criteria, say A + B. These reports provide some value since they will allow you to react faster if (AND only if) one of these meters in these known groups fail because they are on your watch list. However, you're still not able to take ACTION until a meter has actually failed.
What happens, however, when X + Y + A is a group that fails more often? Your existing if-then business rule groupings have not identified this group yet. And this is just one simplified example showing why predictive analytics should be integrated into your processes.
Predictive analytics will analyze the data you have collected and "automagically" identify the factors that impact meter failure. In addition, as you add new data or as data changes over time, predictive analytics picks up and adjusts which characteristics impact either increases or decreases. As a result a predictive model takes into account changes and then produces a report of the meters most likely to fail in the future with a ranking between 0 and 100 percent.
The resulting report might look very similar to the original report, but it is distinctly more valuable. Now you can sort the meters most likely to fail in the future based on weighted patterns that take into account interactions across many readings that may not have been identified in your simple groupings and if-then queries. These new “groups” can be identified, but they will not all exhibit the same characteristics grouped by a single query.
For example, if you look at meters with a predictive propensity to fail at 80 percent or higher, you are likely looking at several different types of meters behaving differently. In other words, they will not meet a simple if-then query. If you tried to accomplish this same result with business rules, you would need to develop a bunch of different if- then queries to capture them all and then maintain and constantly update these queries.
A future-looking model will score and do this grouping for you, you just have to change your reporting process and change how you think about grouping meters. In this case, your predictive models are updated and changed based on changing data inputs and when new data is added, but your process that uses the output from these models and the propensity scores do not change.
The next step is to add in more predictive analytics like survival analysis. In addition to knowing your propensity to fail in the future, survival analysis will tell you the likelihood of WHEN this failure will occur.
With these changes, your predictive analytics will allow you to dispatch crews to fix or replace high value meters or pumps prior to them failing. It is no longer acceptable to simply automate a process. You must convince others in your organization about the benefits of "auto-magicing" your processes.
Now, let's look at how this change might impact other strategic decisions. Let’s say your if-then grouping identified a particular meter from a specific vendor that has a higher failure rate than the same type of meter from other vendors. This may lead you to stop doing business with this vendor or could lead you to decide to dump all this inventory (at a loss). However, using predictive analytics helps you to identify that the root cause of the failure rate isn’t due to a particular vendor. Instead, the fail rate is uniform across all your vendors when their meters are used in the northeast. The reason it looked like you had a problem with a particular vendor was that this vendor’s meters were installed in the northeast at a highly disproportional rate compared to your other vendors.
In this case, the use of predictive analytics could help identify the combination of factors, including very cold weather (regardless of vendor). No need to stop doing business or to dump inventory. Instead you need to advise all your vendors for the need of a different meter to be used in the northeast and you can re-distribute your existing inventory of meters across other regions in which they don’t exhibit the failure rate in the northeast.