How to Ensure Success When It Comes to Predictive Analytics

Predictive analytics, for all its benefits, can be overwhelming, especially for those who up until now had only been familiar with traditional BI. But predictive analytics has proven to be a revolutionary technology for many industries, and companies who have not yet made the leap to predictive analytics are ready to reap the same benefits that many of their peers have already achieved. However, before they can start introducing predictive analytics, these companies need to have a firm understanding of what predictive analytics is.

What is predictive analytics?

Predictive analytics is a way of using data to make predictions about future events. Data mining, statistics, and modeling can all be used to better inform the predictive model so it can more accurately analyze current data, thereby making more accurate predictions about future events. There are enormous benefits to using predictive analytics, but successfully deploying them can be a bit of a challenge. Recently, I was interviewed by John Edwards at CIO for an article about this exact topic, and about the steps that companies can take to help ensure success when they’re ready to start introducing predictive analytics. While John explores seven of these steps in his article, I’m emphasizing the four that I find most critical.

1. Build a solid data foundation

Before any attempt is made to seriously introduce predictive analytics, companies need to first address their data foundation. Analytics is only as good as the data, so it’s absolutely vital to take the time to do this legwork in advance. What kind of data do you have? What sources do you use? Is it well-structured? Is it clean and accessible? Is the data harmonized, timely, or even accurate? Is data among multiple sources integrated in a good way? In order for predictive analytics to work the way expected, the algorithms need access to structured, clean, and accurate data that is up-to-date.

Additionally, it is essential to have both business and subject matter expertise about the data in question. If you’re looking at IoT data, is there a technician who understands what all the different technical data streams of switches and alerts mean? If this is clickstream marketing data, is there an analyst who understands all the user interactions, marketing campaigns, and features that have been logged? Well-developed data science is designed with validity and hypotheses in mind. The team has to understand the data to both develop accurate hypotheses and models for face validity.

2. Use analytics for the tasks most likely to deliver positive ROI

Analytics will make the biggest, most measurable difference when used to improve specific business processes. Companies who eagerly—and perhaps haphazardly—start applying analytics to mountains of data won’t find themselves experiencing any tangible results. Instead, focus your efforts on specific, manageable tasks that will make a meaningful difference. Start using predictive analytics for something small and quantifiable, like automatically reordering of supplies when stocks are low, and achieve success there, then move on to something bigger.

Level of effort also plays a key role in understanding ROI. Sexy, big-vision projects may create initial excitement, but if they require ingestion and integration of dozens of data sources and complex, ensemble models, and if they are a challenge to productionalize, then the business will quickly lose appetite as resources are poured into a vision that was scoped much too large.

3. Define the right approach

There’s a seemingly unlimited number of methods and approaches that can be used to generate accurate predictive analytics, and new methodologies are constantly being introduced. This flexibility is one of the great things about predictive analytics, but it can make it hard to figure out the best approach for you. That’s why it’s so important that you define your business outcome first, so you can choose an approach that aligns with your overall goals. Do you want to understand whether a customer who purchases one product is more likely to purchase a second product so you can determine if you should bundle those products together? Do you want to target your marketing efforts to a specific segment of customers so you can increase revenue? Understanding the core business outcome and the concomitant existing problems that cause pain creates context for the necessary solution. Only then can a data scientist evaluate whether the data is present to support predictive analytics that solves for the problem, or can suggest additional data sources or an entirely different approach to solving the problem. When the business outcomes are not scoped well, then you run the risk of solving the wrong problem even though you’re using the exact same data.

4. Be aware of human biases present in models

It’s tempting to implicitly accept what the model tells you, but doing so ignores the implicit bias that’s inherently built into everything that we, as humans, create. Our CEO Roman Stanek recently wrote about this exact topic, and about how unrealistic it is to assume that AI is bias-free purely because the algorithm itself is not human. But as we humans build the models that enable predictive analytics, we teach those models to respond to our biases in the same way that we respond to them. And I’m not talking about the racial or gender biases that we hear about so often, but less well-known biases like the backfire effect or anchoring, which can be just as damaging.

Enterprises are going through growing pains and learning that predictive analytics isn't something you can dabble in. However, the impact that strong predictive analytics can have on business efficiency, revenue, and product performance is well worth the time, energy, and resources that are needed to ensure success.