In a previous post, we discussed how beginning with a proof of concept can be the best way for an organization to get started with predictive analytics. However, if you’ve never worked with predictive analytics before, it may be challenging to design and execute your proof of concept.
Using an example based on a recent 110 Consulting project, here are the key steps to achieving a successful predictive analytics proof of concept.
Determine Your Goals and Objectives
First, you need to start by defining the problem you’re trying to solve. In this case, the client offered a suite of software products and needed a way to predict which customers were most likely to purchase new products or upgrade existing versions so they could strategically tailor their customer communications and marketing efforts, as well as improve their sales forecasting.
We decided to focus the scope on predicting which customers were likely to purchase one new software product based on their existing purchase history and overall profile. Keeping the scope of your proof of concept limited is important, as it will also help limit the costs and overall timeline. The return on investment for our client would be a clear demonstration of the impact that predictive analytics could have on their overall sales strategy. We also worked with the client to define high-level next steps for all possible outcomes, from needing to refine and re-run the proof of concept to developing a production solution.
Build Your Proof of Concept Team
It’s important to include an executive sponsor and project owner from the client side to make sure the proof of concept has both upper-level support and a regular point of contact. This project already had an executive sponsor at our client’s company, and they chose a project owner to work closely with our team. Our team was small, consisting of just a project manager, business analyst, and developer, in keeping with the limited scope of the proof of concept. Previously a data scientist might have been required to build a predictive analytics model and apply it to the client’s data, but as predictive analytics become more mainstream, business analysts are becoming more frequent creators and users of predictive analytics models.
Our project manager worked with the client project owner to identify the source system data stewards and stakeholders as well as the target audience. In this case, we would be working with data from the client’s Microsoft Dynamics CRM system that tracked information such as company size, industry, previous sales history, and current technology platforms. Data stakeholders included the partner and account managers responsible for updating CRM data, as well as senior-level sales management and the CRM system administrators who also served as data stewards. The target audience was the partner and account managers, sales management, and executive leadership.
Obtain Sample Data Set
The next step was obtaining our sample data set, although this can also be done while building your team. To reduce costs and timeline, we chose to use existing data from the client’s CRM system. We decided to use historical data about a previous software launch and those purchase records, along with the customer profile data, so we could validate our output against real sales results. Our analyst and developer worked with the client’s CRM system administrators to pull data and perform analysis in preparation for integrating data with our model.
Select the Predictive Analytics Model
We wanted to identify the strength of the variables that influenced a customer’s purchase decision to improve the targeting of individual sales efforts as well as overall forecasting and planning, so we chose a decision tree model. Our input included the customer’s existing software portfolio, company size, industry type, and sales history including length of time since initial purchase, discounts offered, etc. at the time of the previous software release. Our business analyst and developer designed the decision tree and integrated that data to break it out into nodes that could show us the different path customers had taken to their purchase decision.
Proof of Concept Results
Once completed, our business analyst could review the model and determine which variables and combinations had most influenced purchase decisions. For instance, they were able to see that a customer who had purchased their first product from our client over 5 years ago and had more than 100 employees was extremely likely to purchase the new product regardless of price, but more recent customers were more influenced by discounts or targeted marketing offers. It was harder for a casual user to interpret these results directly from the model, so we designed a simple web interface that allowed our target audience to view the data via sample visualizations and reports. We were also able to compare the results for the entire model against historical data and validate our model with a very low margin of error.
Stakeholder Review and Next Steps
We reviewed our proof of concept results with our target audience, including sales management, partner and account managers, and executive leadership. They were able to use the reports and visualizations to filter customers into segments based on the factors influencing their purchase decisions, and then plan potential campaigns and forecast overall sales accordingly.
Our executive sponsor, client project owner, and data stewards met with our team to review the results and discuss the ROI. They agreed that using their existing data to highlight the variables most likely to influence their customers’ purchases as well as forecast overall sales would streamline and increase the effectiveness of their sales efforts. The client chose to take the next step of reviewing the level of effort for a production version, as well as discussing ways to expand the model’s range and complexity by including more sales scenarios and additional variables.
We were able to successfully prove the positive impact of predictive analytics to our client by executing on a limited scope with a small, focused team, clear objectives and the right predictive analytics model for our data and deliverables. On the client side, having executive support and collaborating with their data stakeholders and stewards ensured that our proof of concept could easily expand to a production solution.