So last week when I wrote about Predictive Analytics, I got responses from folks saying, “The value from such areas is clearly there. But the challenge is which technique to use and the ever-sliding sword of showing ROI so there is buy in for these analyses”.
In framing the Analytics problem – we need to balance data, SME knowledge, and performance. One of the things I have noticed in my work is when the analysts build models the real skill in creating effective analytic model is knowing which models and algorithms to use. They can use different techniques: neural networks, decision trees, linear regression, naïve Bayes, etc. But these days many analytic workbenches now automatically apply multiple models to a problem to find the combination that works best. One needs to explore different paths – they look at the problem from different perspectives. When these algorithms are combined there is resulting synergy. Once the modeling data sets were finalized, the largest incremental gain was not achieved by fine tuning the training parameters of an individual algorithm, but by combining predictions from multiple algorithms.
So with the myriad tools and techniques that exist, the way to approach this is to ask the questions that are really important for what the company is trying to solve:
- Strategic Customer Questions
- Who are the most/least profitable customers?
- Who are the most/least satisfied customers?
- What is fastest/slowest customer segment?
- What are the reasons for customer attrition?
- What are the costs of customer transactions?
- Strategic Product Questions
- What are our most/least profitable products?
- What are our production costs & how can we lower them?
- What is our cycle time & how can we lower it?
- Strategic Employee Questions
- Who are the most productive salespeople, employee?
- Which managers have the highest retention rates? What do they do?
- What is the cost of turnover?
- Strategic Financial Questions
- How accurate are the financial forecasts?
- How much financial data is used to answer business decisions?
- What impacts the demand of our product?
- What items are affecting our margins the most?
Based on this one has to look at some of the following techniques:
- Classification – predicting an item class, “Decision Tree”
- Association – what occurs together, “Market Basket”
- Estimation and Time Series – predicting a continuous value
- Web and Text Mining – extracting information from unstructured data
- Clustering – finding natural clusters or groups in data
- Deviation Detection – finding changes or outliers
- Link Analysis – finding relationships