Sybase IQ

« Guest Blog: Introduction to Fuzzy Logix and In-database Analytics | Main | Leading Indicators »

April 21, 2010

Guest Blog: Business Benefits of In-Database Analytics

During a recent blog entry Introduction to Fuzzy Logix and In-database analytics, I wrote about high performance analytics and how they can be pervasively deployed. In this entry, we’ll highlight some of the business benefits of in-database analytics with examples from recent client engagements. All of our examples share a few common themes. Each company benefited from putting analytics into the hands of businesspeople and each did so at multiple levels in their organizations. They also leveraged advanced analytics to find previously hidden patterns of behavior that led to changes in their business practices and all witnessed significant business benefits. Let’s look at a few examples that include forecasting, customer targeting and cost of sale reduction, and provider scoring and fraud for insurance.

First let’s look at forecasting. Usually, forecasting is a process driven by sales people and management or perhaps, pushed down via finance. Either from the field up or from the executive suite down, there is always some tweaking along the way to the final ”guestimate.” While these processes are common and indeed necessary (because they capture the viewpoint and experience of the business), they are often time consuming and based on many layers of subjectivity. It can help a great deal to pair a subjective forecast with a quantitative forecast based on transaction data. Think of it as a ”check and balance” for the process. While the idea is sensible, in practice it can be challenging.

A good example is deploying a forecasting system for one of the largest consumer packaged goods manufacturers in America. They have hundreds of SKU’s, over 500,000 sales locations and 18 months of transaction data; about 2 Billion rows in total. Add a matrix management structure and the requirement to allow each manager to forecast their area of responsibility plus related areas and it became clear that a traditional approach to analytics would not work. If we tried to use their already overworked statisticians, who had to move many set of data and then run individual models, we’d never finish in time to deliver the forecast and we certainly wouldn’t be able to make this part of their weekly operational process.

Our innovative solution was to leverage the power of in-database analytics and deploy the forecast via their existing reporting environment. Now managers simply select from a list of drop down parameters, for example they can choose to forecast all sales of a particular brand (or group of brands) in a state then see the expected changes in demand from worst decline to highest improvement. They can even click on button and have the report sent via email to the person who can address the issues. This allows our client to focus their teams on key areas of future risks and start to address issues before they happen. Now their sales teams work with distributors and customers to address the root causes of loss of future sales and in many cases their actions can reduce or eliminate the lost sales. There are many other benefits, such as decreasing variability in their supply chain and cash flow, but we’ll leave that article for later.

As for customer targeting, it can take many forms. Many of our clients want to understand their most profitable customers and then target that part of their business for growth. Many want to personalize ads for each customer and prompt them for their next likely purchase or reduce customer churn. Let’s use an example that combines reducing churn with stimulating next likely purchase. Based on transaction, customer and demographic data, we can use in-database analytics to understand the patterns of behavior that indicate that a customer has a high propensity to attrite. Once we know who has is likely to leave, we can target them with a number of programs including suggesting their next likely purchase. For one of our clients we did just that and saw a 10% reduction in churn. Additionally, once we developed our model for churn, we built a tool for marketing to understand the churn adjusted cost of sale for each marketing program. While this example was for a company with a large call center, we are also working with a telecom company to combine social networking analysis with the programs mentioned above to ensure that those with a high propensity to churn, who are also influencers of behavior, receive special attention because losing an influencer can have a painful add-on effect as those in their network also attrite.

Finally, let’s look at provider scoring and fraud in healthcare. One of the largest insurers in America wanted to score the quality and efficiency of care given by their providers. The process took 6 weeks and required the IT team to run the legacy programs 25 times. Since the scoring was based on patterns in claims data, the volume was simply too large to process it all in one run with their row-based RDBMS and traditional analytics software. This meant they need to move the data from their EDW to an analysis server for each run. Using in-database models, we were able to reduce the time needed to process the scoring models to less than 2 hours (and one model run). So the while the difference in time was 6 weeks to 2 hours, the effect on risk management was dramatic. Now the company could evaluate the doctors giving care to their customers on a weekly basis (or on demand) and immediately identify situations when the care provided was not up to standard.

We are now working on a project to find hidden fraud and expect to see benefits in both speed of processing and fraud prevention. Our client reviews deviation from the norm in charges for services. They analyze 10 practice areas (cardio, OB/GYN, etc.) on a state-by-state basis. We wanted to know why they didn’t just run all practice areas and combinations of charges and identify the areas with the highest potential risk. We also wanted to know why they were only looking at standard deviation and were not using clustering (segmentation) or market basket analysis to find the hidden deviations in the care given to patients. The reason was that they couldn’t build the models without buying expensive and complex software and also didn’t have the processing power. Another issue was that they wanted to put the analysis tools in the hands of their employees in the special investigative unit (SIU) so they could analyze the data and act quickly. To satisfy their needs, we are building a custom set of models for their end-users. The models will deployed will include types of outlier analysis, market basket analysis and classification models (such as multiclass support vector machines). Soon, investigators will simply run the models and then look at the cases of highest likely fraud. They will also have the flexibility to drill down and see the reports based on geography, practice area or even episodes, treatments and practices. We haven’t fully quantified the value as the project is on-going, but I expect this to easily produce over $1 million in benefits to our client.

Hopefully these examples will spur some thoughts about how in-database analytics can be deployed in your company to drive business benefits. We are finding that our clients who leverage these solutions see a return on investment of 10X or more. In fact, one recent client saw a 32X return on investment in 12 months. That’s good proof that companies who deploy analytics will see dramatic results. In our next installment, we’ll dig deeper into the technical reasons that these benefits are larger with in-database analytics than with traditional solutions.

Posted by Sybase IQ at April 21, 2010 3:35 PM