Complexity

One can argue that the essence of business is to provide customers with simple solutions to complex problems. Complexity can be seen in anthills, human brains, and stock markets, but complexity also abounds in the business environment. From engineering to marketing, sales and customer support, business is governed by complexity; and data mining is an essential tool for supporting such complex business functions. Complex systems exhibit behavior that cannot be explained in simple ways, and that makes it hard to manage business systems. The most difficult support needs are those that involve making predictions. For example, how do you predict sales, product ship dates, marketing response rates, or service quality? Let's look at predicting customer support escalations as an example, and explore how data mining can increase the profitability of a customer support business function.

Increasingly, service processes are highly automated and a very large number of cases are handled simultaneously. Consider a complex service business where thousands or millions of cases may be active throughout a year, a day, or even at any moment. Most cases are handled within established boundaries but there are notable exceptions. In exceptional but not so rare cases, a service case escalates. By definition, these escalations are marked by customer pain, and they frequently require executive involvement, and they may cause excessive costs. How can this complex business function be supported to prevent escalations? One approach is to fix the complex system to remove the possibility of escalations. A complementary approach, and this is the approach I discuss here, is to predict escalation risk for each case, and to proactively engage in high-risk situations to forestall escalations. It is possible to predict the escalation risk of each case with so-called predictive analytics solutions.

Think of a recent escalation you have been involved in. In my experience, a moderate proactive investment could have avoided any specific catastrophic escalation. The problem is that if you handle thousands or millions of cases you cannot give each one any significant incremental attention. You need to identify cases in need of attention with laser-like focus. It can be profitable to establish proactive processes which engage only for identified high-risk situations. But how can you determine whether that proactive process is profitable? We're in luck here, because the situation is really quite simple. Profitability depends on only three quantities: prevention cost, escalation cost, and prediction precision.

Some examples will help to confirm this. Say prevention cost is $5,000 for engaging a customer proactively, escalation cost is $80,000 due to penalties, and prediction precision is 35%, then you will make an average profit of $23,000 on each case you address proactively: (0.35 * ($80,000 - $5,000)) - (0.65 * $5,000) = $23,000. At a precision of 10% your profit plummets to $3,000, and below 7% your profit turns negative because you spend more on prevention than you save on prevented escalations. Similarly, if prevention cost is $10 and escalation cost is $75 then your break-even point for predicition precision will be at 13%.

I hope you found this overview of data mining as a tool for complex business problems useful. If you want to learn more you may be interested in a research paper I coauthored on maximizing the ROI of predictive analytics: "Predicting Software Escalations with Maximum ROI" by Charles X. Ling, Shengli Sheng, Tilmann Bruckhaus, and Nazim H. Madhavji. This paper has been accepted for publication at The Fifth IEEE International Conference on Data Mining, Sponsored by the IEEE Computer Society, Houston, Texas, USA, 27 - 30 November 2005. I plan to post again soon about the nasty property of "high dimensionality" and how data mining can help you deal with it.

Posted at 09:21AM Oct 07, 2005 by tilmannsblog in The Predictive Business  |  Comments[0]