Cape Town, 06 Mar 2015
There is now general consensus that we are moving towards a future in which predictive analytics is pervasive. These applications will most likely be ubiquitous and invisible, like the microchips powering today's vacuum cleaners and refrigerators.
Doctors, instead of reading an ECG printout, are likely to refer to a computer with underlying analytical algorithms to make a diagnosis. Sports coaches will refer to a model based on biological readouts from sensors on an athlete's body to assess their fitness and risk of injury. New examples are coming up almost weekly: identification of false alarms in the security industry, increasingly sophisticated traffic advisories, localised weather predictions, use of telemetry data or insurance, and many other applications, intelligent automobile and homes - the list seems endless.
Despite this bright outlook, it remains clear that we have some way to go in unlocking the full potential of predictive analytics. Currently, we seem preoccupied with the traditional application areas: credit risk and scorecard development, fraud detection, churn prediction and next best offer modelling. Although the present high cost of predictive analytics expertise and technology has led to an 'enterprise only' vision for predictive analytics, in the end it could be a very different reality, much like the lowly vacuum cleaner software of today.
But perhaps more significant than the cost is the obstacle presented by the technology itself. While large software houses have produced many well-crafted data mining workbenches, there is often a significant gap between the data mining workbench and usable business solutions. Closing the gap between the workbench and the final business solution requires substantial input by a range of specialists whose skills are in short supply.
An illustration of this kind of problem can be found in the field of fisheries management. For around 30 years, the field has been dominated by applied mathematicians and mathematical statisticians. Even today. a fisheries management meeting feels more like a mathematical statistics conference than a meeting about fisheries science and biology. While the underling population models are clearly useful, they are also complex and arcane. Understanding their inner workings requires a postgraduate degree in applied mathematics and computer science.
Recognising this problem, developers went to work to produce a wrapper for fish stock models which would simplify the application of population modelling in fisheries management. The result today is a program called Stock Synthesis, which uses ADMB and R models to produce serviceable modelling and diagnostics of fish populations in a host of applications globally. One no longer needs to be a genius in both C++ and mathematical statistics to explore fish stock management options. Rather, the software is used to define the basic parameters of the problem and the input data; tasks that do not require skills beyond those of a fisheries biologist.
The generic and expanding reach of Stock Synthesis has significant benefits for fisheries management. Rather than having hundreds of people around the world rethinking the logic of a fish stock model dozens of times a year, it has been done, once. This creates time for discussion at meetings to focus on fisheries biology itself, rather than mathematics and computer science, or high-level statistics. In fact, it is the only way that many countries will ever be able to access high-end tools for quantitative fisheries management. The important step and vision of developing Stock Synthesis may therefore prove to be a pivotal point in fisheries science and management.
The same logic applies to data mining and predictive analytics. Business people aren't usually interested in the complexities of modern data mining algorithms such as the intricacies of the multi-layer perceptron or radial basis function neural networks, or the marvels of support vector machines. Rather, they seek real solutions to real problems. Telecommunications companies faced with a high turnover of clients want the ability to focus their marketing efforts and improve how they treat their customers. Insurance companies want the ability to triage their claims to focus their assessor resources effectively, or source vehicle repairs from suppliers who give value for money and keep their clients satisfied. If possible, they would like to have these solutions without having to hire a team of mathematics professors to do the job.
As a result of the difficulties in implementing cost-effective predictive analytic solutions, an unhealthy mismatch is developing. There are, for example, many users of data mining technology who aren't actually using it for predictive analytics. Instead they are using it for ETL, data pre-processing and reporting. On the other hand, there are many companies who don't have any data mining technology who are in urgent need of predictive analytical solutions for claim and supplier segmentation, credit risk modelling, fraud detection and next best offer models.
In spite of the advancements in data mining tools, the reality is that real business users are in need of a suite of adaptors which fill the gap between complex analytical technologies and the real world of applications and solutions. Companies need a churn adaptor, a credit risk score card adaptor, an insurance claim segmentation adaptor, and a host of other adaptors. Solutions analogous to Stock Synthesis would ensure that we do not need to relearn and redevelop each of these solutions from scratch thousands of times around the world, hundreds of times a year. Rather, it is done once, properly. Once the adaptor is set up, all the relevant mathematical and statistical intricacies should be automatically taken care of. All that is left is for the user to define the parameters of the problem. The adaptor might not suffice permanently, but at worst, a Mark I adaptor would become a foundation from which its generic reach can be extended with Mark II, III and beyond.
These adaptors will be invaluable to the world of data, and should capture the best-of-breed approaches and predictive analytical models. As a leading company offering predictive analytics skills and solutions, OLRAC SPS has embarked on a journey to develop suitable adaptors on a case-by-case basis. With time, we believe this is the direction in which the world of predictive analytics is going, and we aim to be at the forefront when it gets there.
Share