Why ArrowModel?

Agile Scoring



Why ArrowModel?



Why model?

The world is a competitive place. Most businesses who can afford it, develop sophisticated models to solve their most important problems.

And business problems are never exactly the same, so you need to develop custom models, models that really apply to your data and to your specific problem.

Finally, the environment is always changing, so your models need to be updated regularly. If you want a consultant to help, don't hesitate to call us. But if you don't, ArrowModel can help you do the job, all by yourself, for a fraction of the cost.

Why use ArrowModel?

There are many ways to do scoring. You could even use paper and pencil, at least for a teaching exercise. We describe below the main alternative ways of doing scoring that we have encountered in the last 20 years of consulting for Fortune 500 companies.

1 Alternative: General Purpose Statistical Packages

The most common scenario consists of using a general purpose statistical package such as SAS/STAT, SPSS or R, often in conjunction with macros to automate repetitive tasks.

While extremely powerful and very flexible, this approach is rather time-consuming, requires good knowledge of a typically obscure proprietary programming language, and does not scale very well. Trying to make sense of somebody else's models (or even your own model, from a few months ago) is a challenge, as there are often multiple individual programs and datasets.

In contrast to that, ArrowModel keeps everything related to a model in one file, providing a uniform structure around scoring projects. Unlike R, ArrowModel does not require all the data to fit in main memory, and can work with very large datasets. You are still limited by disk space and address space, but the only realistic constraint is time.

With ArrowModel, a scorecard can be built without writing a single line of code. In cases where you need to transform the input data, it is done through SQL, which is widely known.

2 Alternative: Everything but the kitchen sink

Expensive solutions like KXEN Analytic Framework, Intelligent Results PREDIGY Platform or SAS Enterprise Miner are targeted primarily at large organizations. Their client-server architecture requires infrastructure to be in place and working.

In contrast to that, ArrowModel is trying to do only one thing, and to do it well. A lot of thought has gone into deciding what not to include. For example, ArrowModel is built around logistic regression and does not include the "technique du jour," like support vector machines, random forests, or multilevel perceptrons.

The simple ArrowModel user interface allows the analyst to concentrate on fitting the model as opposed to trying out all the available techniques. This is in line with Frank Harrell's notion that "carefully fitting an improper model is better than badly fitting (and overfitting) a well-chosen one" [1]. For most practical classification tasks, logistic regression is as good as, if not better, than other models [2].

ArrowModel runs on your desktop and does not require an internet/intranet connection. Once the raw data is on your local (or shared) drive, you can start working.

3 Alternative: Homegrown Software Solutions

Very few businesses rely on internally developed custom scoring solutions. Unless scoring is the company's main area of expertise, such solutions are likely to be a patch made of Access databases, Excel spreadsheets, VBA components, and scripts holding the whole thing more or less together.

ArrowModel is written in C++, using modern software development tools and techniques. It runs natively on all supported platforms, giving it a significant performance advantage over custom software often written in interpreted languages. As a commercial product, ArrowModel undergoes extensive testing and validation.

Bottom Line

As you can see, in most practical situations ArrowModel allows you to build, test, deploy, document, and monitor predictive models easier and faster than any of the alternatives. See ArrowModel in action...

References

[1] Frank E. Harrell, Jr. Regression Modeling Strategies. Springer, 2001

[2] Paul Komarek. Logistic regression for fast, accurate, and parameter free data mining.