|
Techniques
Modeling Building
A significant amount of data preparation occurs long before any modeling technique is applied. After the information value possibilities are fully extracted from the various data sources, a specific modeling technique can be employed. It is highly unusual for one technique or another to yield a significantly different result at this point in the analysis primarily because every technique is almost wholly dependent on the data inputs for the most accurate predictions.
A variety of techniques are used, including pattern recognition using a maximum likelihood estimator (extended version of logistic regression) and neural networks, but the preferred modeling tool is logistic regression because it allows the user to shape the models to meet palatability constraints. This is especially useful in ensuring the models remain intuitive to the end user as well as provide for a robust outcome-over-fitting to the data is prevented. These models are particularly suited for creating reason codes to assist in the declination of credit and the operational effectiveness of investigative personnel.
These scorecard building tools are designed to take multiple data inputs in a hierarchical manner. Specifically, they allow the creation of multi-stage, sequential models that incorporate different families of predictive characteristics. This capability is particularly helpful when designing a solution to fit a certain operational or cost constraint. For example, in transactional fraud scoring, a model segment summarizing the information value of account and personal data (available at day's end) eliminates the need to carry more than that stage's score into the authorization system. The score can then be combined with recent authorization activity to create a final prediction. As well, additional stages in the model can be fired post-authorization in order to incorporate data for operational decisions that would otherwise be inaccessible due to latency concerns. Effectively, a multi-stage scoring approach allows for maximum model strength while reducing the computational intensity in the in-line authorization process and ensuring expedited but precise decision support.
These scorecard building tools do not suffer from variable-set size constraints, which allows for stronger models given that all variables are always under consideration. Traditional modeling vendors, when the number of candidate variables is very large, typically use a model building process that involves a preliminary filtering procedure. This step is usually a very simple test that measures the individual "strength" of each candidate variable as a predictor, and keeps the "x" strongest variables in order to minimize the computational intensity of modeling building. Such an approach does not take into account any correlation of variables, and practically identical variables will be preferred over others, regardless of the fact that including one renders the others insignificant.
All of these features would not be possible if the model building technique was not an extended version of standard logistic regression. Besides having the ability to use linear constraints, the modified stepwise approach ensures that the algorithm cycles out and cycles in variables so that relationships are fully mined. Standard stepwise techniques suffer from variable ordering issues. A variable could be excluded from consideration (because of statistical insignificance) if introduced at the wrong stage of model development. This is also related to multicolinearity which is handled in the variable selection process.
Finally, these logistic regression scorecard building tools can be deployed in software in order to fully automate the model building cycle. This automation can be run at regular intervals, typically monthly or quarterly. The cycle can be supervised, if desired, and a fully qualified expert can validate the final models before they are put into production. The advantage of an automated modeling cycle is to minimize the reaction time between emerging trends and the appropriate predictions to mitigate their impact.
Strategy Optimization
Other tools that can be used to maximize the benefit from a scoring system include Optimal Rules Discovery. Optimal Rules Discovery uses multiple analytic techniques to identify the rule set that yields the best results for the business. The tool is used to maximize the net benefit the user receives from deploying an analytic solution. Rules discovery includes multiple competing objectives and business scenarios, so that the financial trade-offs can be quantified.
Alternative Uses of Analytics and Decision Support Systems
Because of the focus of our organization on process automation and optimization, case management and ActionAnalytics (operational analytics) can be developed from the same data to predict the best course of action for a given high-risk account. These predictions or profiles would include the benefit for interdiction in the case. Examples:
- the forecasted recovery amount given circumstances of the case;
- the probability that a case requires resources immediately or after subsequent activity occurs;
- the inclusion of alternative data into scores in order to automate case disposition.
The capabilities of the decision support system also allow manual actions to be scripted in software and applied automatically. As an example, moderate risk accounts may normally be called to validate identity and ensure the card is in the hands of the authorized user. Given the high variation in contact rates from one account to another and the time interval between the authorization and the contact, an automated account block/status and removal process is sometimes necessary. The decision support software can calculate the appropriate models, apply and remove the block/status code, and message the automated dialer to make the outbound call.
The use of ActionAnalytics to optimize operational process has proven to yield a 20-50% cost savings and an ROI of less than three months. These results are incremental over any of the other scoring-based benefits.
Interested in hearing more?
Contact Austin Logistics for more information.
|
Resources
|
