Enterprises are increasingly operating in a dynamically changing and fluid environment. They are constantly changing gears just to keep pace. CXOs are constantly looking for ways to overcome or create disruptions in a world becoming increasingly complex. Infosys Consulting Blog gathers a community of subject matter experts who are driving pragmatic conversations around that which is changing and that which needs to be rethought, redefined and redesigned for enterprises to achieve market-leading performance roadmaps.

« Social CRM Tech Trends -- Facebook Integration | Main | Social CRM Tech Trends -Twitter and Data Security »

Analytical Models...Design Right and Monitor Well

Due to ever-changing business conditions and technological advancements, the predictive power of an analytical model can fade unless there is a mechanism to monitor and calibrate it, thereby keeping it in sync with decision variables. This raises few important and related questions. Can models be calibrated to ensure longevity? Is it feasible to design a self-learning model sensitive to business environment? Are there some model design and management principles which can address future performance needs?

 

 

ice cream.jpg

Due to ever-changing business conditions and technological advancements, the predictive power of an analytical model can fade unless there is a mechanism to monitor and calibrate it, thereby keeping it in sync with decision variables. This raises few important and related questions. Can models be calibrated to ensure longevity? Is it feasible to design a self-learning model sensitive to business environment? Are there some model design and management principles which can address future performance needs?

Analytical models vary in context, complexity and implementation. However the core process of building models in terms of getting the right input, applying the relevant algorithm and producing the desired result, remains the same. Getting the right data to the model is an important step but not an easy one. There is a huge surge in data as more devices are getting connected and firms are trying to capture every bit of data by building appropriate infrastructure. Data used for developing models are changing on an ongoing basis. Applying models based on old data structures may not yield an optimized result as customer decision process is evolving and becoming contextual based. An example, you are looking for a restaurant to have a quick lunch with your friend, mobile based apps can provide local search results of restaurants near your vicinity with special offers. You can also review the feedback of restaurants before picking up that menu card. It is a different approach of connecting with potential customers, being context specific and personalized. These light weight models (so called business apps) have a quick design to deploy cycle, mostly driven by business rules and a thin layer of analytical engine. They are usually part of the analytics game plan for many of the leading firms.

On the other hand we have these complex models applied in critical systems such as those used in banks to predict fraud behavior of transactions to alert the customer or the ones used in airline industry to predict maintenance interval of aircraft components. These models have a lot at stake and hence demand complex algorithms with champion-challenger methodology for self-learning and benchmarking. If designed correctly, the model can be calibrated by updating the business rules and associated components without having to update the analytical engine. The best practice is to design the model right and design for future. Some of the well accepted design principles are:

  • Input Strategy: Understand, filter and integrate current and expected future data structures and its relation with business function decisions.
  • Business Rule Aggregation: Design for future by separating the business rules and constraints from the core analytical engine.
  • Analytical Engine:  Implement advance mathematical and statistical algorithms relevant to business function with built-in fallback methods. Ensure the engine is built to handle structured and unstructured data as the case demands. Some of the common modeling techniques are logistic regression, neural network, decision tree, time series analysis, and principal components analysis.
  • Output Strategy: Outputs are presented for quick and effective decision making through appropriate medium and format.
  • Self-Learning components: Design procedures to capture False Positive and False Negative cases for analysis and feedback into the model for self-learning and enhancement.

 

Once a model has been successfully developed and deployed, it is important to monitor and calibrate it on an on-going basis. With multiple models being commissioned and decommissioned, it is easy for an organization to lose track of the same. One of the best practices of model management is to implement a Model Validation Framework at the business function level. Inventory of all the models are listed down with attributes such as model purpose, metrics impacted, financial impact, complexity, validation frequency, model owner, date commissioned. Doing so would provide complete visibility on all the models and their current status.

One of the key questions of model validation is to ask if the model still serves the purpose for which it was developed. A model validation group can oversee the validation process and provide appropriate recommendations. The validation report would contain discussion on technical aspects of the model including discriminatory and predictive power, mathematical/statistical algorithms, benchmarking and back-testing results. This way, the models can be tracked to ensure its relevance and performance as desired by the business.

 

 

 

Post a comment

(If you haven't left a comment here before, you may need to be approved by the site owner before your comment will appear. Until then, it won't appear on the entry. Thanks for waiting.)

Please key in the two words you see in the box to validate your identity as an authentic user and reduce spam.

Subscribe to this blog's feed
 

Recommend on Google

Follow us on

Infosys on Twitter