July 23, 2020 at 11:30 AM
In a previous blog, we looked at assessing your credit models and the challenge of building and deploying models representative of the COVID-19 crisis. At the crux of the challenge was that the fact that:
- Current models are not going to be working that well, due to the current environment
- Standard new models will also not cut it (models with the standard development data periods will not represent the current credit environment and so these models will unlikely work as well as needed).
A solution to this is short outcome/strict performance models, whilst there are some limitations, these will be models that best represent the COVID-19 economic upheaval period.
As I said, you can read more about this in my previous blog.
Whilst these models may help in the short term, during a tumultuous period, it may be necessary to redevelop your models more regularly. Scorecard building projects can be costly in both time and effort. The deployment of a machine learning environment can reduce costs and ensure the models are most representative of the consumers in the current economic environment.
Principa’s Quick-Step machine learning models allow lenders and collectors to adapt and deploy rapidly. Models can be built and rebuilt at regular intervals considering all relevant data observation and outcome periods.
PRINCIPA's MACHINE LEARNING OFFERING
Over the last 4 years, Principa have been creating a machine learning environment where scorecards can be developed, deployed, monitored and redeveloped seamlessly. We’ve deployed this solution at South Africa’s leading collections recovery agency. The components of the machine learning environment are as follows:
1. Data management platform (DMP) – this platform allows us to:
(a) Ingest data from a variety of sources, clean the data and then combine the data
(b) Create features (essentially these are variables that can be used for scoring, for example aggregated variables across a time series or/and at customer level). We use Spark SQL.
(c) Create a data asset. The data asset is a data set from which reports can be pulled and the data is also modelling-ready.
2. ML Retraining Environment (Python) – this environment allows us to:
(a) Retrain models using a variety of methods. Our methodology is to build multiple scorecards and then to choose the scorecard that offers the greatest lift.
3. Stratus ML Model Execution Engine– this platform allows:
(a) Scores to be calculated – models are migrated directly from the ML Retraining Environment to Stratus and then tested.
4. Monitoring Reports – Populations and models need to be monitored.
(a) Our reporting component allows for daily reporting (execution control reports and scorecard monitoring report) which can help determine when data shifts have occurred.
With our quick-step models you are not stuck to one type of model. The modelling approach involves preparing the data, sampling the data, selecting an outcome definition and then modelling. The advantages of our ML retraining environment are that we build models utilising multiple methodologies concurrently. Some of the methodologies utilised include:
The model strengths can then be analysed, and the best model selected for deployment into Stratus – our machine learning execution engine.
Once the machine learning model has been deployed, the data will accumulate within the data asset. The data is modelling ready and extraction and sampling is all that is required prior to model retraining. A traditional model build typically takes 3 months with another 3 months for testing and deployment (sometimes longer). In adopting Quick-Step machine learning you will be able to rebuild models within a matter of days and then deploy and test within a matter of weeks thus reducing the time to market by up to 90%.
If you would like to learn how Principa’s Quick-Step machine learning models can assist your business, please get in touch with us on firstname.lastname@example.org.