As machine learning becomes more prevalent in modeling exercises today, financial institutions are facing a significant challenge. Analysts and pricing professionals want to retain the transparency of conventional models while utilizing the flexibility delivered by Machine Learning models. Machine Learning models have been found to allow a much better fit than conventional statistical models. Because Generalized Linear Models (GLM) cannot capture causality, or the reason a significant event occurs, Machine Learning models are superior for understanding causal relationships. On the other hand, GLM models are ideal for understanding correlation, or the relationships among data variables. An analyst is thus faced with the following challenge:
How do I understand both correlation and causality within my modeling projects?
Earnix has addressed this challenge in version 9 of our software – with the implementation of Hybrid Modelling. Hybrid modelling gives an analyst the ability to build an analytical project that combines the proven predictability of GLMs with the accuracy of machine learning models such as Random Forest (RF) and Gradient Boosting Machines (GBM). Using machine learning techniques like GBM combined with GLM allow a monotonic, continuous, smoothing effect to be applied to data variables – to provide more accuracy in analytical answers. Hybrid modeling delivers superior predictive capability which translates into greater insight into customer behavior. This makes it possible for more personalized custom-built offerings to be proposed to the customer. Better formulated models are thus the next step up in this data-enriched, more defined and tailored effort to meet individual customer needs.
The benefits of combining traditional GLM modeling with machine learning models via a hybrid modeling approach are numerous. These benefits include:
- Improved Accuracy. Traditional GLM models require a significant number of assumptions or hypotheses to be introduced when constructing the model. To achieve the accuracy of machine learning models, GLMs require more “human” work on aspects like feature engineering. Machine learning models achieve this accuracy automatically without human intervention, thus the use of machine learning models alongside GLMs improve overall accuracy.
- Ability to Use More Data. Integrating additional and third-party data sources is easier when using machine learning models.
- Speed to Market. Using multiple models to achieve a desired outcome facilitates faster testing and model building. Minimum human intervention using machine processes allows for more rapid variable selection, transformation and smoothing, and interaction setting. All of this results in quicker time to insight.
- Accelerated Innovation. By solidifying existing processes around claims and fraud with hybrid modeling, entry into new areas such as personalization and recommendation systems, usage-based insurance, and other emerging areas can be explored.
The hybrid modeling approach Earnix has introduced to the market has resulted in many customers wanting to implement the new capability – and significant results have been realized. One large European insurer has adopted the hybrid modeling approach to better understand wealth and income distribution of their customer base, which is measured by the GINI index. They compared the results of using traditional GLM modeling with hybrid modeling and found that the hybrid modeling approach increased the GINI coefficient, or understanding of income distribution from 66.7 to 73.4, or roughly a 7% increase. The understanding of the causal effect of this coefficient was also greatly improved.
As hybrid modeling makes its way into analytical modeling inside of financial services, we will see organizations become more advanced, automated, and accurate in their customer interactions. This will benefit customers by providing more contextually relevant and proactive offers. The customer experience will be greatly enhanced as a result of the analytical insight improvements happening behind the scenes – all because of hybrid modeling.
For more information, please see the following resources: