A fresh take on risk and valuation
Back to all services

Machine Learning Model Validation Framework

Performance... Explained
Designed to meet requirements of Risk Measurement, Validation team and Board level executives
Finalyse offers a classic approach holding true for Machine Learning model validation. Our model validation scope can be tailored to accommodate your requirements: A purely technical and qualitative exercise focusing on whether a certain methodological choice is fit for the intended purpose. A larger scope covering the implementation of a certain methodology. And ultimately, a validation of the whole process through which the model is deployed within the organisation.

How does Finalyse address your challenges?

Bring in the experience of a multidisciplinary team of experts with extensive exposure to regulatory and non-regulatory models.

Unleash the full power of your ML models: ascertain that you can justify their performance, interpret and explain their outputs and find the right balance in case of multi-objective trade-offs.

Ensure optimal communication between your business lines, subject matter experts and the data scientists creating your ML models.

Ease the compliance with the regulatory expectations of having a proper knowledge of - and widespread awareness around the models in use (TRIM, BCBS 2005, CRD Section II, CRR2 Art. 189 and 293).

Bring in our framework and governance set-up capabilities to ensure the best traction for your models.

Ensure independence of your validation from all possible agency pressure.

Validate your model performance

How does it work in practice?

Business Case

You are using an ML model to screen credit applications. You are not aiming at minimizing the level of defaults, but are in fact trying to optimise your risk-return payoff. And you want to do it in a way compliant with the European Regulatory Framework, implying e.g.:

  • No implicit or explicit discrimination on certain demographics
  • Privacy respect / right to be forgotten
  • Actionability of the model and overall integration in the risk management framework and governance

 

Working with Finalyse on your Model Validation Framework will allow you to leverage on the broad experience we have accumulated in the area of risk measurement and management, as well as on varied expert profiles brought together under a team.

Key Features

  1. A healthy mix of quantitative, business and regulatory profiles for all your ML validation needs.
  2. We can bring clarity and explainability when complexity blurs the underlying messages, or go looking for that elusive optimisation in your place – just as long as we don’t validate it as well.
  3. Keeping tracks of the European regulatory progression on ML models as well as more local initiatives to ensure we have the best practices in place.

What Finalyse experts say

19 Mar, 2020
Machine Learning Model explainability –why is it important and methods to achieve it

Machine learning is today’s buzzword and it has gone through some phenomenal changes over the last few years. However, despite widespread adoption, machine learning models remain mostly black boxes. Hence it is essential to have techniques for explaining these black boxes in an interpretable manner. This article looks at the methods that are most used for interpreting machine learning models: SHAP, LIME and CP Profiles. It discusses advantages and drawbacks of each and shows how are they being used in practice.

Read
10 Dec, 2017
Machine Learning in Risk Management

This Expert input works as an overview of the basis of most available machine learning techniques and serves as a great stepping stone to get into the intricate world of ML. Whilst mostly written with credit risk in mind – offering some advice for the use of machine learning to help us model Expected credit loss and its components (PD, LGD, EAD), the list of the potential techniques depicted steps far beyond this relatively singular use and examines a multitude of approaches, ranging from Supervised ML (Decision trees, Artificial Neural networks, etc.) through ensemble ML (Random forests, Gradient boosting), to unsupervised ML (Deep learning, Clustering methods, etc.). The input concludes with several general tips and tricks regarding machine learning.

Read