Alan Turing’s seminal 1950 paper "Computing Machinery and Intelligence", posed the question "Can machines think?" Introducing the idea of the first “learning machine with the potential of becoming artificially intelligent”. Since then machine learning (ML) has found its way into numerous processes; seeking to simplify our lives by making processes smarter, better and faster. Financial risk management is an industry that is rife with opportunities for ML to disrupt in the coming years, one of the most obvious areas being credit scoring.
In this blog we explore some of the main findings of the recently published Bank of England survey on ML, this is followed by our views on the challenges and potential solutions of implementing ML within a credit risk scoring framework.
ML in Credit Risk: Challenging status quo
When we talk about the rise of ML in credit risk, we quite often forget that one of the earliest real life use cases for ML was within this very industry. The primary objective was (and still is) to predict the probability that an obligor will default on their loan using bank customer data. Credit scoring models in the form of simplistic logistic regression models (a simple version of ML) have been used by lenders for over 50 years, yet when we think of ML we quite often think of more complex methods such as artificial neural networks.
ML is defined as a branch of artificial intelligence in which a computer generates rules underlying, or based on, the raw data that has been fed into it. In recent years there has been rapid increases in data availability and computing power, accelerating the rate of ML adoption. As highlighted by a recent joint Bank of England (BoE) & Financial Conduct Authority (FCA) survey. The UK financial services (FS) sector is beginning to take advantage of this and ML is increasingly being used in UK FS. We have summarised the main findings of the report in the diagram below.
Whilst ML is already widely used in banking; deployment within credit risk remains low. This is highlighted in the diagram below. Furthermore, quite surprisingly, credit risk scoring wasn’t mentioned in any of the use cases discussed as part of the BoE/FCA paper.
Source: ML in UK financial services – BoE/FCA (2018)
Navigating potholes: Internal challenges with ML deployment
As an industry, financial services are (and will always be) extremely data-reliant. Hence, this new data-driven economy goes hand in hand with fundamental changes to the structure and nature of the financial system supporting it. ML is a principal driver contributing to this new paradigm.
The use of ML has the potential to generate analytical insights, support new products and services, and reduce market frictions and inefficiencies. For the past several years institutions have been experimenting with ML building challenger credit risk models and observing significant increases in model performance. It is now the time to move ML from the lab to live production and realize this potential. To do so, however, we need to satisfy the high standards set by the credit risk industry with regards to:
- Regulatory requirements for models to be compliant,
- Business requirements for models to be explainable, and
- Analytical requirements for models to have high and stable performance.
According to the BoE and FCA survey, institutions do not see the regulation as a barrier for ML implementation, the main challenges are coming from firms internal constrains with regards to:
- Data & technology, and
- Governance & controls.
Firms have already started to lay the data and technological foundation required to make full use of advanced analytics, however, a bridge also needs to be built between current risk management frameworks and ML model development and validation practices. In essence, the current ML modelling process needs enhancement to accommodate components considered integral to credit risk management:
In addressing these challenges, ML models can offer higher predictive power, deeper analytical insight, increased operational efficiency and comply with regulations.
Let’s take the rating system as an example, this is the engine of any model driven risk function. This set of scoring models has a direct impact on the profit of the bank as misclassified clients who default produce economic loss for the bank. Higher rating system predictability is beneficial to the bottom line because requests can be assessed more accurately, which means acceptance rates can be increased and at less risk as misleadingly rejected but solvent customers are included in the portfolio through new models. Therefore, an improvement in the model accuracy by a few percentage points can save future losses in the millions for large portfolios. Loss prevention is but one area, there are however other potential benefits such as:
- Reduction in the economic and regulatory capital required due to more accurate risk measurement,
- Reduction of the model maintenance costs through self-learning procedures, and
- Increased profits as difficult to model portfolios and segments can be evaluated by new techniques increasing market size.
These are tangible benefits associated with improving the quantification of risks for banks. However, there is clearly a need for risk management frameworks & systems to evolve and the transformation required to address the previously mentioned challenges are costly and time consuming.
The question therefore is: how can I start realising immediate benefits through the use of ML?
Your shortcut to the highway: Zen Risk
At Deloitte, we have developed a cutting edge approach to binding traditional regression with ML algorithms; allowing you to make smarter decisions with ML through a new way of doing analytics. This approach is enabled through our solution called Zen Risk, which enables fast, tailored and compliant end-to-end model design, helping you to yield the benefits of ML without the creation of a black box.