Accelerating the adoption of AI in banking with MLOps

There is rapid adoption of artificial intelligence (AI) and machine learning (ML) in the finance sector. AI in banking is reshaping client experiences, including communication with financial service providers (for example, chat bots). Banks are exploring ways to use AI/ML to handle the high volume of loan applications and to improve their underwriting process. AI/ML technologies are also transforming the operations of financial institutions, providing significant cost savings by automating processes, using predictive analytics for better product offerings, and providing more effective risk and fraud management processes and regulatory compliance.

While AI/ML offer significant opportunities for banking and financial institutions, there are several challenges that these institutions face when implementing these technologies at scale including data quality issues, compliance with relevant regulations and standards, AI/ML model explainability, bias and fairness, AI/ML model governance and management.

Enter MLOps

MLOps (Machine Learning Operations) can address some of the challenges faced by financial institutions when implementing AI in banking projects by providing a systematic approach to managing the lifecycle of machine learning models.

MLOps is an approach similar to DevOps – which is widely adopted in software engineering today. Much like DevOps, MLOps principles focus on collaboration between the various parties that are involved in the process, such as developers, system administrators, and data scientists. It includes standardising practices for data preparation, model development, model deployment, model monitoring.  MLOps typically yield improvements in the code quality, faster patching, upgrades and updates, more efficient releases leading to more trustworthy and scalable AI in banking. 

How can banks apply MLOps?

MLOps practices can be applied to a wide range of AI in banking use cases, as we will explore below.

Fraud Detection

Fraud is a persistent and serious problem for financial institutions, and it is critical to detect fraudulent activities as quickly as possible to minimise losses and protect customers. According to estimates, e-commerce losses to online payment fraud were 41 billion U.S. dollars globally in 2022. The figure is expected to grow further to 48 billion U.S. dollars in 2023. Federal Trade Commission data shows that consumers reported losing nearly $8.8 billion to fraud in 2022. 

Banks and financial institutions are developing and deploying fraud detection models that can automatically analyse large volumes of data, identify patterns, and detect potential fraudulent activities in real-time. These models can analyse transactional data, customer information, and other relevant data to identify potential fraud.

The process of developing and deploying a fraud detection model involves several steps such as data preparation, model development, model deployment, model monitoring.  MLOps can help ensure that each step is carried out effectively and efficiently:

• MLOps practices can help data science teams at banks to automate the process of collecting, cleaning, and preparing the data to be used in the model.
• Data scientists at banks can use MLOps practices to ensure that the model is developed in a consistent and reproducible manner, using best practices for data analysis, feature selection, and model training.
• When AI/ML models are moved to the production environment, MLOps can help ensure that the model is deployed in a scalable and reliable manner, using best practices for containerisation, orchestration, and monitoring.
• AI/ML teams at banks continuously monitor the performance of the models in production, detecting any issues, and making necessary adjustments to improve the model’s performance leveraging MLOps practices.

Personalisation and customer segmentation

Despite significant investment in AI, banks have not been able to effectively apply predictive insights from their machine-learning (ML) models to manage customer personalisation programs and to inform their campaigns. Inconsistent customer data, narrow scope of ML models, one-off use cases, hard-to-replicate models, and limited knowledge sharing are some of the common hurdles faced by banks to successful personalisation at scale. 

Banks need to get better at developing a full suite of ML models capable of driving personalised engagement at every touchpoint. Most models are trained on isolated moments with short-term, product-driven aims, such as boosting mortgage applications or account openings, rather than on identifying drivers of customer lifetime value and shaping their customer interactions based on those insights. MLOps can help ensure that the model is developed using best practices for data analysis, feature selection, and model training.

In most cases, ML models and campaign-management systems often lack feedback loops to connect them, resulting in banks being unable to apply predictive insights from their ML models to inform campaign execution and decision making. 

MLOps promotes collaboration between data scientists and various operations teams, helping to ensure that banks can confidently apply predictive insights to decision making and create personalisation programs. Banks can apply MLOps practices to build an integrated technology stack to orchestrate the insights loop that connects enterprise data to the ML models, feeding the resulting inputs into campaigns.

Credit risk assessment

Credit risk has always been a challenging area for banks, given the multiple factors that form an individual’s risk profile. Credit risk assessment involves analysing a borrower’s credit history, financial statements, and other relevant data to determine their ability to repay a loan.

The process is even more complicated for business borrowers as data across a variety of parameters and time periods must be aggregated and analysed to create a holistic picture of risk. 

AI/ML techniques are reshaping credit risk assessment. ML models are continually fed data, made to extract insights, and then draws predictive insights on new datasets. This is an iterative process which results in the ML models getting more accurate with each round of training. 

However, ML models may still contain assumptions that can pose a significant challenge when analysing noisy historical financial data and may lead to poor model performance. There’s also a risk of overfitting the data, as ML models are more sensitive to outliers than traditional analytics. 

Feature engineering involves selecting and engineering the features that will be used in the ML model. MLOps can help banks and financial institutions to help automate this process and ensure that the features are selected based on their relevance to the credit risk assessment task.

AI/ML models can inadvertently reflect biases and prejudices that exist in the data used to train them, leading to unfair or discriminatory outcomes. This is a particularly sensitive issue in the financial services industry, where decisions based on biassed models can have significant financial and social implications. Imposing constraints on the model to control for model biases or counterintuitive behaviour can also be an onerous task for some ML techniques.

MLOps can help address bias and fairness issues by providing tools and techniques to data scientists at banks and financial institutions for detecting and mitigating bias in machine learning models. MLOps can also help ensure that models are built and tested using diverse and representative data sets.

In addition, decomposing ML models can be complicated, thus creating issues when there is a need to explain the model’s functionality in detail. The black-boxed nature of AI/ML could be a factor holding back implementation, especially among larger regulated financial entities. MLOps can help improve model explainability by providing tools and techniques for interpreting and visualising model results. MLOps practices and tooling can also help AI/ML teams at banks to provide transparency into how models are built, tested, and deployed, making it easier to understand and explain model behaviour.

Adopting a secure open source MLOps platform

Open source software is increasingly playing a crucial role in MLOps for financial services. An open source MLOps platform can provide financial institutions the flexibility to customise their MLOps processes and tools to meet their unique needs. 

Open source software is significantly more cost-effective than proprietary software and often at the forefront of innovation, with a large and active community of developers constantly improving and extending the capabilities of the software. This can help banks and financial institutions to stay ahead of the curve in terms of MLOps capabilities. 

Managing open-source software and all of its dependencies securely is an imperative for financial institutions. That holds true for an open source MLOps platform as well. Banks and financial institutions need secure open source software for building and maintaining AI/ML powered intelligent applications without compromising on their compliance, security, or support requirements. For banks and financial institutions looking to adopt AI/ML at scale with a secure open-source MLOps platform, Canonical offers Charmed Kubeflow and Ubuntu Pro.

Want to learn more about secure open source for financial services? Read our white paper!

You can also watch our on-demand webinar to find out how financial institutions can use secure open source MLOps at scale to achieve enduring business value.

Photo by Clarisse Croset on Unsplash