Machine Learning Operations (MLOps) in Banking | Deloitte US has been saved
A blog post by Chida Chidambaram, managing director, Deloitte Consulting LLP; Chris Thomas, principal, Deloitte Consulting LLP; Zach McLaughlin, senior manager, Deloitte Consulting LLP; Jared Artz, manager, Deloitte Consulting LLP; Rishabh Kochhar, senior consultant, Deloitte Consulting LLP.
With recent technological advancements, data has become an invaluable resource for organizations. Computational techniques like machine learning (ML) have enabled organizations to deliver better value to consumers at lower costs. Within banking, ML algorithms and analytics have improved end-user services, personalized customer experiences, prevented fraudulent claims, and boosted sales.
Firms use large volumes of historical data to train ML algorithms to predict outcomes or uncover patterns that generate crucial insights. For banks, managing the new data velocity and scaling of ML algorithms can become complex. “Stale” models, which are ML models trained on outdated data, can produce inaccurate predictions that can be harmful. Machine learning operations (MLOps) has emerged as a framework to address these complex challenges and enable ML at scale.
MLOps extends DevOps principles to machine learning operations to manage data and machine learning model updates due to rapidly evolving business processes. It is the set of practices, technologies, and skill sets that streamline the deployment, maintenance, monitoring, and training of models while also improving the collaboration between data scientists and operation engineers. There are four key tenets to deliver the foundational need for “continuous training and continuous delivery” for ML:
- Model deployment – This includes a defined and, at times, automated process in which models are designed, developed, and released to production, like many of the software delivery life cycle (SDLC) processes used for current business applications. This tenet’s critical aspect is the process’s repeatable steps, enabling data scientists to deploy models in a scalable, measurable, and repeatable manner.
- Model monitoring – Post-deployment, the monitoring of a model for accuracy and performance is critical. A model may become “stale” over time, and if new real-world data representing new consumer behavior (e.g., COVID-19) is introduced, the results may be non-optimal. A robust monitoring capability enables teams to identify those occurrences and retrain the model to deliver desired insights.
- Model training/ retraining – Once model accuracy varies and/or predefined time passes, models should be retrained using updated datasets so that the model is optimized for the target outcomes. For example, in a survey by the Bank of England, it was found that 35% of banks reported a negative performance of their models after the COVID-19 pandemic.1 The pandemic caused a change in consumer behavior, which generated new data that the models had not been trained on. In these scenarios, promptly updating ML models is critical to ensure that banks are providing the right recommendations to best serve their customers.
- Automation – This is the most critical principle to deliver ML at scale. As firms increase their reliance on ML models to solve problems and improve experiences, the manual effort necessary to develop, deliver, monitor, and retrain models is not sustainable. Automation enables delivery at scale, much as DevOps pipelines do for business applications. The automation of operational steps enables data scientists to focus on building new or enhancing current models to improve insights, performance, or experiences.
With 71 zettabytes of data generated in 2021 and dynamic world events changing consumer behaviors, data used to train models can rapidly become out of date. In a 2021 survey involving 750 business decision-makers, only 8% felt that their companies’ models were sufficiently sophisticated.2 This poses a serious challenge, and MLOps has become a necessity for banks as they serve consumer needs, aim to mitigate security and financial risks, and improve financial performance. MLOps helps banks in three facets:
- Risk mitigation: Outdated ML models can deliver poor predictions that negatively impact consumers or organizational outcomes. Further, these outdated models can increase legal and compliance risks. The automated monitoring and associated retraining can mitigate them.
- Efficiency: With MLOps, internal data science and ML teams are empowered with automation frameworks and tools that allow them to focus on building the best models instead of manual, repetitive deployment, and monitoring tasks.
- Scale: MLOps can automate the deployment, monitoring, and retraining of ML models in a repeatable, replicable method. This allows organizations to manage a larger number of models, allowing them to scale rapidly as new data, events, or use cases arise.
Banks serve a wide range of consumers, ranging from individuals and small businesses to large corporations and governments. Machine learning has been instrumental in enabling banking institutions to process larger volumes of data faster than human rule-based systems to uncover subtle and hidden patterns in diverse data sets. For example, a global bank doubled its fraud detection capabilities by deploying ML models to identify anomalies from historical transactions.3 Also, a top European bank is using its rich datasets and cloud computing to build ML models that prevent fraud.4 In both examples, these ML models paired with human expertise have delivered better predictions to best serve their consumers.
With constantly changing consumer behaviors and needs, banks may find themselves at a disadvantage if their model outputs aren’t current. MLOps can help banks deploy better real-time models and improve their top-line results with key insights to cross-sell their products to the relevant consumers. With a rise in digital-only banks, MLOps can be a key enabler to help traditional banks stay relevant in retail banking by providing personalized services and experiences to their customers.
A major financial services institution in the United Kingdom leveraged MLOps to scale its data science abilities to build and launch ML services in a faster, scalable, and autonomous manner. Using Continuous Integration Continuous Delivery pipelines in the cloud, the financial institution was able to reduce the time from ideation to value realization by 60% and standardize the development and deployment of ML models across teams.5
In Europe, a collaboration of five Dutch banks used MLOps to standardize development and deployment of AI and ML to monitor payment transactions for signals that could indicate money laundering or the financing of terrorism. Being cloud-native, the banking consortium found it even easier to leverage ML, with collaboration and joint ownership of models across tech, business, and IT.
Today, traditional banks face an increasing threat by modern digital-only banks that operate with significantly lower capital expenditures, more personalized services, and faster product delivery. For traditional banks to stay competitive and increase their value by leveraging customer insights, MLOps is a critical requirement. However, MLOps adoption across the banking industry is low and is often overlooked until scale challenges emerge. The low adoption rate provides first movers an opportunity to occupy a strategic advantage, though. By adopting MLOps, these companies enable their data scientists to rapidly deploy models and better meet the dynamic needs and regulations within the global marketplace. A strong MLOps competency will future-proof a bank’s business and distinguish it as a world-class institution delivering top-tier experiences for its consumers.
Endnotes
1. Akinwande Komolafe, “Retraining model during deployment: Continuous training and continuous testing,” neptune.ai, February 21, 2023.
2. Petroc Taylor, “Total data volume worldwide 2010–2025,” Statista, September 8, 2022.
3. Wall Street Journal Custom Content, “At Capital One, enhancing fraud protection with machine learning,” March 1, 2023.
4. Barclays, “Machines can’t do it by themselves,” January 16, 2020.
5. Maira Ladeira Tanke et al., “How NatWest Group built a scalable, secure, and sustainable MLOps platform,” AWS, April 26, 2022.
Get in touch
Chida Chidambaram
Managing Director | Deloitte Consulting LLP
Chida leads the firm’s Cloud AI/Machine Learning practice and delivers cloud and machine learning strategy, implementation, and operation in all public clouds. Chida works with clients to create operating models and transformation roadmaps and strives to align people, process, and technology. He brings strong thought leadership experience to engagements and thrives in supporting executive stakeholders achieve performance improvement and modernization goals across the financial services, consumer, automotive, transportation, hospitality and services industries. Chida is a serial tech entrepreneur and an avid community builder in the startup and developer ecosystems. He holds a BS in Physics and an MS in Computer Applications from India and holds an MBA from The University of Illinois Urbana-Champaign.
Chris Thomas
US Cloud Leader
Chris is a principal at Deloitte Consulting LLP and is Deloitte’s US Cloud Leader. He has over 20 years of strategy consulting and hands-on transformation experience in the cloud and core technology domains across industries and globally. Chris has extensive experience partnering with senior executives to enable business outcomes by shaping and implementing large-scale technology transformations, cloud-centric operating models, strategic cost optimizations, global outsourcing programs, and workforce of the future initiatives. He brings the perspective that humans and machines must develop a symbiotic relationship, each with specialized skills and abilities, in a unified workforce that delivers multifaceted benefits to the business. Chris writes frequently about emerging technologies and banking, and his observations appear in the Wall Street Journal’s CIO Journal, Deloitte Tech Trends, and financial-services industry forums and technology journals. He holds a B.S. from Miami University (Ohio) and a M.S. from Northwestern University.
