Scaling analytics in financial services

Financial institutions face a host of challenges in scaling up analytics across business areas, leading to increased timelines for deployment of up to 18 months, additional costs from inefficiencies, higher attrition rates, lack of business value delivery (for example, because market models are not available for use in strategic business campaigns), and more strategic projects abandoned across key business use cases, such as AI and machine learning (ML) initiatives.

At a time of rising interest rates and increased economic uncertainty, financial institutions must address operational costs and increased workloads, as McKinsey data show that model inventory sizes of US banks have increased by approximately 31 percent since 2021.

A 2022 McKinsey survey of model-risk-management (MRM) executives at 27 North American banks revealed the reasons for this increase in model inventory size, including an uptick in bank M&A, which led to the integration and amount of model landscapes, an expansion of the scope of MRM, and major redevelopment efforts.

Our McKinsey survey revealed that model development and validation teams face peak workloads due to these challenges, along with keeping pace with business-as-usual activities.

Addressing the time to market for model life cycle activities

Our survey indicates that the current time-to-market life cycle for modeling can be between 15 and 18 months, and pain points can be found throughout the process. (See sidebar, "Accelerating model life cycle time to deployment by 60 percent.")

Accelerating model life cycle time to deployment by 60 percent

A leading US bank launched a model life cycle transformation initiative to accelerate model delivery and enhance the analytical capability. It’s previous time to deployment took approximately 15 months, and a large proportion of models failed to be used in production because of the long delivery time across business lines.

To resolve these challenges and accelerate the pace of AI and ML development, a blueprint for the defined model life cycle transformation was developed, along with a diagnostic of the current pain points. Twenty pain points were identified from the ideation through the model monitoring phases, with a baseline efforts estimation. The blueprint included these:

a model standardization that highlighted a risk-based approach across the model lifecycle, with a focus on different business domains
an operating model that outlined leading practices in process optimization—for example, process parallelization, phased rollout, use case clustering, and the RACI (responsible, accountable, consulted, and informed) framework
a data and technology enabler that outlined leading practice enablers along with governance model
a talent strategy that outlined the leading practice strategy across planning, hiring and onboarding, talent management, and development

Following implementation, the estimated time to production for AI and ML use cases was reduced by about 60 percent (from around 15 months to six months). The organization was able to identify the priority opportunity areas with levers to execute based on the leading practices, and the blueprint provided a multiyear road map for efficiency.

Ideation and preparation: insufficient formalized model intake processes, a lack of early engagement from data engineers, poor clustering of use cases (especially with a bloated model inventory and early engagement)
Data acquisition and preparation: cumbersome data access control without role-based access; an absence of data domains, data products, and feature stores; insufficient use of whitelists at the enterprise level, exacerbated by late legal and compliance engagement
Model development: limited automation of clear-scoped tasks, such as documentation, testing, package scans, and updates; insufficient delineation of model-specific, risk-based procedures for documentation and test plans; insufficient code collaboration and assetization; excessive friction as a result of unstandardized computing environments
Validation and approval: limited automation of clear-scoped tasks, such as documentation, testing, package scans, and updates; delayed engagement with the first level of design (LOD) during development; inconsistent adoption of staged validation
Deployment testing: limited automation of clear-scoped tasks, such as documentation, testing, package scans, and updates; friction resulting from inconsistent computing environments; limited use of parallelization in development and validation
Model monitoring: extensive manual monitoring process; insufficient accountability for model performance

Four types of efficiency levers

To address these model life cycle challenges, modeling and analytics leaders within financial institutions can accelerate value delivery on strategic model use cases and free up capacity across model life cycle activities by deploying four types of efficiency levers, potentially reducing the time to market by 50 percent.

Automation, data, and technology enablers. Focus on the reuse and assetization of key components to industrialize the process, moving to a single environment for development, validation, deployment, and automation.
Delivery model and operating rhythms. Design standardized processes and protocols with increased compression and parallelization of activities across the model life cycle, along with model inventory management.
Clear, detailed standards and procedures. Establish a set of overarching objectives for the model development process, with actionable and specific guidance for developers across the life cycle.
Capability and skill-building plans. Establish clear roles but ensure enough cross-training and translation capabilities across the team to facilitate collaboration and interaction.

Phases of the model life cycle transformation

The model life cycle transformation has four key phases, and each phase should be strategically managed from concept to deployment. This process begins with a road map and communication, including an understanding of pain points and an estimation of the baseline efforts, such as metrics, cycle times, supporting infrastructure, and workflow management tools.

In the design phase, enablers are chosen to prioritize quick wins, and materials are designed to train impacted groups. The change story is developed based on the target state and the current situation in cooperation with executive sponsors.

Next, the rollout involves the implementation of enablers designed through pilots—for example, a sample of use cases end to end. Training is conducted and initiatives are refined iteratively based on the lessons learned during pilots and ground reality. The impact is measured in the form of, for example, efficiency gains, and this is communicated to key stakeholders.

Finally, in the scale-up phase, initiatives are deployed to the remaining use cases in the model inventory.

Successful model life cycle transformation

In addition to deploying the four types of efficiency levers, financial institutions that have demonstrated success in model life cycle transformations have followed some guiding principles:

Leaders from all key stakeholder groups across the end-to-end life cycle are actively involved.
Each stakeholder is aligned with the vision and comes to the table without any biases.
An 80/20 approach is applied, acknowledging that there will be cases where the transformation will not yield efficiency results.
Tangible progress should be communicated to build the confidence of the leadership and the working teams and to focus on quick wins.
A culture transformation is critical to realize the full potential.

Much is at stake when financial institutions use model life cycle transformations to upgrade modeling and analytics. A significant reduction in times to market for AI and ML use cases can yield a 20 to 40 basis point ROA increase for leading institutions.

In addition, these efficiency levers can increase transparency and consistency, reduce the risk of errors and attrition, and improve team health.