Unified multi-modal neural network for improving credit scoring accuracy

The client is a prominent international banking group with headquarters in New York City, offering digital banking, payments, wealth management, and personal and business loans across multiple markets.

The bank was planning to enter the Indian market with retail banking products, facing several critical challenges:

• Thin credit histories – Large population segments have limited or nonexistent credit histories
• Regulatory differences – Indian banking laws differ significantly from American credit scoring systems
• Model translation issues – U.S.-built credit scoring models don’t translate well to emerging markets
• High default risk – Analytics team forecasted elevated loan default risks in the new market

The company needed an advanced AI-driven credit scoring solution to accurately assess borrower reliability in markets with limited credit data while maintaining profitability.

The original approach treated three data sources (card transactions, debit account transactions, credit history) as mutually independent, using separate neural networks and logistic regression for final scoring.

Solution: We developed a unified multi-modal architecture that processes embeddings from all data sources simultaneously, allowing the model to identify correlations between different data types that were previously missed in the independent approach.

Initial experiments with direct raw data input to a unified model resulted in long training times and difficulty identifying the impact of individual improvements under time constraints.

Solution: We implemented an embedding-based approach where each data source is transformed into vectorized representations before feeding to the main neural network, enabling faster training and better parallelization of development processes.

Not all clients have complete data across all three sources (card transactions, debit accounts, credit history), creating gaps in the scoring process.

Solution: We designed constant embedding mechanisms that use average embeddings for missing data sources, ensuring the model can still make accurate predictions even when some client data is unavailable.

Supporting multiple models instead of a single end-to-end solution created development and maintenance complexity while requiring fast iteration cycles.

Solution: Our modular architecture allows for independent improvement of individual data source models while maintaining the unified prediction layer, enabling parallel development and faster experimentation with new approaches.

The original architecture made it difficult to incorporate additional data sources without major system redesign, limiting future expansion capabilities.

Solution: We built a flexible framework where new data sources can be easily added as additional embedding inputs, allowing the bank to enhance scoring accuracy by incorporating new data types without architectural changes.

Balancing model complexity with inference speed requirements while maintaining high accuracy standards for real-time credit scoring decisions.

Solution: We optimized the multilayer perceptron architecture for fast inference while preserving the quality improvements from multi-modal data fusion, ensuring production-ready performance with significant accuracy gains.

Fast training: Such a model is trained quite quickly due to its architectural simplicity and the
absence of recurrent layers

Parallelization: This approach allows for the acceleration of the development of a general solution due to the parallelization of the processes of improving individual models

Multi-model support: The challenge is that we must support a set of models instead of just one end-to-end model.