To address the BOP categorisation, BSG proposed a model to predict the transaction category using data science and machine learning, based on transaction details. The model proposed would predict transaction categories with an accuracy of 82%. Customers would then have an opportunity to approve the category or provide guidance for reclassification.

The BSG team proposed an iterative approach to building the model:

• Iteration 1: Data exploration – understand the data available and determine the relationships within the data that will assist in predicting a BOP category
• Iteration 2: Predictive model development – using R, the team built a proof of concept (POC) random forest model, which leveraged relationships within the data to predict a BOP category
• Iteration 3: Model accuracy optimisation – by developing additional sets of features and adjusting the model tune grid, the team were able to develop a model with high accuracy
• Iteration 4: Model validation and documentation – with development complete, the team documented the process undertaken to develop the model to ensure repeatability 

By removing the manual processes and back and forth communication requirement, the ability to predict the transaction BOP category would significantly reduce the time required to successfully categorise transactions. This time-saving would result in operational cost savings and overall business unit efficiencies. 

To address the query volumes, the BSG team proposed using data science, machine learning and robotic process automation (RPA) to reduce the manual intervention required to resolve queries. The model used natural language processing (NLP) to evaluate the emails received and identify those relating to inward cross-border trace progress. It then analysed the email, extracted relevant data (such as account number, customer details, etc.). This data was then passed onto the RPA process, which then queried the transaction data from the system and appropriately responded to the customer. 

Once again, the BSG team proposed an iterative approach:

• Iteration 1: Data transformation and model development – the team broke down the email into its components (subject, header and footer) and appropriately cleaned the data to develop the NLP model. Using Microsoft Azure, the team developed a minimum viable product (MVP) version of the model
• Iteration 2: Refine model and create content extractor – the team developed a topic model based on the subject line, which was then stacked against the original model, improving its accuracy. The content extractor used business rules and regular expression in python and R to extract the relevant information (such as account number, customer details etc.)
• Iteration 3: Deploy and integrate model – the model was then deployed as an API and integrated into the Microsoft Flow process that acted as the solution orchestrator 
• Iteration 4: Documentation and knowledge transfer – the developed process was well documented and time was spent with the bank’s internal team to ensure future modifications and updates could be made internally, as needed

BSG was responsible for developing the Microsoft Azure functionality, using an API, while an internal team at the bank was responsible for developing the RPA functionality.