About the project

The project, named "Automatic Mapping of Incoming Payments" covers a core process in the insurance business: handling around 37.000 transactions with the overall amount of ~ € 271 Bn.  Incoming payments are frequently disorganized in the bank account and the goal of the project is to ensure more secure and reliable financial transactions. We intend to utilize an AI-driven model to assist with repetitive tasks and mitigate the workload for our finance team.

Tracking payments poses a significant challenge for Munich Re because the business partners initiating transactions often do not use the most direct payment methods. The uncertainty of the payment source, typos in incoming quarterly payments (e.g. in the banking reference text) and the involvement of brokers are complicating and impeding the efficiency of the finance team in handling incoming payments.

The project proposes using bank transaction info and SAP system data to categorize different payment types. The current financial predictions made by the finance team are grounded in their collective experience, intuitive judgments and over 6000 rules. The concept involves attempting to generate a fingerprint based on a predictive model that considers past matches made by the finance team. This aims to accurately associate each transaction with the business partner having the highest probability. Your assignment is to develop a method based on transaction history and train a computer program to identify business partners. Implementing this method not only resolves current issues but also saves time through a more seamless and automated business process, enhancing communication efficiency and addressing gaps in information.

Software Stack

In the Software/IT domain, our project uses a strong combination of tools and technologies. This includes analytics tools like R, Python (PyTorch, TensorFlow, scikit-learn), and databases such as Microsoft SQL Server, Oracle, and SAP HANA. Frameworks and tools such as SAP Systems, Azure Data Factory, Azure Machine Learning, and Microsoft Power Platform are crucial for the project's success. Additionally, visualization tools like Power BI and web apps (React/Angular, Django) play a vital role in how the project is implemented.

The suggested data-driven predictive model is divided into three main work packages (WPs):

-        In WP1, we break down the problem, synthesize context, formulate hypotheses, conduct exploratory data analysis, consider biases, and go through the process of developing the model.

-       WP2 is dedicated to pattern recognition techniques, covering supervised learning, logistic regression, random forest, support vector machines, K-nearest neighbors, naive Bayes, sequence-to-sequence models, and gradient boosting algorithms.

-       WP3 focuses on evaluating the model, including real-time monitoring, anomaly detection, revalidation, and preparing it for production at scale.