As an outcome of the inaugural AIdeathon held last year, three winners were shortlisted to progress to a proof-of-value stage. This video showcases the successful completion of the three proof-of-value projects. Further details are provisioned in the sections below.
Forecasting the flow of funds involves obtaining a pragmatic assessment of an asset’s future financial position. It is also just as important to ensure that our clients are cognisant of the various elements that determine the overall direction of an asset’s price. For such endeavours, both quantitative and qualitative analyses are essential.
In order to harness the benefits from both analytical techniques, it is first necessary to examine a few important considerations. For example, “how should funds be designed to ensure continual and sizable assets-under-management (AUM) flows?”, and “how can AUM projections be accurately forecasted so that accurate financial forecasts can be generated?”.
We recognised the need for a prediction engine that is premised on all the above considerations, with a focus on enhancing our understanding of clients’ requirements by precisely forecasting their AUM funds flow. With much model development effort, the team managed to yield a positive outcome. Their Funds Flow model was able to successfully provide a quarterly forecast of the monthly AUM fund flows for most of the Group’s financial products. Moreover, with the aid of various charts generated by a combination of regressors and classifiers from the predictive model, sharper investment decisions could be made for our clients as we are able to use data to quantify the key factors driving the fund flows.
Now, UOB is significantly better positioned to serve our clients as we are well-resourced to meet their needs by presenting to them data-driven quantitative support to assist with their investment objectives. Businesses too have clearer guidance on which funds to focus on for marketing, and how to market them.
The collective effect of the Fourth Industrial Revolution and the coronavirus pandemic has engendered a massive acceleration in digital transformation to create a socially-distanced, digitally-connected world. Unfortunately, several undesired ripple effects were triggered by this kaleidoscope of events – one of which was a marked increase in financial scam cases.
When a customer first suspects he has been scammed, he must first report the scam transaction, which may have happened a few days or a few weeks ago. In that time, there is a strong possibility that the scammer would have transferred the money to another country; making efforts to trace and recover the funds far more challenging.
Time is of the utmost essence – this was the consensus drawn by the team entrusted to conduct a detailed study and reach a solution. On top of this, the sought-after solution should leverage upon the large volumes of data that is available. What followed was the development of a Scam Prevention model that was designed using big data analytical approaches derived from AI methodologies.
The model works by learning from past customers’ transactional behaviour in order to stop unauthorised or fraudulent transactions immediately. Augmented with prior knowledge of fraudulent accounts and information received by the authorities, a stack of ensemble models analyses the transaction patterns and profiles of customers to filter out suspicious transactions in real-time.
Given that AI is especially effective at spotting complex patterns and trends that humans may miss, our operators can promptly respond to any suspicious activity, before a would-be scammer has sufficient time to move the money out to another country.
For a team of highly trained officers at UOB’s Trade Operations Centre, their monthly routine consists of enduring the laborious process of manually registering hundreds of complex trade documents into the Bank’s processing system. Apart from the sheer amount of time and effort spent on manual data extraction and input, there also runs the risk of human error incurred.
Designed for intelligent document analysis, the Trade Docs model was conceived to automate the laborious tasks so that officers may focus on the technical subtleties requiring advanced knowledge. By applying various methodologies emanating from natural language processing (NLP), the model can extract key entities and clauses to classify clauses that are important to document processing.
It is easy for officers to derive benefits from the Trade Docs model, as it can be deployed as a service once it has been trained on a large corpus of annotated data. Using images digitally received from the customer, all officers have to do is to retrieve the document. The Optical Character Recognition (OCR) engine automatically takes over from there by digitising the trade document into a machine-readable digital text file. Unstructured data from this file is then ingested by an AI model that employs a plethora of NLP algorithms for various computations; namely, key clause extraction and classification, named entity recognition, and sentiment analysis. The model has an accuracy of 90% across the various document review tasks. Finally, outputs from the model are presented onto a Streamlit dashboard for users to examine.
Two key objectives were fulfilled by the AI-enhanced workflow. Overall document readability has been improved, and the model was able to provide explainability so that the intricate details of each document may be readily consumed by our officers. On account of these, the Trade Docs model has managed to improve both the accuracy and turnaround time for Trade Operations.
