Large corporations and FIs leverage their size to successfully employ advanced ML tech to support their strategy and business decisions. The technological gap between them and regional banks and SMEs is widening, as the latter are unable to fund and tap into the research themselves, leaving them at competitive disadvantage. The case point is the SME funding: it is known that SMEs play a critical role in the economy – they provide 57% of workplace in developing countries; In Europe there are 21 million SMEs employing 100m citizens. The regional banks play crucial role in funding the SMEs, having personal relationships and sometime also statutory obligation to support regional development (e.g. Landesbanken in Germany). Still, there is a 4.8 Trillion funding gap within SME sector of developing countries, 630bn in Europe, and 59bn in the UK. One of the reason is the slow verification process and high uncertainty associated with SMEs funding done by these institutions.

We want to level the field by bringing advanced ML-based forecasting and optimisation services to our customers (FIs) and their customers (SMEs), so the funding process is more efficient for SMEs and less risky for the banks, so the funding gap can be reduced and so help to increase prosperity in developing countries and sustain growth in developed economies.

What it does

We created Fusion Optimum which supports and reconciles management of balance sheets of 3 economic agents: the bank, its line of business (e.g. serving SMEs), and bank's customers (e.g. SMEs). Fusion Optimum, using ML algorithms, allows to forecast unexpected events affecting balance sheets of a bank or SME (e.g. loan prepayments, ad-hoc withdrawals, delayed receivables, defaults), and considers this forecast while proposing optimal decisions to increase profit and reduce the risk of missing financial goals.

How I built it

Fusion Optimum is built as a set of services leveraging the Fusion API to access portfolio data of a bank stored in Fusion Risk. The services are using open source ML libraries (scikit-learn, leras, tensorflow) to implement ML-based cash flow forecasting (supervised learning) and balance sheet optimisation (deep reinforcement learning). Fusion Optimum does also offer a strategic decision support service based on stochastic programming optimisation software provided by our partner Detech Technologies.

Challenges I ran into

This is a challenging problem only large financial institutions have been able to solve for their balance sheets. The cash flow forecast, based on supervised learning, requires a proper feature engineering and labelled data to train the model. The balance sheet optimiser based on deep reinforcement learning is not relying in a training sample but needs to be fed with probability densities describing behaviour of the balance sheet and endogenous factors impacting it (market, customers, etc.). Another challenge is the size of data to be handled, esp for the stochastic programming approach the curse of dimensional has to be handled. The stochastic programming approach also requires a lot of parameter inputs which we had to rationalise to make it practical.

Accomplishments that I'm proud of

We have already created the prototype of the cash flow predictor, and integrated the stochastic programming optimiser of Detech. We built drill downs and attribution capability to track optimisation results. Together with ETH Zurich we have defined a novel balance sheet optimisation model, based on reinforcement learning paradigm, which we have not seen any equivalent published so far. We have engaged and got interest of our customers (multiple banks in Europe and Russia), who want to collaborate with us in development of Fusion Optimum. They are willing to provide their real portfolio data, expertise and time, to develop the ML based forecasting and optimisation.

What I learned

Dreaming and prototyping using a dummy use case is relatively easy but tackling deep real life problems, like this is the case for Fusion Optimum, and aiming to be a market ready product within the next few months, is a real challenge. I have learned that prototypes have to be validated by customers, our customer research phase and design thinking approach has been very useful to fine tune the proposition

What's next for Fusion Optimum

We will bring to market the optimisation service based on the Detech optimiser. We will develop ML based predictors for non-maturing products in collaboration with our customers (the banks), using their data and expertise to deliver high quality ML models. We will deliver the first of its kind balance sheet optimiser based on deep reinforcement learning (neural network) algorithms. These services will be exposed via FFDC and also available to customers of our customers (e.g. SMEs) via white labelling.

Built With

Share this project: