Deep Pockets - Loan Offers for SMEs using Deep Learning

• 

  Deep Pockets - Best Loan Offers for SMEs using Deep Learning Optimization

• 

  Eliminate bias and get instant best offer recommendations on loan requests by optimizing the balance sheet of the bank

• 

  Deep Pockets - Demo - Best Product Offers

• 

  Tensor Flow - Neural Network Agent

Inspiration

8 out of 10 SME loan applications are refused due to lack to history data. For the applications that are approved, the lead time is very long due to manual assessments and loan pricing. On the other hand, retail customers get loans instantly with many price comparison apps providing easy access and great customer experience. We want to bring the retail experience to the SME sector which is currently underserved.

What it does

The app uses deep learning based optimization of the bank's balance sheet to instantly suggest 3 best offers given the ranges and constraints specified by the loan officer. Deep reinforcement based learning is used to solve a constrained optimization problem over high dimensional parameters. Once the model is trained, the Optimizer can be used in real-time to suggest the best loan products.

In the future, the App can be linked to guided SME onboarding process integrated with financial literacy program to bridge the lack of data issue to build better accuracy in the pricing models, which can be further integrated into the Optimization routines.

How we built it

The core application service has been created using Keras with TensorFlow - the architecture is based on our collaboration with ETH Zurich aiming at creation of a reinforcement learning based balance sheet optimiser (https://arxiv.org/pdf/2009.05034).The UI has been built using inVision online platform (https://www.invisionapp.com). The agent is able to use the FFDC ALM Data Service as its training data source.

Challenges we ran into

The challenges included implementation of a sufficiently sophisticated learning environment, which could simulate a realistic market behaviour and related balance sheet dynamics. Another challenge was to get sufficient data to train the agent - we have used our previous work on Deep Scenarios to generate synthetic training data sets (see https://devpost.com/software/zzz-byhx0q) and the FFDC ALM Data Set service as the source of the real data seed. The obvious main challenge was to optimise the architecture and calibration of the agent so it comes up with acceptable proposals. The neural networkrk behaves however surprisingly well - nevertheless we expect more challenges due to increased sophistication of the simulated balance sheet looking forward (e.g. addition of complex derivative financial products).

Accomplishments that we're proud of

As far as we know this is the first practical use of a deep learning based balance sheet optimisation. We have achieved it in collaboration with academia in a relatively short time with just a handful of dedicated team members. We were also able to connect and use our FFDC ALM Data Set Service.

What we learned

There is a steep learning curve when attempting to use deep learning, even with the Keras API. Another lesson learned was to leverage better existing market simulators instead of developing economic models from zero.

What's next for Deep Pockets

We want to start proof of concepts with several of our banking customers, aiming at serving esp. their SME portfolios. Related challenges of credit scoring for people with non-existent credit history are being developed - we will also be able to integrate the existing partner solutions in this area. The aim is to have an online learning agent (following the reinforcement learning paradigm), which is able to provide immediate recommendations based on the current market state.

Built With

• python