Machine Learning in Finance

The financial sector is an extremely exciting and highly complex environment. The skilful use of machine learning enables financial institutions to make good decisions and manage risks optimally.

Key Technology of the 21st Century

In recent years, machine learning has established itself as one of the disruptive technologies that has fundamentally changed our understanding of tackling complex challenges. There are many impressive examples of applications in speech processing, in generative technologies (e.g., images, texts, speech), or in pioneering game strategies (e.g., chess). But there are also classic problems in mathematical finance that can suddenly be solved under much more realistic assumptions.

Neural networks are widely used in the field of artificial learning. Inspired by biological models, they consist of a structured arrangement of interconnected neurons that convert signals into decisions according to the chosen architecture. The connections between the neurons are weighted; the higher the weight, the stronger the signal transmitted. Appropriate external stimuli are used to gradually improve the weights and to train a first-class instance for the specific task in a reasonably short time.

  • Optimisation of business, investment, and hedging strategies
  • Asset-liability-management (ALM) and quantitative risk management
  • Valuation of derivatives
  • Scenario generation
  • Automation and digitalisation
  • Forecasts (e.g., credit migrations and defaults, fraud detection, marketing)
  • Utilisation of crowd intelligence

At the heart of our innovation projects are dynamic decision problems along a term structure, where the environment changes randomly and our decisions affect the system. We model the balance sheet of a business model as a state and optimise its structuring over time. We also must ensure compliance with several rules. In technical terms, we analyse high-dimensional stochastic optimisation problems with constraints and frictions. The learning process we use is inspired by reinforcement learning. There is no oracle that knows the best solution. However, we are all capable of evaluating a proposed solution, while considering the risks involved. This is enough to trigger a learning process and, in due course, discover effective strategies under realistic conditions. By far the greatest challenge here is the construction of adequate training data. It is also essential to control model risk.

Over the past few years, we have successfully developed and implemented a number of innovative solutions for banks, issuers of derivatives, energy and trading companies. Our optimisation projects have focused on treasury departments of banks, trading strategies of investment funds, hedging of derivatives, production plans of hydroelectric power plants, energy storage, and the inventory management of commodities.

Deep Treasury Management for Banks. (2023).

Englisch, H., Krabichler, T., Müller, K. J., and Schwarz, M.
Frontiers in Artificial Intelligence. Vol. 6.

A Deep Learning Model for Gas Storage Optimization. (2021).

Curin, N., Kettler, M., Kleisinger-Yu, X., Komaric, V., Krabichler, T., Teichmann, J., and Wutte, H.
Decisions in Economics and Finance, Vol. 44, pp. 1021–1037.

Hedging Goals. (2023, Preprint 2021).

Krabichler, T., and Wunsch, M.
Financial Markets and Portfolio Management.

A Case Study for Unlocking the Potential of Deep Learning in Asset-Liability-Management. (2023, Preprint 2020).

Krabichler, T., and Teichmann, J.
Frontiers in Artificial Intelligence. Vol. 6.