Financial institutions such as life insurers or pension funds have a fiduciary duty to manage long-term commitments to their customers. They initially collect premiums/contributions from their clients, and in exchange promise to provide a guarantee or a higher return at a further date. A large part of risk management in such companies is focused on making sure that the invested assets are managed in such a way as to guarantee the honouring of all commitments (pensions, annuities, insurance claims etc.). One such tool is an asset-liability model. Such models try to simulate three things simultaneously:
- Evolution and performance of company's investments
- Evolution of liabilities
- The impact on profitability by different trading strategies in response to outflows and inflows throughout time
The aim of this project is to write an Asset-Liability model that uses agent-based modelling concepts and LLMs to extract insight on the company's portfolio. All technologies used are open-source and widely used (Pyhon and specific packages such as Pands, Numpy, Datetime etc., Ollama).
Ultimately, OSEM should be able to take as input:
- Description and size of investments
- Description of economic factors such as interest rates, credit spreads etc.
- Description and size of outflows
- Description of the factor influencing the size and timing of outflows
- Description of planned new business that a company wishes to include into the projection
OSEM should be able to produce a projection of what would happen to the balance sheet of this company in 5, 10, 20, 50 years.
The development of the OSEM model is still very much an ongoing project. However, a draft of the methodology document can be found in the Documentation folder as OSEM pdf or a OSEM Jupyter notebook.
Specific deep dives into topics related to the methodology are available as Jupyter notebooks or pdf-s. The topics covered so far are:
- Yield-curve calibration/projection Term structure example
- Equity pricing Equity pricing example
- Fixed income pricing Bond pricing example
- Summary charts Summary charts example
So far, we have produced a simple example that can be ran using the script main.py. A hypothetical portfolio of 3 equities and 3 corporate bonds and a single fixed liability cash flow profile. This example will grow as more development is committed to the main branch.
To run the example:
- Download the OSEM repository
- Open the Anaconda command prompt
- Make the OSEM repository the root folder
- Run the script main.py
- In the folder Output, look at the cash flow time-series
A simple summary of what goes into the run and what are the outputs can be found in the Summary pdf
OSEM uses a simple LLM agent for trading decisions. In particular we are excited to try mimicing human decisionmakers in the future versions. This was inspired by for example in Generative Agent Simulations of 1,000 People. This slows the model considerably since each decision must use the LLM inference call which is slower than deterministic logic. To make sure the agents run, first run llama3.2 using Ollama.
If you use copilots to further develope the code, we have started assembling a modelingfile that can be used to refine the prompts. The file is located in the llm_modelfile folder.
Send us an email at gregor@osmodelling.com with feedback, comments, ideas on what we could do better etc. Do you want to help us on this open source project? Check our wiki page also on GitHub Wiki
