Simple Uses for LLMs – Preparing for ALCO

The fifth instalment of our series exploring simple GPT-4 use cases focuses on its capacity to generate documentation for an ALCO committee. This article delves into a specific scenario involving a small bank or building society, and how GPT-4 can assist in creating insights into interest rate risk.

Our approach begins with a basic spreadsheet showing different time buckets, the corresponding assets and liabilities, the net gap position, and two flat yield curves at 5% and 5.01%.

Using this data, the discount factors from the yield curves are used to calculate the net profit or loss resulting from a 0.01% interest rate shift, namely this is a simple delta report, a common tool in banking for managing interest rate risk.

In this example, we see a bank with long-term assets funded by short-term liabilities, creating both interest rate exposure and potential liquidity risk.

Initially, I asked GPT-4 to analyse the spreadsheet. However, the results were unsatisfactory. GPT-4 struggled with the Excel format, producing an incorrect and misleading narrative.

Adopting a different strategy, I copied rather than uploaded the relevant section of the spreadsheet directly into GPT-4's prompt window. This action generated an image of the spreadsheet and displayed the numerical data in the chat. After removing the numbers from the chat and leaving only the image, I requested GPT-4 to analyse it.

Despite this modification, the process remained challenging due to inconsistencies in the spreadsheet's formatting, such as unclear labelling and the negative numbers being in brackets. After improving the spreadsheet's consistency and including negative numbers rather than brackets, I initiated a new session with GPT-4. This time, the result was more coherent.

Progressing further, I requested GPT-4 to create bar charts and relevant commentary for the ALCO committee. Although time-consuming, this exercise yielded a workable method for engaging GPT-4 effectively.

From this experience, you can see that GPT-4 can act as a co-pilot but it still needs a lot of interaction so it is not an automatic process. However, if you use it in conjunction with its text generation capability, preparing relevant and useful information is achievable faster. It would not surprise me if, within 12 months, given the improvements in LLMs the process could be automated leading to a very large jump in productivity.

Key Points

1. Consistency of data uploaded to the LLM: GPT-4 does not deal well with any slight inconsistency in the data format, for example, if there are blank cells, unaligned rows or columns or bracketed numbers. This contrasts with using text in GPT-4 where it can work with, and make sense of, relatively jumbled ideas, poor syntax and grammar to come up with very good working documents. When LLMs can do this with numbers we will have a step change in the ability to use them in finance.

2. Handling Spreadsheets with GPT-4: GPT-4 exhibits better performance with images of spreadsheets rather than the files themselves, why this is, I don’t know. Is it due to it being a language-based model?

3. Uploading the image without a prompt: this led GPT-4 to analyse what it received before I initiated the prompting, it seemed to slow the model down giving it time to think about the information rather than just trying to complete the prompt and in this limited example the subsequent output was better as a result.

4. Process Efficiency: The procedure of using GPT-4 for such tasks can be initially tedious. However, mastering the process is a valuable skill, promising time savings for future repetitions.

5. The Necessity of Human Oversight: Similar to the early days of spreadsheet usage, there's a risk of assuming AI-generated content is infallible. Vigilance is essential to identify potential errors. GPT-4 can generate plausible but potentially incorrect charts and commentary. Therefore, human validation is crucial to ensure accuracy before using the output in reports.

6. Evolution of LLMs: GPT-4's current capabilities significantly surpass those of its predecessors, suggesting a near future where AI can seamlessly turn data into meaningful analyses and reports. This is where you just upload the data and the charts and written reports are generated automatically for the whole ALCO deck.

7. Benefits of LLMs in Routine Tasks: By handling routine analytical work, AI models like GPT-4 free up human resources to focus on complex decision-making and risk evaluation through collaborative discussions.

The dialogue with GPT-4, detailed below, provides insight into the challenges and solutions of integrating LLMs into financial documentation and analysis tasks.

----------------------------------------------------

Me:

GPT-4: You've uploaded an image that appears to contain a table with financial data related to asset and liability management, including time buckets, asset and liability amounts, net gap, interest rates (Curve 1 and Curve 2), and net profit and loss (P&L).

This data seems to be part of an asset-liability matching or a gap analysis, which is typically used by financial institutions to manage interest rate risk. The table includes different time buckets ranging from "overnight" to "10 years". In each time bucket, the assets and liabilities are compared, and the 'Net gap' row shows the difference between them. Below that, the table shows two different interest rate curves and the net P&L calculated from these rates, which indicates the financial impact of the interest rate changes on the net gap over time.

If you need any specific analysis or have questions about the data, feel free to ask!

Me: please produce a bar chart of the net gap by time bucket

GPT-4:

Me: now a bar chart of the P&L by time bucket

GPT-4:

Me: Now the commentary for an ALCO meeting, this should explain the interest rate risk

GPT-4: ALCO Meeting Commentary on Interest Rate Risk

Introduction: Our current asset-liability position indicates varying degrees of interest rate risk across different time buckets. The net gap analysis, alongside the interest rate curves, offers a clear view of our exposure.

Net Gap Analysis: The net gap distribution shows significant short-term liabilities exceeding assets in the "overnight" to "6-12 months" buckets, indicating a potential liquidity risk if interest rates rise. Conversely, from "1 year" to "10 years," we observe a surplus of assets over liabilities. This longer-term positioning could be beneficial if interest rates decrease but poses a risk if rates increase, as our interest income would be adversely affected while paying out at higher rates on liabilities.

Interest Rate Curves and P&L Impact: The first curve (Curve 1) represents a baseline scenario with current interest rates, while the second curve (Curve 2) represents a slight upward shift in rates. The net P&L across time buckets shows that as the maturity extends, the impact of an interest rate rise becomes more pronounced, culminating in significant potential losses, particularly evident in the "2 year" to "5 year" buckets.

Short-Term Concerns: In the immediate term, we must address the liquidity risk due to the negative gap. Any unexpected increase in short-term rates could increase our cost of funding significantly, which is not fully offset by corresponding increases in asset yields.

Long-Term Strategies: For the longer-term, our current position could benefit from a hedging strategy to mitigate the risk of rising rates. This might include entering into interest rate swaps or options to lock in current rates for future liabilities or to convert fixed-rate assets to floating rates.

Conclusion: It is imperative to re-evaluate our asset-liability matching strategy to ensure it aligns with our risk appetite and market expectations. While we have some cushions due to the positive gaps in longer maturities, the potential for interest rate hikes suggests a more conservative approach might be warranted to protect our net interest margin.

Recommendations:

1. Increase the proportion of floating-rate or short-duration assets to reduce the negative gap in the short term.

2. Consider derivative instruments for hedging long-term interest rate exposure.

3. Continuously monitor the interest rate environment and adjust our ALM strategy accordingly.

This commentary would need to be tailored with more specific details from the financial institution’s own data and strategic positioning, but it provides a framework for discussing interest rate risk at an ALCO meeting.