FINANCIAL MODELLING
Financial analysis is likely to involve the use of spreadsheets (Excel being the most well-known), some playing a significant role in an investment or financing transaction (business or security valuations, merger models, LBOs, project finance, hedging etc). Many will be time based, others tables and databases. This section will focus on the former.
A.I. now plays a major part in providing model support, often embedded in Excel as an add-in but also building models from scratch (as is discussed in an article in this section) - Codex lists 350 Excel models it says it could produce in the areas of finance, accounting and tax. Whether A.I. models follow best practice and have the most efficient formulae needs to be assessed.
FINANCIAL MODELLING: Building an Excel model with A.I
EXCEL DEVELOPMENTS
Financial Modelling I - Building an Excel Model with A.I. (GPT 5.3 and Sonnet 4.6)
This article tested how good two A.I. models released in February 2026 (GPT 5.3 / Codex / OpenAI and Sonnet 4.6 / Claude / Anthropic) were at building or replicating a Convertible bond pricing Excel model (see Appendix in the article ‘Security Valuation – Convertible Bond pricing’). The exercise, carried out in late February 2026, approached the testing in two steps:
- Prompting the bond price assumptions and parameters to be used and asking for an Excel model (in the hope the calculated price was the 115.00 calculated by the 'user')
- Uploading the Excel model and asking for an identical Excel model (the A.I. models used Python in both tests).
After the first prompt, the output was human-reviewed and questions or amendments given, before a second attempt etc. A second run was carried out with all the prompts in the first run combined. Codex reached the target 115.00 price in both tests sooner with minimal chat, although Claude’s thinking was more ‘interesting’ to observe.
A number of observations can be made from this surprisingly entertaining exercise:
- Prompting from the user has to be accurate and complete: any subsequent prompt can take the LLM in the wrong direction and lead to bloated formulae
- Prompting needs to be read completely by the LLM: quite often simple instructions were ignored and were only followed after a prompt.
- Python output didn’t always agree to Excel output, even though that was a requirement
- Excel output cells did not always have embedded formulae, but hard coded values from Python (despite this being requirement in all prompts)
- Best practice Excel model design is not always followed:
A follow up exercise would be to check their formulae, since they do seem rather long in places!