Re: [Quantlib-users] Has anyone tested ChatGPT or FinGPT for QuantLib functionality

[Ben W. <ben...@ma...>]

I am looking at this as a concept. Taken a bit of a look at FINGPT but from what I have seen it is just a training set for finetuning open source LLM’s on finance related q&a. Good for sentiment analysis, financial news etc.  Mostly done on smaller models. BBG GPT is closed source and from what I understand is not setting the world of fire…. They have basically held it back to keep researching. BBG are also quite famous for their vapor-were  - where they try and kill the incentive for innovation elsewhere.


GPT-4 is not bad at coding Quantlib with a bit of guidance – I built a pretty good wrapper for FX options in very quick time as well. GPT4 has a good broad base financial product understanding and can make inferences that get you a long way. But the real benefit of GPT4 is when you understand what you are asking and GPT does the grunt work in generating code. If you know what the answer should look like, then GPT4 is really useful. 

But I suspect the question that you are asking – can we use natural language to execute Quantlib code reliability and consistently. This part I have been doing a lot thinking about and requires chunk of infrastructure wrapping around Quantlib. Aside from a high level abstraction above the Quantlib layer, you will need a combination things good prompt engineer,  RAG (retrieval augmentation generation) and that requires vector databases and good documentation,  a finetuned model that understand the instructions to create a properly formatted query.  

 

One of the big  issues is that prompting and RAG is about providing context to  the query and the LLM’s that we can run locally have a limited context windows. To get a longer window, we need bigger models and now you are getting into a hardware issue. Multi-agent framework might be able to solve some of these issues. The other big idea that is not really fully explored is the idea of meta programming where the output from an LLM returns functions that can be executed – this is adding custom vectors to the output  that translate to function calls to do certain things. Right now GPT4 is the only platform really doing this, but is not great and not customisable. 

Regards

 

Ben

 

From: sanfranciscofogg <san...@gm...> 
Sent: Thursday, October 26, 2023 12:27 AM
To: Qua...@li...
Subject: [Quantlib-users] Has anyone tested ChatGPT or FinGPT for QuantLib functionality

 

All,

   After going through the FinGPT documentation, it occurred to me that QuantLib equivalent output might result from a properly formatted query.

   The code which might be output from such a query, was the specific interest.

    Are there any participants of FinGPT who might offer commentary?

 

Cordially,

S.F.Fogg

   

      ★ San Francisco Fogg ★    

               Mercury Algorithmics

               Fogg Spirits

               ecurie Foggio