A builder shares a concept for a deterministic, node-based graph system that LLMs can both understand and construct for complex time-series financial calculations, arguing that black-box LLM math output is untrustworthy for finance agents. The project is still a proof of concept, with longer-form deep dives planned for community feedback.
TLDR: you can't trust a large language model to build complex time-series calculations. Exploring a graph-like system for modeling events that an LLM can both understand and build. This video has a lemonade stand example with a couple of variables. Concept video link is in the comments. ------- So here's the thesis: no one would trust a large language model to run complex time-series financial calculations years in the future. The challenge right now is that a simple calculation like 5 × 4 is not something that an LLM actually calculates. It just has that as a pattern stored in memory. The next order is something like a mortgage calculation, where it's probably spinning up a little code block. There's no guarantee it does it the same way every single time, and there's always the chance that it gets the code wrong. Slim, for something as simple as a mortgage calculation, but a possibility And then there's the next level, which is complex models, things happening at different dates, and a bunch of dependencies inside of there. Could be something as basic as a software-as-a-service model with customer cohorts, different pricing tiers, and churn rates. And even if it gets it right, most of the time, the results are what are getting stored. The actual steps that it took to build that become lost over time. You, as the user, can't easily plug into that ecosystem to validate the calculations. It's a trust exercise, and that's why you're not seeing, in my opinion, a lot of successful AI agents for finance. If anything, they're spinning up Excel models as its source of truth, but these models lacked context. It doesn't know why you put a 5 in this cell or a 26 in that cell, or what it represents other than the label beside it. Years ago, I started building a node-based graph-type system for running financial calculations. Dumped a ton of my own money into building it out, and while it was technically solid, the learning curve was a barrier for anyone who was interested. The cool thing now is that an LLM can easily understand core concepts, and if you give it an infrastructure it can use to model these scenarios, then you've got a pretty big unlock. I'm still deep in proof of concept right now, but I've been building this in a bit of a bubble. It's time that I start sharing it more, getting feedback, and figuring out where my blind spots are. Going to be doing some longer-form videos and deep dives into my process, but wanted to just get some content out there and start sharing my ideas. Concept video link is in the comments.
The author built an AI trading agent but found that using LLM for direct trading decisions is unstable, leading to a deterministic system where LLM translates strategies into explicit rules for execution.
Summarizes a deterministic, constraint-based approach for building AI agents in regulated finance, where the LLM only generates prose, numbers are cryptographically sealed, and auditability is ensured through separated layers.
This paper reviews and audits execution realism in LLM-based trading research, proposing clearer reporting standards for reproducibility and evaluation comparability.
The article discusses various techniques to make LLM-powered agents more deterministic, such as golden sets, guardrails, consensus mechanisms, regression tests, coded logic, and hyperparameter tuning, and asks for additional successful methods.