Short answer: Jake Trefethen says multimodal models do a good job of visual price action analysis from chart images, even with indicators, though they struggle with time frames. Jordan Metzner says more thinking time trades off against trade speed.
Watch this answer from Episode 8 (starts at 11:41) · Episode 8 notes and transcript
From Episode 8: Can Public Make You a Better Investor?, starting at 11:41.
Jake Trefethen [11:41]: No. Yeah. It's it's really interesting. I guess, curious to hear whether you tried out any kind of multimodal prompts with ChatGPT. So, did like, sending a picture of the price chart versus sending the metrics. Yeah. So, in experiments with this that I've done, something that's really interesting on your question around the model is if you use a multimodal model, you can actually send it a picture of the price chart and even multiple time frames of the price chart, and it does a really good job. The models have improved a lot over the last couple years, but it does a super good job actually of doing, like, visual price action analysis. Interesting. Like, you can even overlay indicators and whatnot, and it can actually process pretty efficiently. And so, yeah, like, that combined with something like a more agentic model, like a Sonnet three five or like Sonnet four, like, thinking type model that kinda Mhmm. Puts a lot of tokens behind its decisions is kinda would be kinda interesting for sure. As, like, an easy next step.
Jordan Metzner [12:51]: Well, that's that's a trade off. Right? Like, if you wanna give the model more time to think, then you trade off kind of how fast you can run trades. So Yeah. It really would depend on your strategy of how, like, frequently you wanna trade. I tried to build something that was, high frequency, I guess, the sense of, like, for me, every fifteen, thirty seconds is pretty high frequency. But, yeah, if we were just trading once a day or even, every hour or something like that, you might be wanna give a model, five, ten, fifteen minutes to make a decision and to really, like, think deeply. Yeah. One question about the charts. Does it work the same if you give it, like, the JSON based history of the price action? Because one thing I did was the my first build, I just gave it, like, how much money I had. And then I started to improve it by giving it my trade history, which kinda gave it some influence of, like, where I've bought it before.
Jordan Metzner [13:41]: It doesn't have I didn't give it full price history, but to your point, I understand like, yeah, I guess it can understand like, kind of the dips and valleys and whatnot.
Jake Trefethen [13:46]: Yeah. Definitely giving it context of any kind helps. Right? Like, definitely, when I played with this, it had a lot of trouble with time frames. So dialing in and out of, are we on a two minute here or two hour? But, yeah, I mean, giving it any historic context is definitely helpful. And, I mean, I think I saw something today that one of the models I think maybe it's a Claude model is now like a million token context window. So it they can handle us an insane amount of historic context, multiple time frames all in one shot now. So something to think about there too. Loading up
Jordan Metzner [14:22]: Yeah.
Jake Trefethen [14:23]: Pretty much every time frame.
Jordan Metzner [14:24]: Yeah. To your point, yeah. I wonder like how much more data I don't know where the diminishing returns are, but how much more data you can give it on the price behavior to get more insight, depending on your trade strategy as well. Right?
Jake Trefethen [14:36]: Yeah. I mean, I've heard like a quick rule of thumb is somewhere around 75% of the stated context windows when you get degradation. But just for like frame of reference, the Great Gatsby is like 60,000 tokens. So a million is a lot.
Jordan Metzner [14:51]: I wonder if I could just do like a Great Gatsby trade strategy, you know. Yeah. Okay. Cool. Any other questions maybe from Sam, Emily, Jake about kinda what I built here?
