19/06/2026
Why ML Dwarfs LLMs in Crypto Patterns Detection.
I wanted to share a project I’ve been building over the last four years, talk about a major pivot we just went through, and get some feedback from the community.
I started building a tool to automate technical analysis for crypto trading. The idea was simple: manual charting takes too much time, and human emotion (revenge trading, fear, greed) blows up accounts. I wanted a bot that could mathematically scan the market for geometric structures (Head & Shoulders, Wedges, etc.) and alert me.
Instead of just drawing a "Bull Flag" on a chart, CoinPattern now uses a Machine Learning model to analyze historical volume and volatility. It grades the probability of a successful breakout. It doesn’t just tell you what the pattern is; it calculates the statistical probability of what happens next to help filter out fake-outs.
My backtests take around 40 minutes to complete in a single month running on Python and around 3 minutes on C++. The biggest issue is getting around the accuracy when running on C++. I'm also facing an issue on my server resources. Python is RAM hungry. In fact, my ML side, which runs the ML agreement by running the hourly crypto price recommendations to be consumed by the Geometric patterns logic is run in batches.
This is a problem for me because an hour ends before all the crypto coin batches have been run. So this means I have turned off my hyper parameter tuning for now. If possible, running the heavy mathematical calculation on C++ alone would save me all this trouble. The challenge is it's inaccurate compared to Python.
So, how would an LLM run these calculations and even use a weeks data/candles to predict. I certainly feel, most LLMs value macro data more than the live data. Well, my risk is 1% and a 2% take profit. My backtests indicate a return of 61%. Live signals had around 41% returns on April. If you'd like to test the software, I can provide the link.
I use a scalping strategy with tight risk management, automatically exiting positions as soon as a strict profit target is met.
I forgot to mention that we’re using WebSockets for the lice Crypto Pattern Detection. This means we expect better returns on the live detection because backtests get the entry and the exit prices on an hourly basis.
[email protected]
https://www.coinpattern.app
