Imagine how ChatGPT learned English — by reading billions of sentences until it understood patterns, context, rhythm. Kronos did the same thing, but instead of sentences it read candlestick charts. Billions of them, from 45 stock and crypto exchanges around the world.
The result is a model that has developed something like an intuition for how prices move. Point it at a market and ask what happens next — it has a guess, and a pretty good one.
In independent tests, it outperformed the previous best forecasting model by 93% on predicting price direction. There's even a live Bitcoin forecast running right now, updating every hour.
Until recently, this kind of tool only existed inside hedge funds with teams of quantitative analysts. Kronos is free, open, and available to anyone.
If you run a business that touches financial markets — a fintech product, a trading tool, a treasury management service — this is the kind of foundation you could build on without starting from scratch.
Even if markets aren't your world, the story here is bigger: specialist AI models trained on specific domains (not just general knowledge) are getting very good, very fast.
Candlestick / K-line — A compact visual summary of how a stock or crypto moved during a time period: where it opened, closed, peaked, and dropped. Traders have used them for centuries.
Foundation model — An AI trained on a huge dataset that can then be adapted to many specific tasks. Think of it as a very educated generalist who then specialises.
Zero-shot — When an AI performs well on a task it was never explicitly trained for. Like hiring someone who's never worked in your industry but picks it up immediately.
MAE (Mean Absolute Error) — A way of measuring how wrong a forecast is on average. Lower is better.
If you're building anything in fintech or investments, it's worth knowing this exists: https://github.com/shiyu-coder/Kronos