Kronos: an open foundation model for candlesticks, used as a second opinion

Repo: shiyu-coder/Kronos · MIT

Most AI models in my market pipeline write sentences. Kronos is different: it reads price bars. It's an open-source foundation model for financial candlesticks — K-lines — trained on OHLCV data from more than 45 exchanges.

Research and education, not investment advice.

How it works

Kronos has two parts. A specialised tokenizer turns open, high, low, close and volume into discrete tokens — the price-bar equivalent of words. An autoregressive Transformer, pre-trained on those tokens, then predicts the next candles. A KronosPredictor class handles normalisation and sampling, returning forecast bars with future timestamps.

The released sizes range from a 4-million-parameter "mini" model up to about 100 million parameters, small enough to run on ordinary hardware.

Where it sits in my pipeline

After the nightly scan, Kronos produces a short forecast for each candidate and the results are stored alongside the scan. An enrichment step then re-scores each signal's confidence: when Kronos's direction agrees with the setup, the signal gets a small bonus; when it disagrees, the note says so. The forecast level is drawn on every generated chart, so a reader sees the model's view next to the bands and targets.

Why it's a nudge, not a vote

• Forecasts are samples, not facts. Temperature and sampling settings change the answer; the same input can produce different paths.
• The scanner's edge is thin already. The stock-level backtest showed a modest profit factor. Adding a model that sounds confident can make a weak signal feel stronger without making it better.
• It needs its own backtest. "Kronos agreed" should be tested like any filter: does it remove more losers than winners on archived signals?

So Kronos adjusts confidence a little and adds context. It never promotes a setup on its own.

Worth trying if…

…you have price data and want a learned baseline to compare your rules against. A foundation model that disagrees with your indicator is useful information either way — as long as you measure which one was right.