Kronos: Forecast the next price candles from past market data

What it does

Kronos is a family of models trained only on financial candlestick charts, also called K-lines: the open, high, low, close and volume of each time step. A tokenizer first turns those numbers into discrete tokens (small symbols a model can read), then a Transformer trained on those tokens predicts what comes next. The creators say it was trained on data from over 45 global exchanges. The mini, small and base sizes are published on Hugging Face; the large one is not.

How to use it

Input → action → output: load a model and its tokenizer from Hugging Face, create a KronosPredictor, and pass it a pandas table of past open, high, low and close prices with their timestamps, plus the future timestamps you want. It returns a table of forecast open, high, low, close, volume and amount for those future times.

Common use cases

• Give it the last few hundred hourly candles of a trading pair and get a table of predicted candles for the next day.
• Run the batch prediction method on price histories for several assets at once and get one forecast table for each.
• Fine-tune it on your own market data with the released scripts, then compare its forecasts with what actually happened.