What Is Kronos?
Kronos is an AI model that learned the language of stock charts. It is the first open-source foundation model built specifically for financial candlesticks, the K-lines that record open, high, low, close, volume, and amount for each period, and it was trained on data from more than 45 global exchanges. Where general time-series models treat a price chart as just another sequence of numbers, Kronos is designed for the high-noise reality of financial data.
It is a Python project under the MIT license, and it had 47,010 stars when we checked on November 6, 2026. The work behind it was accepted by AAAI 2026, and the paper is on arXiv. That academic grounding matters here, because a forecasting model is only as trustworthy as the method behind it, and Kronos publishes its method in full.
How It Reads a Chart
Kronos is a family of decoder-only foundation models, and it works in two stages. First, a specialized tokenizer quantizes continuous, multi-dimensional K-line data, the OHLCV values, into hierarchical discrete tokens. That is the step that turns a chart into a language the model can read: prices and volumes become tokens the way words become tokens in a language model. Second, a large autoregressive Transformer is pre-trained on those tokens, which lets one model serve many different quantitative tasks instead of needing a separate model for each one.
The point of the two-stage design is that financial data is noisy in ways other sequences are not. A tokenizer built for K-lines can separate signal from noise before the Transformer ever sees it, and the pre-training across 45+ exchanges gives the model a broad sense of how markets move, rather than the quirks of a single ticker or a single exchange.
The Model Zoo
Kronos ships as a family of pre-trained models in different sizes, all available on the Hugging Face Hub. Kronos-mini has 4.1 million parameters with a 2,048 context length, Kronos-small has 24.7 million parameters with a 512 context length, and Kronos-base has 102.3 million parameters, also with a 512 context length. Kronos-large, at 499.2 million parameters, exists but is not open source.
The size ladder is practical, not decorative. Mini runs almost anywhere and is fine for trying the workflow, small and base are the workhorses for real forecasting, and the context limit matters: small and base cap out at 512 steps, so your lookback window should stay inside that. The predictor handles truncation automatically if you pass more, but you get the best results by keeping your input inside the limit.
Forecasting in a Few Lines
Forecasting with Kronos is straightforward. You load a pre-trained tokenizer and model from the Hugging Face Hub, create a KronosPredictor, prepare your historical K-line data as a pandas DataFrame with open, high, low, and close columns, volume and amount are optional, and call predict with your lookback window and prediction length. The predictor handles preprocessing, normalization, prediction, and inverse normalization, and returns a DataFrame with forecasted values for the future timestamps you asked for.
The README's worked example loads a CSV of 5-minute K-line data, uses a lookback of 400 periods, and forecasts 120 periods ahead. If your market data arrives as a CSV file, a quick pass through a CSV to JSON converter is an easy way to inspect the structure before you load it into a DataFrame. Going the other way, once you have a forecast DataFrame, a JSON to CSV converter turns the predictions back into a spreadsheet you can open anywhere. There is also a predict_batch method that forecasts many series in parallel on a GPU, and a complete runnable example that plots the forecast against the ground truth.
The Live Bitcoin Demo
The fastest way to judge Kronos is the live demo, which shows a forecast for the BTC/USDT trading pair over the next 24 hours. The README opens on the same kind of chart the demo produces: the model's prediction plotted against what actually happened, so you can see the gap with your own eyes instead of taking a benchmark number on trust.
Beyond the demo, the repo ships fine-tuning scripts so you can adapt Kronos to your own data, with a worked example that uses Qlib to prepare Chinese A-share market data, fine-tune both the tokenizer and the predictor, and run a simple backtest. The example is deliberately a demonstration, not a production trading system, and the README says so plainly.
The Disclaimer That Matters
Both the video and the description say it, and the article should too: Kronos is a research model, not financial advice. Its forecasts are probabilistic predictions from a model trained on historical market data, not recommendations to buy or sell anything. The backtesting example in the repo is a simplified top-K strategy, and the authors warn that a real quantitative workflow needs portfolio optimization, risk factor neutralization, transaction costs, slippage, and market impact modeling before it means anything in production.
That honesty is part of why the project is worth writing about. A model that forecasts prices is easy to oversell and easy to misuse, and Kronos documents its limits alongside its results. Use it to learn how foundation models read markets, to test forecasting ideas, and to build research pipelines. Do not use it as a signal to trade real money.
