Almost every price-prediction tool ever built starts from the same blank slate. You pick one asset, gather its history, and train a model from scratch to fit that single series. A new model called Kronos throws that habit out. It asks a stranger question instead: what if a market chart is just another language, and you could pre-train one big model to speak it fluently, the way GPT was pre-trained on text?
That is the whole idea. Kronos is an open-source model released by a group of academic researchers and accepted to AAAI 2026 under the title "A Foundation Model for the Language of Financial Markets." It has been called, only half-jokingly, "the GPT of finance." The label is loose, but the architecture genuinely earns the comparison, and understanding why is a neat lesson in how the whole foundation-model idea travels beyond text.
Predicting the next candle is, mechanically, the same problem as predicting the next word. You just need a way to turn a candle into a token.
The shiftWhat "foundation model" actually buys you
The phrase foundation model gets thrown around, so it's worth being concrete. A traditional quant model is bespoke: trained on one market, useful for that market, useless the moment you change assets. A foundation model is trained once, on almost everything, and then transfers to problems it never saw in training. It is the same reason one ChatGPT can write both a sonnet and a cover letter.
Kronos was pre-trained on roughly 12 billion candlestick records from more than 45 global exchanges: equities, crypto, and beyond. That breadth is the point. By the time you hand it a chart it has never seen, it already carries a general sense of how markets behave: how volatility clusters, how trends exhaust themselves, how a quiet range tends to precede a breakout. You are fine-tuning intuition, not building it from zero.
The clever bitTurning candles into words
Here is the obstacle. Language is already made of discrete symbols (words, characters, subword tokens), so a GPT just predicts the next one from a fixed vocabulary. But a candlestick is continuous: five raw numbers, Open, High, Low, Close and Volume (OHLCV), that can take any value at all. There is no vocabulary to predict from.
Kronos solves this in two stages, and this is the part worth slowing down for:
- A tokenizer turns candles into "words." A dedicated neural network quantizes each candle's continuous OHLCV values into a small set of discrete, hierarchical tokens: coarse ones that capture the broad shape, finer ones that fill in the detail. This is the exact analogue of breaking a sentence into subword tokens.
- A decoder-only Transformer predicts the next token. With candles now expressed as tokens, you can train a plain autoregressive GPT: given the sequence so far, predict the next token, then the next, rolling forward. Decode those tokens back into OHLCV, and you have a forecast candle.
Strip away the finance and it is a language model. The genuine innovation isn't the Transformer, which is standard. It is the tokenizer that makes the market speakable in the first place.
The nuance most people missIt doesn't predict a number. It samples a future.
This is the most important idea in the whole piece. Kronos does not output a single confident price. Like an LLM writing a sentence, it samples, and it exposes the exact same knobs you would recognise from any chatbot:
Temperature (T) controls how adventurous the sampling is. Top-p nucleus sampling trims the improbable tail. And sample_count lets you draw many independent future paths at once, then average them or study how far they spread.
The payoff is that you get a distribution of possible futures, not one brittle line. The width of that fan is itself information, an estimate of uncertainty, which is why the same machinery doubles as a volatility forecaster and a generator of realistic synthetic market data for stress-testing other strategies.
What shipsThe model family
Kronos is a family, not a single model, and it is released under the permissive MIT licence, commercial use included, which matters for anyone who wants to actually build on it. Weights live on Hugging Face; code and a live BTC/USDT demo are on GitHub.
| Model | Parameters | Context | Note |
|---|---|---|---|
| Kronos-mini | 4.1 M | 2048 | Smallest, longest window |
| Kronos-small | 24.7 M | 512 | Lightweight |
| Kronos-base | 102.3 M | 512 | The workhorse |
| Kronos-large | 499.2 M | 512 | Tokenizer weights closed |
By large-language-model standards these are tiny: half a billion parameters at the very top end, against the hundreds of billions in a frontier LLM. Markets carry far less raw information than human language, and it shows in the model sizes.
ResultsThe numbers the authors report
On the paper's own benchmarks, Kronos posts real gains over both dedicated time-series foundation models and from-scratch baselines:
RankIC is the metric to notice. It measures ranking skill (does the model correctly order which assets will outperform which) rather than nailing an exact price. For anything trading-shaped, getting the order right matters far more than the decimal.
Read this twiceThe honest caveats
A model that backtests brilliantly is not a money printer, and it's worth being blunt about the limits:
- A short memory. The small, base, and large models see only 512 candles of context; longer histories get truncated. Its working memory of the past is genuinely limited.
- The benchmarks are the authors' own. Impressive in-paper numbers routinely soften in live, adversarial, out-of-sample markets.
- Markets fight back. They are non-stationary and reflexive: the moment an edge is widely known, crowding erodes it. Regime shifts, slippage, and transaction costs quietly eat theoretical alpha.
- Shape is not profit. Forecasting the plausible shape of price action is a different, easier problem than predicting direction profitably after costs.
The takeawayThe mental model to keep
Kronos is to a candlestick chart what GPT is to a paragraph: a pre-trained, general-purpose next-token predictor. The move that makes it novel is tokenising OHLCV so that the entire language-model toolkit (autoregressive pre-training, temperature sampling, transfer learning) suddenly applies to markets that were previously modelled one at a time.
Whether that cashes out as durable trading edge is the open, and much harder, question. But as a demonstration that the foundation-model recipe generalises beyond text, images, and code, reaching into the strange, noisy grammar of price itself, Kronos is one of the more quietly interesting releases of the year. And for students, it's a clean, real example of the single most transferable idea in modern AI: tokenise it, then predict the next token.
Get the next one in your inbox
One AI tool, one prompt, three quick wins, every week. Free, and under 3 minutes to read.
Subscribe Free →