Large language models learned to write by predicting the next word in huge amounts of text. A newer family of models applies the same idea to markets: they learn to predict the next candle from huge amounts of price history. Kronos is one of the most prominent open-source examples, and SignalEye uses it as an independent check on its chart-reading AI. This guide explains how such models work, why their output is a set of possibilities rather than a single forecast, and what they can't do.
From language models to "market language" models
A candlestick chart is a sequence: each candle has an open, high, low and close price, plus volume. Researchers noticed that this sequence can be treated a bit like text. If you convert each candle into a token, a discrete symbol, you can train a Transformer, the architecture behind modern language models, to predict which token comes next.
General-purpose time-series foundation models already existed, but they were mostly trained on data like electricity demand, traffic or weather. Financial candles behave differently: they're noisy, their volatility changes abruptly, and the relationship between open, high, low, close and volume carries information. That's the gap Kronos was built to address.
What Kronos is
Kronos is described in the paper "Kronos: A Foundation Model for the Language of Financial Markets" (arXiv 2508.02739, 2025), and its code and several model sizes are published as open source. Key points from the paper:
- Training data. It was pre-trained on more than 12 billion K-line (candlestick) records from 45 exchanges, covering many asset types and timeframes, not only crypto.
- Specialised tokenizer. A dedicated tokenizer turns each multi-dimensional candle (OHLC prices and volume) into discrete tokens organised in a coarse-to-fine hierarchy, so the model can represent both the rough shape and the finer detail of a candle.
- Autoregressive prediction. A decoder-only Transformer then predicts future tokens one step at a time, which are decoded back into candles.
- Several tasks. The authors evaluate it on price forecasting, volatility forecasting and generating synthetic price series, and report improvements over earlier general time-series models on their benchmarks.
It's worth being precise about what this means. Benchmark improvements measure forecasting error on historical test data. They're not a claim that the model makes money trading, and the paper doesn't make that claim.
Why sampling many paths matters
Ask a forecasting model for "the" next twelve candles and it will give you an answer, but markets are noisy enough that any single predicted path is very likely wrong in detail. The more honest approach is to treat the model as a generator of plausible futures.
Because Kronos predicts tokens probabilistically, it can be run many times on the same history, each run sampling a slightly different path. The collection of paths says something more useful than any single one:
- If most sampled paths end higher, the model leans bullish; if most end lower, bearish.
- If the paths split roughly evenly, the model has no real view, which is valuable information in itself.
- The spread of outcomes gives a sense of expected volatility.
SignalEye uses exactly this. For each candidate setup it samples many future paths over a fixed number of candles and measures the share that end in the signal's direction. Only when the chart-reading model and a sufficient majority of the sampled paths agree is the signal sent. The current horizon and agreement threshold are shown on How it works.
Why combine it with a chart-reading AI?
The two models look at the market in fundamentally different ways:
| Vision AI | Kronos | |
|---|---|---|
| Input | An image of the chart with indicators | Raw candlestick numbers |
| Strength | Context: levels, structure, pattern quality | Statistical regularities learned from billions of candles |
| Typical failure | Misread prices, over-interpreted patterns | No notion of "why"; weak when conditions are new |
| Output | A trade plan with reasoning | A distribution of possible future paths |
Requiring agreement between two independent methods filters out some setups where one of them is simply wrong. It also means fewer signals overall. That's a deliberate trade-off: fewer, better-supported setups instead of more noise. For the chart-reading side, see how AI analyzes crypto charts.
The limits of price forecasting
Any honest discussion of AI forecasting has to cover what it can't do.
Markets change
Models learn from the past. When market structure shifts, through new participants, new regulations, a crash or a mania, patterns that held in the training data can stop working. Statisticians call this non-stationarity, and it's the central difficulty in financial prediction.
Price-only models are blind to events
Kronos sees candles, not news. A hack, a listing, a macroeconomic surprise or a large liquidation cascade can move price in ways no candle history anticipates.
Good forecasts aren't automatically profitable trades
A forecast can be directionally right and still lose money if the stop is too tight, the move too small to cover fees, or the timing off. Profitability depends on the whole trade plan and on risk management.
Agreement isn't certainty
If 75% of sampled paths point up, that's the model's view under its learned assumptions, not a 75% real-world probability. Model probabilities aren't guaranteed to be calibrated to actual outcomes.
Edges get competed away
Open-source models are available to everyone. Any simple, persistent pattern they reveal is likely to be exploited, and to fade, over time.
How to think about AI forecasts as a trader
- Use them as a filter, not an oracle. They're most useful for saying "the evidence doesn't support this trade".
- Prefer distributions to point forecasts. A single predicted price is false precision.
- Judge by out-of-sample results. What matters is how signals performed after they were published. That's why we publish every closed signal on our track record.
- Keep risk fixed. No forecast justifies a larger position than your risk rules allow.
Summary
Kronos shows how far forecasting has come: a single open model, trained on candlesticks from dozens of exchanges, can produce plausible future price paths for many markets. Used as one input among several, sampling many paths and requiring agreement with an independent chart reading, it makes a signal pipeline more selective. It doesn't make markets predictable, and anyone who says otherwise is overselling the technology.