Can AI predict stock prices accurately?

No model predicts prices with certainty. Foundation models like Kronos produce probabilistic forecasts — a distribution of plausible future candlesticks — which is useful for framing risk and scenarios, not for guaranteed outcomes.

What does zero-shot forecasting mean?

Zero-shot means the model forecasts a ticker it was never fine-tuned on. Kronos was pre-trained on OHLCV data from 45+ global exchanges, so it generalises to new symbols directly from their recent price history.

What data does the forecast use?

Kronos Forecast pulls live OHLCV candles for the selected symbol and timeframe, or you can upload your own CSV of open, high, low, close and volume bars.

AI Stock Price Prediction

Most "AI stock predictors" fit a small model to one ticker and hope the pattern holds. Kronos takes a different route: it is a financial foundation model for candlesticks, pre-trained on OHLCV data from 45+ global exchanges, that forecasts a symbol it has never been fine-tuned on. This page explains how that works — and what a probabilistic forecast can and cannot tell you.

Run a forecast View live demo

How the model forecasts price

1 · Pre-training on candlesticks

Kronos is trained on OHLCV sequences from 45+ global exchanges, so it learns the general grammar of price action — trend, volatility clustering, mean reversion — rather than one ticker's quirks.

2 · Zero-shot inference

You feed in the recent bars for any symbol. The model tokenises that history and continues the sequence, no per-ticker training or retraining step required.

3 · Probabilistic paths

Sampling temperature and top_p control how wide the distribution is. Instead of one line, you get multiple plausible futures plus a p10/p90 confidence band.

Why zero-shot matters for stocks

Training a bespoke model per ticker is slow, data-hungry, and prone to overfitting a single regime. A foundation model amortises that work: the price dynamics it learned across thousands of instruments transfer to a new symbol from its recent history alone. In practice that means you can point the terminal at an equity, ETF, or crypto pair and get a forecast in seconds, with the same sampler settings across all of them.

Three model sizes are available — Kronos-mini (4.1M parameters), small, and base (102.3M). Larger models capture more structure at the cost of inference time.

Reading a probabilistic forecast

The median path is the model's central expectation, not a target price.
The p10/p90 band shows how uncertain the model is. A wide band on a volatile symbol is information, not a defect.
Temperature and top_p widen or narrow the sampled distribution. Low temperature gives a tighter, more conservative fan.
Horizon length compounds uncertainty — accuracy decays the further out you project.

Common questions

Can AI predict stock prices accurately?: Not with certainty, and any tool promising that is misleading you. What a model like Kronos gives you is a distribution of plausible futures, which is useful for framing risk and scenarios — never a guarantee.
What data does the forecast use?: Live OHLCV candles for the symbol and timeframe you select, or your own CSV of open, high, low, close and volume bars.
Is it free to try?: Yes — the free plan includes 8 forecasts per day on Kronos-mini. Compare plans.

Try it on your own ticker

Pick a symbol, choose a timeframe, tune the sampler, and project the next candles.

Open the terminal GitHub Paper

Research preview. Forecasts are not financial advice.