Kronos (shiyu-coder/Kronos, AAAI 2026) is a decoder-only foundation model pre-trained on more than 12 billion candlesticks from 45+ exchanges. A tokenizer quantises each OHLCV bar into discrete tokens and an autoregressive transformer continues the sequence, so a forecast is a sampled future rather than a point estimate.

This service draws many samples in one batched pass and keeps them all. The median across samples is the headline number, the 10th-90th percentile becomes the shaded band, and the candles drawn are the single sampled path whose terminal return sits closest to the median — a future the model actually produced, not an average of several.

Known bias, and why the lookback is 128 bars

Kronos z-scores each context window before tokenising, and it mean-reverts within that window. Walk-forward testing on this deployment measured a correlation of about −0.75 between how stretched the last close sits inside the window and the return the model predicts. Feed it 400 daily bars of a bull market and it forecasts a decline almost every time.

So the lookback was tuned rather than assumed. Over eight US large caps at eight origins each, on a five-day daily horizon. These hit rates come from that narrow sweep and are used only to rank the four windows against each other — the headline accuracy figure is the broader measurement further down, which is much lower:

128 bars is both the most accurate and the only effectively unbiased setting, so it is the default. Indicators such as SMA200 and the 52-week range are computed over a much longer window than the model ever sees.

The band is widened on purpose

Kronos' raw samples are badly over-confident. Left alone, a 63-bar Bitcoin forecast came back spanning eight percentage points — a quarter of BTC inside a range narrower than a normal fortnight. So every path's deviation from the median is scaled up, bar by bar, until the fan matches the volatility that symbol has actually been realising. The median path is never moved, so this costs no accuracy.

Measured over 10 symbols at 10 origins each, on daily bars:

A true 10th-to-90th percentile band would cover 80%. Calibration roughly doubles coverage where the raw fan was worst, but still falls short at long horizons — so the band is better than it was, not perfect.

Do not trade the BUY/SELL label

Directional accuracy measured between 50% and 56%. That is a coin flip. An earlier, narrower sweep suggested 66% and it did not survive a broader sample, which is exactly why the number above is the one published here. The composite verdict is useful as a summary of what the model saw, not as a prediction; its confidence figure is capped just above the measured hit rate for that reason.

Treat the spread as the only genuinely useful output. This is research tooling, not investment advice.