ERA OF MINDS

Kronos AI trading model: what the 93% claim means, and what happened when we ran it on Bitcoin

By Era of Minds ·

Short answer

Kronos is a free, MIT-licensed model (shiyu-coder/Kronos) that reads candlestick data as tokens and predicts the next candles. It is real, it is small enough to run on a laptop CPU, and the paper was accepted at AAAI 2026. The viral "outperforms every model by 93%" line is a relative gain on one correlation metric (0.0267 vs 0.0138), from a large model you can't download. In our own test, Kronos-small made 40 Bitcoin forecasts 24 hours ahead and was beaten by the simple guess "the price won't change" 27 times out of 40. Use it to learn and experiment, not as a trading signal.

Era of Minds cover: The 93% claim, tested on BTC. Kronos AI model K-line tokenization diagram from the README, next to our 40-forecast BTC/USDT test: Kronos error $2,225 vs no-change $1,052.
Figure: the K-line tokenization diagram from the Kronos README (MIT), next to our own walk-forward test on BTC/USDT hourly candles from Binance. Mean absolute error at 24 hours: Kronos-small $2,225, "no change" $1,052.

What is the Kronos AI trading model?

Kronos is a foundation model built only for price charts. No news, no text. It takes candlesticks (open, high, low, close, plus volume and amount if you have them) and treats them like a language.

It works in two steps. A tokenizer squeezes each multi-number candle into discrete tokens, split into a coarse part and a fine part. Then a decoder-only Transformer, the same family as GPT, is trained to predict the next token. Run it forward and you get a predicted price path. Run it several times and you get several paths, which you can average.

The team pre-trained it on more than 12 billion K-line records from 45 exchanges and 7 time frequencies, per the paper by Yu Shi and six co-authors. It was accepted at AAAI 2026. The repo was created in July 2025 and had about 41,850 stars and 6,940 forks when we checked on October 20, 2026. It is trending again this week because of a viral X post (more on that below).

Four sizes are listed, three are downloadable from Hugging Face:

  • Kronos-mini: 4.1M parameters, 2,048-candle context, Kronos-Tokenizer-2k.
  • Kronos-small: 24.7M parameters, 512-candle context. About 1.19 million downloads in the last month.
  • Kronos-base: 102.3M parameters, 512-candle context.
  • Kronos-large: 499.2M parameters. Not open-sourced.

What does "outperforms every model by 93%" actually mean?

The post that went viral said Kronos "outperforms every model by 93%." The paper says something narrower: Kronos "boosts price series forecasting RankIC by 93% over the leading TSFM and 87% over the best non-pre-trained baseline."

RankIC is a rank correlation between a forecast and what really happened. 1 would be perfect. 0 is no relationship. In the paper's appendix, the average price-series RankIC is:

  • Kronos-large: 0.0267
  • Kronos-base: 0.0258, Kronos-small: 0.0254
  • Time-MoE base, the best other foundation model: 0.0138
  • Best model trained from scratch (TimesNet): 0.0143

0.0267 divided by 0.0138 is 1.93. That is the 93%. So the claim is true as arithmetic, and it is also a gap between two very small numbers. As Satyajit's code-level write-up points out, this "price series" RankIC is measured along a single forecast path for each sample, which is not the cross-sectional stock-ranking number most quants would assume. And the headline score belongs to Kronos-large, the one size you can't download.

The paper reports other wins too: 9% lower MAE on volatility forecasting and 22% better fidelity on synthetic K-line generation. Those are arguably the more useful results.

How do you run Kronos for free?

You need Python 3.10 or newer and no GPU. This is what we did on our box:

git clone https://github.com/shiyu-coder/Kronos.git
cd Kronos
pip install -r requirements.txt

On a CPU-only machine, install the CPU build of PyTorch first (pip install torch --index-url https://download.pytorch.org/whl/cpu) to skip the large CUDA download. Then, from the repo folder:

from model import Kronos, KronosTokenizer, KronosPredictor
tokenizer = KronosTokenizer.from_pretrained("NeoQuasar/Kronos-Tokenizer-base")
model = Kronos.from_pretrained("NeoQuasar/Kronos-small")
predictor = KronosPredictor(model, tokenizer, device="cpu", max_context=512)

pred_df = predictor.predict(df=x_df, x_timestamp=x_ts, y_timestamp=y_ts,
                            pred_len=24, T=1.0, top_p=0.9, sample_count=5)

x_df is a pandas DataFrame with open, high, low, close (and optionally volume, amount). x_ts and y_ts are the past and future timestamps. Keep the lookback at 512 candles or fewer for small and base. sample_count is how many paths it samples and averages. The repo also has predict_batch for many assets at once, a web UI in webui/, and fine-tuning scripts (the Qlib pipeline in finetune/ expects multiple GPUs via torchrun).

For Bitcoin we pulled hourly BTC/USDT candles from Binance's public klines endpoint. With 400 candles of context, a 24-hour forecast with 5 paths took 7.8 seconds on an 8-core CPU.

Can the Kronos AI trading model predict Bitcoin? Our 40-forecast test

A single forecast tells you nothing, so we ran a small walk-forward test. We took 1,999 hourly BTC/USDT candles (July 28 to October 20, 2026, UTC). Starting from 40 points between August 21 and October 19, spaced 36 hours apart, we gave Kronos-small the previous 400 candles and asked for the next 24 hours, averaging 5 sampled paths each time. Then we compared its 24-hour close with what happened, and with the laziest possible forecast: "the price stays where it is."

  • Mean absolute error at 24 hours: Kronos $2,225, no-change $1,052. Kronos was about twice as far off.
  • Beat the no-change guess: 13 of 40 times.
  • Right direction (up or down): 22 of 40, or 55%. With 40 tries, that is well within coin-flip range (two-sided binomial p ≈ 0.64).
  • Size of moves: Kronos predicted an average 24-hour move of 2.3%. The real average move was 1.3%.

That last line explains most of the error. Kronos produces confident, realistic-looking paths that wander further than Bitcoin actually did. On 24-hour closes, a model that over-predicts movement loses to one that predicts none.

Our test is small: one asset, one model size, one 2-month window, hourly bars, 5 paths per forecast. More paths or Kronos-base could change the numbers a bit. But we were not the only ones to look.

What did independent tests of Kronos find?

Every outside test we found landed in the same place:

  • sumit-ai-ml/kronos-market-eval (run August 4, 2026, Kronos-base): on 100 BTC/USDT 24-hour forecasts, accuracy was 118% worse than the no-change baseline and direction was 50 of 100. On 60 SPY sessions it was 42.5% worse. Its uncertainty band covered the real price 21% to 43% of the time, where 80% would be honest. Fine-tuning lowered token loss and made forecasts worse.
  • Florinel Chis on Medium: 749 forecasts across 25 forex and CFD markets, using only data from after the training cut-off. Direction was right 50.2% of the time, and the trades lost money before costs.
  • Mercurio (June 12, 2026): scored 376 trend-following stock entries from a two-year window against Kronos's 24-hour forecasts. Information coefficient -0.05 (p = 0.29), directional hit rate 42.8%. Used as a veto gate, even blocking only the worst-scored 10% of trades cut the two-year return from +53.1% to +15.9%.

None of these say the model is broken. They say that asking it "will this one asset go up tomorrow?" is the wrong job. The Kronos README agrees: its demo signals are "raw predictions" that would normally go into a portfolio optimisation model, with costs, slippage and risk management on top.

Is the Kronos live BTC demo real?

Yes, but it is stale. The viral post mentions a "live BTC demo." When we opened the demo page on October 20, 2026, it said Last Updated (UTC): 2026-07-04 11:00:26, so the forecast shown is more than three months old. The page explains its method: Kronos-mini, the last 360 hours of BTC/USDT 1-hour candles from Binance, and 30 Monte Carlo paths, turned into an "upside probability" and a "volatility amplification" probability for the next 24 hours.

It is a good way to see what the output looks like. It is not a live trading feed, and as the tests above show, those probabilities were not well calibrated.

What is Kronos actually good for?

  • Learning how tokenized time series work. The tokenizer is the clever part, and the code is short enough to read in an afternoon.
  • Volatility and scenario generation. The paper's volatility and synthetic-data results are where it beats baselines in ways that matter for risk work, such as stress-testing a strategy on realistic fake price paths.
  • A feature, not a signal. If you build quant models, Kronos output can be one input among many, tested out of sample with costs.
  • A baseline to beat. If your own model can't beat Kronos and a no-change forecast, you learn something cheaply.

What it is not: a free hedge-fund model that prints money. If that is how a post sells it, check the numbers.

We have seen this pattern before. A viral claim, a real tool behind it, and a gap between the README and outside tests. It was the same story with Headroom, where 39% fewer tokens did not always mean a smaller bill. Our Gemini 4 Argon "I must not do that" explainer traces another viral line back to its source. For more small open-source models and tools that run on ordinary hardware, see earshot (CPU-only audio for coding agents) and ManimGX.

Why is Kronos trending on X again?

Kronos is not new. The paper came out in August 2025. What changed is one post. @quantscience_ posted on October 18, 2026 (12:13 UTC, 16:13 Dubai) that someone "built an AI that reads candlestick charts the way GPT reads English." When we checked on the morning of October 20 (Dubai), it had about 128,340 views, 1,680 likes, 3,942 bookmarks and 184 reposts. A @RoundtableSpace follow-up on October 19 had about 46,500 views. A wave of copy posts followed in English, Persian and Spanish, plus replies from people who tested it and came back unimpressed.

The bookmark count is the tell. Nearly 4,000 people saved it to try later. If you are one of them, the setup above takes a few minutes. Just run your own walk-forward test before you trust a single forecast.

This article explains a research model. It is not financial advice.

Key facts

  • Model: Kronos, an open-source foundation model for financial candlesticks (K-lines). Tokenizer plus decoder-only Transformer.
  • Repo: shiyu-coder/Kronos. MIT. Created July 1, 2025. About 41,850 stars / 6,940 forks (checked Oct 20, 2026).
  • Paper: arXiv 2508.02739, Yu Shi et al., accepted at AAAI 2026. Pre-trained on 12B+ K-line records from 45 exchanges.
  • Sizes: mini 4.1M, small 24.7M, base 102.3M (downloadable on Hugging Face); large 499.2M not released.
  • The 93%: average price-series RankIC 0.0267 (Kronos-large) vs 0.0138 (Time-MoE base).
  • Run it: Python 3.10+, pip install -r requirements.txt, KronosPredictor.predict(...). CPU works: 7.8 s for a 24h, 5-path BTC forecast in our run.
  • Our BTC test (40 forecasts, 24h, Kronos-small): error $2,225 vs $1,052 for no-change; beat no-change 13/40; direction 22/40.
  • Live demo: Kronos-mini, 360h context, 30 paths; last updated 2026-07-04 UTC when checked Oct 20.
  • X buzz: @quantscience_ post Oct 18 ~128,340 views / 1,680 likes / 3,942 bookmarks (checked Oct 20).

FAQ

What is the Kronos AI trading model?

Kronos is an open-source foundation model for financial candlesticks (K-lines), from the shiyu-coder/Kronos repo on GitHub. A tokenizer turns open, high, low, close, volume and amount into discrete tokens, and a decoder-only Transformer predicts the next tokens, the way a language model predicts the next word. It was pre-trained on over 12 billion K-line records from 45 exchanges. The paper was accepted at AAAI 2026. It forecasts price paths; it is not a trading bot.

Is Kronos really 93% better than other models?

Only on one metric, and in relative terms. The paper says Kronos boosts price series forecasting RankIC by 93% over the leading time series foundation model. In the appendix that is an average RankIC of 0.0267 for Kronos-large against 0.0138 for Time-MoE base. Both are very small correlations, and Kronos-large is not publicly released. It does not mean Kronos is right 93% of the time.

How do I run Kronos for free?

Clone github.com/shiyu-coder/Kronos, install Python 3.10+ and pip install -r requirements.txt, then load NeoQuasar/Kronos-Tokenizer-base and NeoQuasar/Kronos-small from Hugging Face and call KronosPredictor.predict with a DataFrame of open, high, low, close (volume and amount optional) plus past and future timestamps. It runs on a normal CPU. On our 8-core box, a 24-hour BTC forecast with 5 sampled paths took 7.8 seconds.

Can Kronos predict Bitcoin prices?

Not well enough to trade, in our test. We ran Kronos-small on 40 BTC/USDT 24-hour forecasts from August to October 2026 using Binance hourly candles. Its mean absolute error was about $2,225, against $1,052 for simply assuming the price would not change. It beat that no-change guess 13 times out of 40 and got the direction right 22 times out of 40, which is close to a coin flip. Other independent tests on BTC, SPY, forex and US stocks found the same pattern.

Which Kronos model should I use: mini, small or base?

Kronos-mini has 4.1M parameters and a 2,048-candle context (it uses Kronos-Tokenizer-2k). Kronos-small has 24.7M and Kronos-base 102.3M, both with a 512-candle context and Kronos-Tokenizer-base. Kronos-large (499.2M) is listed but not open-sourced. Small is the most downloaded on Hugging Face, with about 1.19 million downloads in the last month when we checked on October 20, 2026.

Is the Kronos live BTC demo up to date?

No. When we opened shiyu-coder.github.io/Kronos-demo on October 20, 2026 it said Last Updated (UTC): 2026-07-04 11:00:26. The page uses Kronos-mini, the last 360 hours of BTC/USDT 1-hour candles from Binance and 30 Monte Carlo paths, and shows an upside probability for the next 24 hours. Treat it as a picture of how the model works, not a live signal.

Is Kronos AI a trading service or a scam?

The shiyu-coder/Kronos project is MIT-licensed code and model weights. It does not take deposits, manage money or sell signals. If a website asks you to fund an account for Kronos AI trading, it is not this research project. The repo README itself says its signals are raw predictions that need portfolio optimisation, cost modelling and risk management before any real use.

Sources

Keep reading