Insightconsultance.ai

Kronos Review in 2026: The Open Source Financial Market Foundation Model

9 min read

Muhammad AliMuhammad AliFounder, consultance.ai
Kronos Review in 2026: The Open Source Financial Market Foundation Model

Kronos is an MIT-licensed foundation model for financial candlestick data, with three free model sizes from 4.1M to 102.3M parameters that run locally. Independent out-of-sample tests in 2026 across crypto, forex and Indian equities found little or no directional signal after costs, so it is best used as a research baseline or a synthetic path generator, not a trading strategy.

What this covers

  • Kronos is a decoder-only model pre-trained on over 12 billion K-line records from 45 exchanges.
  • Code and the mini, small and base weights are MIT licensed, so commercial and private use is allowed.
  • Kronos-large (499.2M parameters), behind the paper's best result, is not released.
  • Running it costs compute, data and staff time; there is no hosted API or vendor fee.
  • Independent 2026 tests found direction accuracy near chance and no profit after trading costs.
  • It has little use for private equity or private credit teams, whose assets do not produce candlesticks.

Kronos is an open source foundation model for financial candlestick (K-line) data, published by Yu Shi and colleagues at github.com/shiyu-coder/Kronos under the MIT License. It forecasts open, high, low, close and volume series and runs locally from free Hugging Face weights, but independent out-of-sample tests published in 2026 find little or no tradeable directional signal, so treat it as a research tool and not a strategy.

That is the short answer. The rest of this review covers what the model is, what the license lets you do, what it really costs to run, what the outside testing shows, and where it fits for a fund, family office or deal team. Every fact below comes from the project's own repository, paper and model pages, or from independent write-ups, all fetched on 7 November 2026.

What Kronos is

The repository describes Kronos as "a family of decoder-only foundation models, pre-trained specifically for the 'language' of financial markets." The design has two stages. A tokenizer first turns continuous OHLCV bars into hierarchical discrete tokens. An autoregressive Transformer is then pre-trained on those tokens, the same way a language model learns to predict the next word.

According to the paper on arXiv (2508.02739, submitted 2 August 2025), the pre-training corpus holds over 12 billion K-line records from 45 global exchanges. The README notes the paper was accepted at AAAI 2026, and that fine-tuning scripts were released on 17 August 2025. The repository was created on 1 July 2025. When we checked the GitHub API it showed 47,178 stars and 7,799 forks, with the last push on 13 April 2026.

The project also ran a live demo that forecast BTC/USDT over the next 24 hours. When we loaded it, the page read "Last Updated (UTC): 2026-07-04 11:00:26", so the demo has not refreshed for about four months.

Hands turning the pages of a tabbed diligence document under a desk lamp

AI for finance teams

Screen more deals without adding analysts

First system live in about six weeks. You own the code.

Book a discovery call

What it actually does

In practice Kronos is a Python library with one main entry point, the KronosPredictor class. You hand it a pandas DataFrame with open, high, low and close columns (volume and amount are optional), the historical timestamps, and the future timestamps you want filled. It returns a DataFrame of forecast candles. The predictor handles normalization and truncation for you.

The model is probabilistic. Parameters for temperature (T), nucleus sampling (top_p) and sample_count let you draw several future paths and average them. A predict_batch method runs many series in parallel on a GPU, provided they share the same lookback and forecast length.

The paper claims more than point forecasts. Its abstract reports a 93% improvement in price forecasting RankIC over the leading time series foundation model, 9% lower error in volatility forecasting, and a 22% improvement in the fidelity of synthetic K-line sequences. Synthetic data generation is a fair use to keep in mind, and we come back to it below.

The model family

Four sizes exist. Three are downloadable.

ModelParametersContext lengthReleasedHugging Face downloads (as fetched)
Kronos-mini4.1M2,048Yes, MIT515,921
Kronos-small24.7M512Yes, MIT2,493,366
Kronos-base102.3M512Yes, MIT792,283
Kronos-large499.2M512NoNot public

One detail matters for anyone reading the headline numbers. An independent analysis on luminalog.com points out that the paper's best price forecasting result came from Kronos-large, which is the one size you cannot download.

The license and what it allows commercially

The repository's LICENSE file is the standard MIT License, copyright 2025 ShiYu. The Kronos-mini, Kronos-small and Kronos-base model pages on Hugging Face are also tagged MIT. MIT permits commercial use, modification, redistribution and private use, provided the copyright and license notice travel with the code. It comes with no warranty.

For a fund that means you can run Kronos inside your own environment, fine-tune it on proprietary data and keep the result private. You do not owe the authors anything beyond the notice. The license does not give you any assurance about accuracy, and the README is explicit that its fine-tuning pipeline "is not a production-ready quantitative trading system."

What it costs to run

There is no vendor bill. The weights are free, there is no hosted API, and nothing you send leaves your machine. The real costs are compute, data and people.

Compute

The README asks for Python 3.10 or later, and the dependency list is short: PyTorch 2.0 or later, pandas, numpy, einops, safetensors and the Hugging Face hub client. The released models are small by current standards, at 4.1M to 102.3M parameters. An independent review on andrew.ooo describes Kronos as small enough to run on a laptop. The raysize.eu backtest ran eight full walk-forward runs in about six hours on one consumer graphics card.

Fine-tuning is heavier. The repository's scripts are written for multi-GPU training with torchrun, in two stages: the tokenizer first, then the predictor. Plan on GPU hours for that step, either your own hardware or rented cloud capacity.

Data

Kronos does not ship with market data. For inference you bring your own OHLCV bars. The fine-tuning example uses Microsoft's Qlib with Chinese A-share daily data, and the README says that for other sources "you will need to adapt the data loading and preprocessing logic." If your desk already licenses clean bar data, this cost is close to zero. If not, the data is the largest line item.

People

The README also warns that many code comments in the finetune/ directory were generated by an AI assistant and "may contain inaccuracies," and recommends treating the code as the source of truth. Getting from the demo pipeline to anything usable needs someone who reads PyTorch and knows how to build an honest backtest.

A shelf of neatly stacked bound reports in raking side light

Build library

See what we have already shipped

Working systems, the stack behind them, and what they cost to run.

Open the library

What independent testing shows

This is the section that matters most. The paper's results are in-sample benchmarks chosen by the authors. Several people have since tested Kronos on data from after its training period, and the results are consistent.

Crypto, five-minute to one-hour bars

The raysize.eu study ran a walk-forward test on Binance spot BTC, ETH and SOL from 1 September 2025 to 31 August 2026. Kronos-small showed real but small skill on the next five minutes of Bitcoin: a rank correlation of +0.086 over 1,456 forecasts and a 53.3% hit rate. That skill was gone by 30 minutes. The average five-minute move was 0.094%, against a spot round trip cost of 0.2% to 0.5%, and 0 of 43 simulated trading cells made money after fees. The larger Kronos-base did no better.

Forex and CFDs, four-hour bars

A Medium write-up by florinelchis ran 749 out-of-sample forecasts across 25 forex and CFD markets from June to August 2026. Direction calls were right 50.2% of the time, with a 95% confidence interval of 47% to 54%. Simply assuming every market would rise was right 54% of the time. The forecasts where the model's sampled paths agreed most strongly were no better than the rest.

EURUSD hourly

Quant for Free reviewed a test of 15,121 scored predictions on hourly EURUSD. Kronos's error was higher than a "no change" forecast at every horizon from 1 to 16 bars. The reviewer flags the limits fairly: one instrument, two months, synthetic volume, and hourly FX data that may differ from what the model was trained on.

Indian equities, cross-sectional ranking

Rohan Patil tested whether Kronos could rank stocks in the NIFTY 200. The published recipe showed no measurable cross-sectional signal, and about 70% of the model's score came from reversion toward its own normalization window. A faint signal in the hidden state did not beat a random-weights control.

What the maintainer says

The author has been direct about this. In GitHub issue #33, opened in August 2025 after a user reported that forecasts lean toward reversal, the maintainer replied that the model's signal behaves like other data-driven statistical models, "exhibiting features of inversion/regression," with losses coming mainly from trending markets. In a follow-up comment, written in Chinese, he described Kronos as being at roughly the "GPT-2" stage of financial pre-training, without post-training or reasoning, and said it covers only the prediction step of a quant pipeline. Signal filtering, entry and exit rules, style constraints and tail risk management still have to be built.

Issues #299 (May 2026) and #325 (June 2026) remain open with further user reports. None of this means the paper is wrong. A RankIC improvement can be real and still sit too close to zero to survive trading costs, which is the point luminalog makes about the 0.0267 versus 0.0138 comparison behind the 93% headline.

Who should use Kronos

  • Quant researchers who want a pre-trained baseline for candlestick data and a feature generator to test against their own signals.
  • Risk and model validation teams who want synthetic but realistic price paths for stress scenarios. The Medium test found the paths "look like real price action" even when direction was wrong, and the paper reports gains on synthetic data fidelity.
  • Teams learning the method. The codebase is small, the license is permissive, and the fine-tuning pipeline is a clear worked example of applying the language model recipe to market data.

Who should not

  • Anyone looking for a trading signal out of the box. No independent test we found shows a profit after costs.
  • Private markets teams. Kronos needs liquid, frequently priced OHLCV series. A buyout fund's portfolio companies, a private credit book or a real estate allocation do not produce candlesticks, so the model has nothing to read.
  • Teams without someone to own the backtest. The independent studies all show how easily evaluation choices move the result. Without that discipline, an attractive equity curve is easy to produce and hard to trust.
A quiet corner of a modern office at dusk with city light through the window

Implementation services

Put AI into the workflow you actually run

Diligence, close, reconciliation, reporting — in your environment.

See how it works

Where it fits for a fund, family office or deal team

For a hedge fund or a family office running a liquid book, Kronos is a reasonable addition to a research stack. It is free, local and auditable, which matters when data cannot leave the building. Its likely value is as one input among many, or as a source of simulated paths, and not as the signal that decides trades.

For private equity, private credit and M&A teams, the honest answer is that Kronos solves a problem they mostly do not have. The expensive work on a deal team is reading data rooms, tying numbers across documents, building models from messy financials and writing the memo. Our notes on AI for financial due diligence and AI for private equity firms cover where language models do earn their keep in that workflow. A family office with both a liquid sleeve and direct deals may find a narrow use for Kronos in the first and none in the second; our page on AI for family offices walks through that split.

When an in-environment build is the better answer

Kronos gets one thing right that many vendor tools do not. It runs inside your own environment, on your own data, under a license that lets you keep everything you build. The gap is everything around the forecast: data pipelines, validation, the controls a compliance team will ask about, and the workflow that puts a result in front of a decision maker.

That is the work consultance.ai does. We build custom finance AI systems inside the client's environment, the first system is live in about six weeks, and the client owns the code. Sometimes an open model like Kronos is a component. More often, for deal teams, the bigger return is in document and diligence workflows, which is what our AI deal desk systems are built for. If you want to talk through whether a build makes sense for your team, you can book a call, or browse the free prompt packs in the consultance.ai library first.

Sources

Bring one workflow to a 30-minute audit.

Book an audit