Kronos is an open-source foundation model trained on billions of candlesticks. We asked two practical questions: can it tell which Tokyo stocks will beat the others, and can it warn when the Nikkei is about to get rough? We wrote the rules down before looking at any result, tested it only on months it had never seen, and made it beat simple free rules after trading costs. It did not.
Kronos reads the last 40 daily candles and imagines many possible next 12 days, like a weather forecaster drawing a fan of possible futures. These are real forecasts it made for the Nikkei 225. It opens on 2 August 2024, the Friday before the market's worst day in decades. Slide through the dates and compare its fan with what happened.
A model can look brilliant on data it has already memorised, and a test can be quietly bent after the fact. We guarded against both. Every rule and pass mark went into PREREG.md, committed before any test data was downloaded. Kronos learned from data up to June 2024, so every test uses later months. A final stretch of months stayed sealed, to be opened only if Kronos passed.
Tap a step to read what happened in it.
Hover or tap a bar.
A broken exam can make a bad student look good. Before grading Kronos, we checked the grading code on cases where the right answer is known in advance.
A score of 0 means no predictive power. Whiskers show the margin of error.
Passed. It found nothing in random fake prices, and found a hidden signal of 0.05 when we planted one.
Round trip for a large Japanese stock, in basis points (1 bp = 0.01%).
10 bp kept. Tokyo's minimum price steps put a realistic round trip at 7 to 14 bp, under the 15 bp line that would have forced a change.
Scrambling which stock gets which signal gave zero, as it should. Scrambling which day gets which returns did not: some stocks stay strong or weak for months, and that pattern survives a date shuffle. We recorded this before grading and changed no rule.
If Kronos could see a storm coming, a trader could bet smaller before it hits. We compared its 12-day volatility forecasts for the Nikkei with three textbook formulas over 435 trading days.
Yearly volatility in %. Higher means rougher. Hover to read any day.
Average forecast error (QLIKE). It punishes missing a storm more than a false alarm.
¥1,000,000 in Nikkei micro futures, sized by each forecast. Profit in thousands of yen.
Failed. Kronos lost to every simple formula. It kept underestimating storms, so sizing by it took bigger bets and the deepest drop (−20.5% against −15.4%). The extra profit came from carrying more risk in a rising market, not from skill.
Every trading day from November 2024 to April 2026, Kronos ranked about 500 of Japan's largest companies by expected return over the next 5 or 12 days. We tested six ways of reading its forecasts. To pass, a version had to add information beyond simple price rules, make money after costs once we discounted for trying six versions, avoid being one lucky streak, and work in most months.
Extra return per year from holding Kronos's top fifth of stocks instead of all of them, after 10 bp trading costs.
Daily ranking score (rank IC). Above zero means its ranking matched what happened.
¥100 in its top picks against ¥100 spread over all stocks, after costs.
Failed. Its ranking scores averaged between −0.001 and +0.008, nowhere near significant, although the test could detect scores down to about 0.01 to 0.03. It added nothing beyond simple price rules. Five of six versions lost to simply holding every stock, and the best one broke even. That matches an independent test on US stocks.
Used out of the box on daily candles, Kronos-small gave a trader nothing usable on Japanese large caps or the Nikkei between mid-2024 and April 2026. Simple formulas forecast risk better, and its stock rankings were noise once costs were counted.
We tested the small model without fine-tuning, on daily bars, over about 18 months, through a free data plan. A fine-tuned or larger model, intraday data, or other markets could behave differently. The sealed final exam was never opened, because nothing passed.
Everything except the licensed market data is in the repository. You need Python 3.12 with uv, a free J-Quants account, and a Kaggle account with GPU access. Each step below writes files that the next step checks.
git clone https://github.com/SIMETRAL55/kronos-zero-shot-japan cd kronos-zero-shot-japan uv sync && bash scripts/fetch_kronos.sh && uv run pytest
uv run python scripts/fetch_dev_data.py nikkei uv run python scripts/fetch_dev_data.py jquants --free # needs JQUANTS_API_KEY in ~/.config/kronos-explore/jquants.env
export KAGGLE_USERNAME=your-name uv run python scripts/push_h2_dev.py # H2, about 12 minutes uv run python scripts/build_h1_windows.py && uv run python scripts/kaggle_dataset.py create # then the five H1 jobs listed in README.md, about 23 GPU-hours
uv run python scripts/eval_h2.py h2dev uv run python scripts/eval_h1.py costs && uv run python scripts/eval_h1.py p5 && uv run python scripts/eval_h1.py dev
The free plan's two-year window moves forward every day, so an exact rerun of our dates now needs one month of the Light plan. A rerun on your current free window tests the same idea on slightly different dates.