Candlestick forecasts for AI agents
LLM agents can't read candles. Kronos can.
We turn Kronos candlestick forecasts into fact lists your agent can use, paid per call with x402.
- Candles are real hourly prices, loaded live in your browser.
- Median path is the middle of Kronos's 30 sampled futures.
- Inner band: half of the paths ended up in here (25th–75th percentile).
- Outer band: 80% of the paths ended up in here (10th–90th percentile).
- Faint lines are the individual sampled paths: one plausible future each.
- Dashed line is the invalidation level, where the up view would be wrong.
Breakdown
What an agent receives for the same run. Built from the cached run file, not live.
Did it work last time?
One example, not a benchmarkA past hold-out run: Kronos-small forecast BTC-USD from Oct 10 16:00 to Oct 11 16:00 UTC, without seeing those 24 hours.
How it works
- 01
Read the candles
Closed candles come from public exchange data. Kronos is an open-source foundation model trained on candlestick data.
- 02
Sample many futures
Kronos draws 30 possible paths instead of one guess, so the spread shows how sure it is.
- 03
Turn paths into facts
We count the paths to get direction, chance up, an expected range, and an invalidation level.
- 04
Your agent pays per call
The API answers
402 Payment Required. The agent pays in USDC, retries, and gets JSON. No account or API key.
What an agent gets
Direction and chance up
direction, prob_up: the share of paths that finish higher.
Expected move and range
expected_move_pct, band_80_pct, expected_range in price terms.
Invalidation level
invalidation_level: the price that says the model was wrong.
Strike probabilities
p_close_above for mapping to Kalshi and Polymarket contracts.
Confidence
confidence, with a note when model sizes disagree.
Full provenance
model, sampling, as_of, and a disclaimer on every response.
Pay per call
Testnet| Endpoint | Returns | Price | Status |
|---|---|---|---|
GET /api/v1/sample?symbol=&tf= | Delayed fact list (2h+ old) | Free | testnet |
GET /api/v1/chart?symbol=&tf= | Delayed per-step percentiles for charting | Free | testnet |
GET /api/v1/forecast?symbol=&tf= | Latest fact list with per-step percentiles | $0.02 | testnet |
POST /api/v1/forecast | Forecast on your own OHLCV (up to 512 bars) | $0.03 | planned |
GET /api/v1/ensemble?symbol= | Two models × two lookbacks, with agreement score | $0.05 | planned |
GET /api/v1/pm-map?market= | Model probability vs Kalshi / Polymarket price | $0.05 | planned |
GET /api/v1/health | Data freshness and worker status | Free | testnet |
GET /api/v1/track-record | Hit rate and band coverage | Free | coming soon |
FAQ
What is Kronos?
An open-source foundation model for financial candlesticks, released under the MIT license by its original authors (Shi et al., arXiv 2508.02739). This service is not affiliated with them.
Is the forecast on this page live?
The candles are live. The forecast overlay is the free delayed run, at least 2 hours old. The newest hourly run is what the paid API returns. The overlay hides itself once its forecast window ends.
Why pay when Kronos is free?
No setup: no PyTorch, weights, or data plumbing. You get cached, consistent runs, clean closed-bar data, and the fact-list layer on top. If you'd rather run it yourself, the skill is free.
How does x402 payment work?
Your agent calls an endpoint, gets 402 Payment Required, pays a few cents in USDC, and retries to get the data. There's no account or API key.
Is there a track record?
Not yet. The track-record endpoint is coming soon. Until then, we show one past hold-out run and label it as one example, not a benchmark.
Is this financial advice?
No. Kronos outputs are probabilistic model estimates for research.
Give your agent eyes on the chart.
Agents pay per call with x402. No account or API key. Read the guide or grab the free skill today.