Skip to content

Tycheon

Kronos forecasts the path; Tycheon tells you how much to trust it.

Tycheon is the open-source calibrated forecasting and risk layer for financial time series: conformal prediction intervals you can check, exogenous covariates beyond OHLCV, regime-weighted routing across Kronos, TimesFM, Chronos and honest baselines, translation of forecasts into VaR, Expected Shortfall and drawdown probabilities, and leakage-proof walk-forward evaluation with a public leaderboard.

v0.1.0 - evidence so far is synthetic

Conformal calibration with measured holdout coverage, a regime-weighted ensemble, a risk layer and a leakage-guarded walk-forward evaluation are built. On the bundled synthetic series no model is distinguishable from the random walk, and the leaderboard says so. Multi-asset portfolio risk is always uncalibrated. See calibration, risk and the benchmark methodology.

Kronos-small beside the random-walk baseline

Start here

Page What is in it
Methodology How calibration, routing, risk translation and evaluation are meant to work, and what we refuse to claim.
Safety and compliance Point-in-time discipline, no live execution, untrusted external text, data licensing.
Model cards One card per forecaster: source, licence, training data, limitations, calibration status.
ADR 0001 Why Apache-2.0 plus a proprietary ee/, and where the line sits.
ADR 0002 Why Kronos is vendored, and why its sampler is replaced.
ADR 0003 The two-clock data model and the forecast contract.

Install

pip install tycheon                  # core: baselines, calibration, risk, backtest
pip install "tycheon[kronos]"        # Kronos + torch
pip install "tycheon[timesfm]"       # TimesFM
pip install "tycheon[chronos]"       # Chronos
pip install "tycheon[serve]"         # FastAPI + MCP server

Development setup, the dev stack and the contribution gate are in CONTRIBUTING.md.

The three invariants

Everything in this project follows from these.

  1. Point-in-time. Every data read takes an as_of and refuses anything published after it. Leakage tests are mandatory for every data path.
  2. Uncertainty is not optional. Every forecast carries intervals or quantiles, calibration status, the model mix that produced it, its as_of and a model card reference.
  3. Honest baselines. Every evaluation reports the random-walk baseline and a Diebold-Mariano test, published whichever way the result goes.

What Tycheon is not

  • Not a consumer trade-signal product.
  • Not personalized investment advice.
  • Not a market-data vendor. Providers are pluggable and you bring your own data licence; Tycheon never redistributes licensed exchange data.
  • Not a live execution system. v1 is paper and simulation only, and only through Keelgate-governed tools.

For research and risk analytics. Not investment advice. Nothing produced by this software is a recommendation to buy or sell any security. Past calibration does not guarantee future coverage.