Methodology

This page states what Tycheon intends to do and, just as importantly, what it refuses to claim. It was written before the implementation, in Phase T0, so that later phases can be held to it, and it is updated as they land. Phase T1 built the point-in-time data layer and the forecasters; Phase T2 built calibration, the covariate residual corrector, routing and the risk layer (see calibration and risk); Phase T3 built the walk-forward evaluation and the leaderboard (see the benchmark methodology). Every section below says which is which.

Why this layer exists

A time-series foundation model gives you a trajectory. On its own that is not usable for risk:

• a point forecast has no stated uncertainty, and a model-reported interval is usually not calibrated out of sample;
• OHLCV-only inputs ignore the news, fundamentals and macro context that moves prices;
• one model is rarely best across regimes;
• a forecast is not a risk number until it is translated into one;
• and a backtest with any lookahead in it is worse than no backtest, because it is confidently wrong.

Tycheon addresses these five in order: calibration, covariates, routing, risk translation, evaluation.

Calibration

The intervals Tycheon publishes are conformal, built on a held-out calibration set drawn strictly before the forecast as_of. Target coverage is explicit (for example 90%), and realised coverage is reported alongside it.

What gets published with every forecast:

Field	Meaning
quantiles / interval	The predictive distribution, not a single number.
calibration_status	calibrated, stale or uncalibrated, with the window used.
coverage_observed	Realised coverage on the most recent evaluation window.
model_mix	Which models contributed, with weights.
as_of	The timestamp the inputs were restricted to.
model_card	Reference to the card for every contributing model.

A forecast whose calibration cannot be established is labelled uncalibrated. It is never silently presented as if it were calibrated.

Financial series break exchangeability: volatility clusters and regimes shift, so split conformal gives only approximate coverage. Tycheon treats this as a first-class problem rather than a footnote, and reports reliability diagrams and coverage-over-time rather than a single headline number.

Covariates

News, fundamentals and macro series enter as exogenous features, each carrying its own publication timestamp. The rule is absolute: a feature may only use information published at or before the forecast as_of. Restatements are handled point-in-time, which means the value a feature carried at that moment, not the value the vendor shows today.

All external text is untrusted data. It is parsed for features; it is never followed as instruction.

Routing

Different models win in different regimes. Routing detects regime (volatility state, trend persistence, liquidity) and weights the ensemble accordingly. Weights are fit on data before as_of only, and the mix is published with the forecast so a user can see which model spoke.

Baselines are not decoration here: random walk, drift, seasonal naive, ARIMA and GARCH sit in the same registry as the foundation models and compete on the same terms.

Risk translation

From a calibrated predictive distribution Tycheon derives:

• VaR at configurable levels,
• Expected Shortfall (the average loss beyond VaR),
• drawdown probability over a horizon,
• stress scenarios (historical and hypothetical shocks),
• portfolio aggregation with correlation structure.

Each number inherits the calibration status of the forecast it came from. A VaR derived from an uncalibrated forecast is labelled as such.

Evaluation

Walk-forward, with:

• expanding or rolling windows and an explicit embargo between train and test so labels cannot leak backwards;
• point-in-time inputs per fold, re-resolved at the fold test-window as_of;
• a cost model (spread, slippage, fees) applied to anything expressed as a strategy;
• the random-walk baseline reported every time, with a Diebold-Mariano test on the forecast-error difference and the test statistic published, not just a verdict;
• multiple-comparison discipline: the more configurations we try, the more we must discount the winner.

We publish results when the baseline wins. A leaderboard that only shows wins is marketing, not evaluation.

Status (v0.1.0): all of the above is implemented except the multiple-comparison discount: p-values are not adjusted for how many models and configurations were tried, and the benchmark methodology lists that as a limitation. Read isolated wins with suspicion.

What Tycheon does not claim

• Not that calibrated intervals make a forecast accurate. Calibration is about honesty, not skill.
• Not that past coverage guarantees future coverage. Regime change breaks coverage, which is why it is monitored rather than assumed.
• Not that any output is a trading recommendation. Tycheon is research and risk analytics infrastructure, and not investment advice.