Why regime, not point forecast.

Procurement and trading desks hit a wall on the same thing: a stable, calibrated signal about whether the distribution around today's price just shifted. Amanoki publishes that signal — a four-state regime label on every 30-minute JEPX bar, and a forecast distribution over those states at 60, 120, and 240 minutes with its held-out Brier score in the same response. This page is the one-screen answer to "why this, in this form".

Point forecasts self-destruct

A perfect short-term price prediction, priced and sold, gets arbitraged away the moment enough customers act on it. That is the self-referential trap in the point-forecast form: success narrows the edge, the narrowing edge churns customers, and churn pushes the vendor toward more aggressive claims. A regime label and a spike probability don't collapse that way — both are summary statistics of market behaviour, not directional trading signals.

Regimes are commercial, not statistical

The four states (low / normal / high / scarcity) map to procurement and trading actions the way price levels don't. A scarcity bar on a cold winter morning moves a retailer's monthly procurement cost more than ten forecasted prices do. The FSM stabilises the scarcity boundary (absolute ¥/kWh threshold) and the volatility boundary (hysteresis over a z-score ladder) in a way that smooths single-bar noise without lagging too far behind real transitions.

Calibrated probability beats direction

regime-forecast returns P(regime_{t+H}) — a full distribution, not a top-pick prediction. The same response carries the held-out Brier score and the unconditional baseline Brier so callers can judge calibration for themselves.

Weather is the exogenous lever

Current weather (temperature, wind, irradiance, cloud cover, humidity) drives the regime model alongside the FSM's own lagged state. Weather doesn't react to market participants, so conditioning on it doesn't self-destruct. As more weather signals (Open-Meteo today, higher-resolution NWP later) land, the same endpoint shape absorbs them — callers don't rewrite against a v1 to v1.5 model upgrade.

What you're paying for when tiers open

Rate limit, archive depth, and early access to new forecast families. The free tier and paid tiers return the same content for any request both are allowed to make, so if a paid-tier customer publishes their Brier score, a free-tier caller can verify it against the same endpoint.