Treating every hour as an average hour throws away the structure. SPECTER learns the handful of regimes your grid actually lives in.
It groups your grid's behaviour into a few real regimes rather than treating every hour the same. Full-covariance regime clustering.
Regimes don't change at random; there's a pattern to the transitions. SPECTER learns it. Hidden-Markov regime dynamics.
What it learns at one node bootstraps a new site that behaves like it. Regime embeddings transferable across sites.
The current regime tells FENRIR how hard to hedge and TARSIS which pattern to expect. Feeds regime context to FENRIR and TARSIS.
Our validation methodology is a walk-forward backtest against standard baselines with Diebold–Mariano significance testing — designed to run on your own history during onboarding so you can see how it performs before you rely on it.
These are the real methods behind SPECTER, each in one plain line. We name the method as a credibility signal — we don't publish the recipe.
Learns your grid's distinct behaviour regimes, not just an average.
Learns how the grid moves between those regimes.