The shape of price
For the adviser who keeps their own financial model and already has the level.
We do not redistribute market data — realised prices belong to ENTSO-E, FX rates to
the National Bank, and storage spreads we already publish free, in full.
What nobody has and we do produce is the shape: captured price and value factor
by technology and zone, the cannibalisation profile over the horizon. We do not transfer the
data, we transfer the result computed from it — the inputs are public anyway, and we
attribute them. The level is bought from Aurora or AleaSoft; the shape is bought from
nowhere.
Not sold until published coverage reaches target on the series being sold. That is a
calibration matter, not a calendar one. And a client's settlement data never enters
this — it is theirs, with its own threshold, at 15 plants.
Operations
Forecasting and nomination support, per MW installed. It runs in parallel and with
no effect: the model writes what it would have proposed, to a journal — it
submits nothing, zero operational risk. We publish no price for something we have
not demonstrated.
What we know so far, measured: over 16 months of backtest the effect came out between
−8,209 and +30,169 RON — an interval that crosses zero. The rule we follow is that
the decision threshold is written before the run, not after the result; that
threshold is not written yet, and we say so rather than imply it exists.
What we believed, and what we measured. In July we blamed the loss on stale output
data: the monthly settlement note arrives 46 days old on average, and
auto-calibration on recent output looked like the main lever. We then measured, over 19
months and 552 days, the same model with and without it: the revenue difference is
−11 ± 28 RON per day one day ahead and +7 ± 24 two days ahead — zero, on
both horizons. So near-real-time output is not the gate for the forecast; it
remains useful for plant monitoring, which is a different thing. What blocks Operate is
the sample: a single month at the real gate, negative, and a decision threshold that must
be written before the next run.