Forecasters at mid-market power desks spend 40–50% of the shift reconciling SCADA telemetry, ISO settlement feeds, weather, and overnight positions before judgement starts. The model is usually fine. The loss originates in the manual rebuild between forecast, dispatch, and the bid stack.
At most power desks, the forecast lives in one tool, dispatch in another, the trading position in a third, and the bid stack in a fourth. A misfire in one propagates through manual rebuilds, spreadsheet copies, and Teams chats. A day later, a settlement spike or a BSUoS hit appears and gets attributed to “the forecast” — when the loss was actually downstream of it.1
The audit walks back from the cost. Below is the typical attribution we observe across desks once we trace it properly: the model contributes less than a third of avoidable imbalance loss. The rest lives in the reconciliation, the rebuild, and the handoff.
For imbalance-exposed desks, asymmetric loss matters more than average error. The model refit improves MAPE by a few points; settlement exposure barely moves.
Reconciliation + the forecast-to-dispatch handoff + manual bid rebuild together accounted for 76% of avoidable cost across the desks we audited.
UpSynQ holds the single forecast that feeds dispatch, trading, and the bid stack, with asymmetric loss functions per horizon. Your engines stay; the loop closes.
60-minute call with whoever owns forecast and trading. 4-page diagnostic returned within 48 hours, including an imbalance-cost attribution walked back from your last four weeks of settlement.
Live SCADA, or handed-over schedule? In what format, with what latency, and where does it diverge from what dispatch sees?
The attribution above, but for your desk specifically. Built from your last four weeks of settlement data and operational logs.
We trace the path: forecast tool → dispatch → bid stack → settlement. Each step where data is rebuilt, reformatted, or re-interpreted by hand is named.
ERCOT 5-min, BSC half-hourly, ISP 15-min. The cadence determines how loss compounds per cycle.
The customer is a vertically integrated power generator running roughly 2,100 MW of thermal generation across four stations, with captive coal mining feeding the largest. Five workflows lived in five tools: demand forecasting, schedule intelligence, dispatch and procurement, fuel value-chain optimisation, and ash management. None of them talked to the others.
We built the decision OS as a single 96-block engine sitting above the existing systems. Schedule intelligence at the centre, with forecast, dispatch, fuel, and market participation feeding into it and learning from it. The same product spine now ports to merchant generators, battery IPPs, and VPP operators in Western markets.
One of the founders will respond within two business days with available 60-minute slots and a short note on the operational logs we would want to look at together.