FATHOM Forecast. Probabilistic load and generation, from feeder to portfolio.

The model stack is deliberately heterogeneous. Long horizons (D+1 to D+7) lean on a gradient-boosted ensemble with engineered weather features and explicit market-regime classifiers — the right shape of model for problems where the signal is dominated by calendar effects, weather climatology, and regulator-induced regime changes. Medium horizons (1–24 hours) use a temporal convolutional network that handles the diurnal-plus-ramp dynamics of net demand; this is where the rise of EV charging and heat-pump load is most visible.

Short horizons (sub-hour to ten minutes) use a small transformer that consumes raw SCADA at 100ms resolution and emits forecasts at 1-second resolution. This is the head that earns the "sub-second" claim — and the one that, when miscalibrated, will silently degrade every dispatch decision downstream. We retrain it nightly and we publish its calibration report to the customer's reliability inbox at 06:00 local time every weekday.

A meta-learner reconciles the three heads and emits a single calibrated quantile distribution. Calibration targets are CRPS (continuous ranked probability score) and pinball loss at each percentile; both are visible on the customer's metrics dashboard. The reconciliation logic itself is a small linear model — kept deliberately simple — because the meta-learner is the component most likely to be inspected by a regulator post-incident and we want it to be defensible without a PhD in the room.

Across the production fleet (eight live UK assets and two US ERCOT pilots as of April 2026) the day-ahead median-percentile MAPE on net demand is 1.71%, with the worst-asset month at 2.4% and the best at 1.2%. Intraday (1-hour-ahead) median MAPE is 3.39%; the figure that matters more — the pinball loss at P95 — has fallen 18% since 0.8.0 thanks to the new short-head transformer. Calibration plots, available in every customer's monthly report, are flat to within ±1.5% of the diagonal at every percentile.

The number we are most proud of is not a MAPE — it is the percentage of dispatch decisions downstream of FATHOM Forecast that, in retrospect, would have been improved by a better forecast. On the eight UK assets that figure is 4.2%, which sounds high until you realise the industry baseline (a leading point-forecast vendor) is 11.8% measured the same way. The remaining 4.2% is dominated by genuinely unforecastable events — unannounced asset trips, sudden weather mesoscale anomalies — and we are honest about that.

FATHOM Dispatch — LANE-2. Autonomous co-optimisation of battery, demand response and curtailment across multiple revenue stacks. Solves a rolling-horizon mixed-integer stochastic programme in under 1.2s at P99 on a typical 200 MW / 400 MWh portfolio. Consumes the full P5–P95 bands from this module.

FATHOM Sentinel — LANE-3. Asset health and grid anomaly detection trained on substation telemetry, weather, and a 12,400-asset historical fault library including 2,180 confirmed failures. Median 19-day lead time on transformer winding faults across a 14-month UK DNO trial.