Strategic Proposal · Confidential

AHOY × Edgecom Energy

Turn signals into decisions, and decisions into action.

From a hand-tuned model that beats IESO once — to a self-improving platform that keeps improving your model, continuously, across your whole fleet.

Ontario IESO Forecast CorrectionProof of Value → Fleet-Wide Deployment
ForCTO & CEO, Edgecom Energy
Prepared14 June 2026
The Thesis

You already proved you can beat IESO — by hand, once, for Ontario.

AHOY makes that automatic, continuous, and fleet-wide.

We are not selling a better forecast. You build excellent models. We sell the self-improving platform around your model — one that keeps improving it as conditions drift, retargets to every asset without new ML hires, and proves every change is safe and explainable.

What we measured in your own artifacts

We opened your model & data — here's the ground truth

SignalWhat we found
Your modelStandardScaler → XGBoost, 267 features, 966 trees @ depth 4, Optuna-tuned, leak-aware residual design. Mature and competent.
IESO baseline today2.0% MAPE, MAE 321 MW — good, but systematically under-forecasts ~250 MW, structured by hour (HE07 +552 MW, evening peak HE17 +388 MW).
Where it hurtsPeak-hour P95 error 930 MW; under heat waves P95 ≈ 1,031 MW — the tail blows out in the costliest hours.
Frozen eval window
Jul 1–14, 2025		IESO MAE 423 MW (2.27%), RMSE 538, P95 ≈ 1,019 MW. Data clean — 6 zeroed rows, no nulls.

Computed directly from best_model.pkl + ontario.parquet — exact figures, not claims.

The Opportunity

The high-value target is the heat / peak P95

Where the tail error reaches ~1 GW — the hardest, costliest hours, where even a well-tuned single model struggles and an autonomous search has the most room.

IESO systematic bias
−250 MW
Consistent under-forecast. Structured by hour → highly learnable (11.3% of residual variance from hour alone).
Peak-hour P95 error
930 MW
Evening peak (HE17–21) is where error and operational cost concentrate.
Heat-wave P95 error
~1.0 GW
Tail blows out in extreme heat — the headline target for AHOY's autonomous loop.

Honesty note: we did not fabricate a "model beats IESO by X%" figure — that needs your feature/eval code. The same features (residual_roll_*) carry leakage risk we'd audit as part of the loop.

The Seven Questions — answered

What AHOY brings, in one line each

  1. Most value: autonomous, continuously-improving model loop — fleet scale + drift resilience + audit-grade governance.
  2. Edge vs your stack: you run a human + Optuna on one XGBoost; we search features + HPs + model families, with a leak-safe eval gate + champion/challenger + explainability.
  3. POC proves: from your champion + data, AHOY autonomously yields a challenger with lower weighted RMSE — peak & P95 no worse — and explains what it changed.
  4. Data we need: your feature/eval code + champion scores, a 2nd asset/region, cost-per-MW of forecast error, edge/MLOps constraints.
  5. Pilot trigger: a reproducible, guardrail-intact improvement over your champion on ≥2 windows, verifiable on your harness, mapped to dollars.
  6. Scaling path: POC → paid pilot → enterprise SDK with fleet-wide rights; value compounds as per-asset savings multiply across the fleet.
  7. Risks: headroom, eval leakage, overfit, build-vs-buy, commercial — each with a mitigation; we stage price to proven value.
How the platform keeps improving your model

Not a bigger model — the dimensions a single tuned model leaves on the table

Where a single tuned model leaves value
  • Wrong loss. Optimizes mean squared error, not the weighted-peak + P95 KPI → we train on the real objective (peak-weighted, tail-aware).
  • Tail under-served. ~1 GW heat/peak P95 → monotonic constraints + regime specialists for extreme heat.
  • −250 MW bias. A calibration layer recovers it — our no-ML correction already cut error −26% on the July window.
What the platform adds
  • One family → many. It searches & ensembles (LightGBM / CatBoost / temporal) — diverse errors cut peak variance.
  • Point → probabilistic. Calibrated P95 intervals, and it optimizes the guardrail directly.
  • Static → self-improving. Retrains on drift through the champion/challenger gate — keeps improving as conditions change.

Honest framing: single-model average-RMSE gains are modest. The improvement is on the operational KPI (weighted-peak + P95), on a 2nd out-of-window regime, provably leak-safe, and produced autonomously — not a one-shot accuracy miracle.

The Ask

Three decisions to move this forward

The operational layer for Physical AI — perceive, decide, execute.

AHOY × Edgecom Energy
1 / 7