Blanc Quant Services · Forecast Assurance

Before you adopt an AI forecast, prove that it beats the one you already use.

Blanc Quant Services independently evaluates a candidate forecasting system against your incumbent, under rules frozen before results are known. The incumbent is mandatory, the evidence is reproducible, and our fee is independent of which model wins.

This is a design-partner engagement, not a market-proven product. We are opening one Forecast Assurance Decision Review for a utility with a live forecasting adoption or renewal decision. The method is demonstrated publicly below; the commercial track record starts with the first engagement.

Discuss a Forecast Evaluation

01 · The costly decision

The decision this protects

A utility is about to buy, renew, or deploy a third-party forecasting system — load, price, or net-load. The vendor's demo beats a naive baseline and looks impressive. But the only comparison that changes the capital or operating decision is narrower: does the candidate beat the forecast you already operate on, on your series, under rules fixed before anyone sees the score? Getting that wrong shows up later in reserve margins, hedging, rate cases, and resource plans.

Forecast Assurance is built for the moment a real decision is on the table:

02 · What Forecast Assurance does

An independent adoption decision, sealed before pressure

Objective forecast benchmarking is not new — public frameworks such as EPRI's Forecast Arbiter already support impartial, repeatable evaluation, and large firms sell broad forecasting consulting. Blanc Quant Services does something narrower and more decisive:

A preregistered adoption decision conducted by an independent utility engineer — the incumbent mandatory, the acceptance rules frozen before scoring, and the result sealed before any commercial pressure can rewrite it.

The rules are frozen in a machine-verifiable evaluation contract and cryptographically hashed before the candidate produces a single forecast: dataset, horizon, allowed information set, baselines, rolling-origin folds, the primary metric, required operating segments, and the exact adoption threshold. The verification pipeline computes the metrics, and the decision engine applies the frozen gate. If inputs are missing, timestamps don't align, leakage checks fail, hashes don't match, the result can't be reproduced, or a stated claim conflicts with the canonical result, the run fails closed. No model, vendor, or stakeholder can influence the outcome of a failing configuration.

03 · Why the incumbent is mandatory

You are not choosing between a model and a strawman

The candidate is tested against the forecast you already trust, on identical timestamps, with each side's information set recorded. A model that beats last-value and seasonal-naive baselines has proven almost nothing about whether it beats a professionally tuned operational forecast. Making the incumbent mandatory is what turns a demo into a decision.

04 · The three verdicts

Adopt, reject, or inconclusive — and inconclusive is honest

Adopt

The candidate clears the frozen gate against your incumbent. The evidence supports replacing or augmenting the current forecast for the defined decision.

Reject

The candidate does not clear the frozen gate. It may still be strong in absolute terms — but not enough, under these rules, to justify replacing the incumbent.

Inconclusive

Evidence quality, comparability, missing inputs, leakage, alignment, or a protocol deviation prevents a defensible decision. We say so rather than manufacture a verdict.

05 · The public demonstration

Forecast Proof: TimesFM vs. PJM

Run on public data, published in full — including the rejection.

PJM published day-ahead forecast (incumbent)0.894
TimesFM 2.5-200M, zero-shot (candidate)1.226
Seasonal naive (24h)1.741
Last value2.743
Exponential smoothing2.755

Mean seasonal MASE, lower is better. TimesFM reduced error roughly 30% versus the strongest simple baseline and beat every simple baseline — yet produced roughly 37% more error than PJM's published forecast, winning only 2 of 5 folds. Because the rules were frozen before scoring, the tested configuration was rejected: this configuration did not justify replacing the incumbent.

Limitation, stated next to the verdict: PJM's published forecast and the tested zero-shot TimesFM configuration did not have equivalent information sets. PJM may incorporate operational inputs, including weather, that TimesFM did not receive. The result rejects the tested configuration under this protocol; it does not prove that TimesFM is universally inferior, and PJM's forecast was not recreated internally.

Read the full Forecast Proof, with the chart, the frozen-contract hashes, and the evidence package →

06 · What you receive

An executive decision, and the evidence under it

Forecast Adoption Decision Memo

The executive deliverable: the adopt / reject / inconclusive verdict, what it does and does not prove, performance by operating segment, and the limitations — written to be defended internally to boards, procurement, and regulators.

Forecast Assurance Evidence Package

The sealed technical deliverable: frozen contract, data manifest, candidate and incumbent forecasts, fold definitions, leakage and alignment checks, raw fold results, metric calculations, adversarial review, limitations, the machine-verifiable decision record, a hash manifest, and a reproduction runbook.

07 · What you supply

What the review needs from you

Datasets are exchanged under agreement after we scope the engagement — never through the form below.

08 · Scope, timeline, and fee

Forecast Assurance Decision Review

$25,000
fixed fee · independent of the verdict
2–3 weeks
after validated data receipt
1 · 1 · 1
one candidate · one incumbent · one decision

Excluded unless separately contracted:

09 · Independence & data handling

How the verdict stays trustworthy

10 · What happens next

Discuss a Forecast Evaluation

Tell us about the decision. If there's a real adoption or renewal question and the timing fits, we scope the engagement and send a one-page brief. We reply within two business days.

No datasets here — metadata only. We store your inquiry to reply; we do not share it.

Prefer email? CEO@blancquantsystems.com