Prove your demand forecast is actually worth trusting.
ForecastProof scores your forecast against the naive baseline (just repeating last period), on your own held-out demand, and hands you one number: Forecast Value Add. Not another accuracy percentage you have to take on faith. A deterministic figure that says whether the forecast beat doing nothing, and by how much, before you commit a single purchase order to it.
Book a pilot See a sample ForecastProof report
THE PROBLEM
A demand forecast comes out (maybe from a planner, maybe from an AI forecasting tool nobody fully audits), and procurement is about to commit real money to it. The dashboard says 85% accurate, everyone nods, the order goes out. But nobody answered the one question that actually matters: did that forecast beat the naive baseline, the do-nothing move of just repeating last period? An 85% forecast sounds great until you learn that last-value naive scored 84% on the same SKU, which means all that machinery bought you one point, and your CFO is now asking why you are paying for it. Forecast accuracy across vendors has plateaued in the 81 to 90 range, the agent-washing backlash has buyers skeptical of any unaudited AI number, and there is still no standalone product that just tells you the honest answer.
HOW IT WORKS
ForecastProof sits behind your forecast, it does not replace it. You give it the demand history and the held-out actuals (the periods the model was not fit on). It builds the naive baseline (last-value persistence, or seasonal naive when your SKU has seasonality), scores both your forecast and that baseline against the actuals using WMAPE (weighted, so it stays honest even when some periods have zero demand), and computes Forecast Value Add as the difference. Positive means your forecast earned its keep. Zero or negative means it did not beat repeating last period, and you just learned that for free instead of after a quarter of bad orders. The whole calculation is deterministic arithmetic, no language model touches the proof number, and if the data is malformed or too short it fails closed and tells you, rather than inventing a figure to look complete.
WHAT YOU GET (the report)
For every SKU, one honest line procurement can act on: - the model forecast and the naive baseline it was measured against, - model WMAPE and naive WMAPE on your held-out demand, - Forecast Value Add, in error points and as a fraction of the naive error removed, - beats naive, a plain yes or no, and - the accuracy KPI for the same period, to frame it.
A ForecastProof report you can put in front of a finance partner, not a dashboard you have to defend on vibes.
WHO IT'S FOR
Demand planners and S&OP leads at mid-market manufacturers, distributors, and retailers; supply-chain teams who bought a forecasting tool and genuinely cannot tell if it beats a spreadsheet; and forecasting-AI vendors who need to certify their own accuracy claim to a skeptical buyer. If someone is about to ask you to prove your forecast is better than doing nothing, this is for you.
PRICING (the ladder)
- Pilot, a fixed-price paid proof on your own SKUs and demand history. You see the FVA numbers before you commit to anything. - Value-metered, per SKU or per forecast cycle, once the pilot shows how many of your forecasts quietly fail to beat naive. - Platform, ForecastProof wired into your S&OP cadence so no forecast reaches procurement without an FVA score attached.
IP is licensed, never assigned. The verifier stays ours, the certainty is yours.
THE PROOF (dogfood)
We run this discipline on ourselves. ForecastProof is the same verifier-first architecture that already powers AYA's regulatory pilot and its bookkeeping close verifier, a checker that sits behind a producer and reports, with a deterministic number, whether the producer beat a naive baseline. Any internal forecast we lean on is a candidate for the same pass. We are not selling a "does your forecast beat naive" proof we would not run on our own numbers.
Honest status, today: the FVA math is built and verified offline on constructed series (a trend-tracking model beat last-value naive by about 14 WMAPE points, a deliberately bad model scored negative Forecast Value Add, and malformed input failed closed instead of emitting a number). It is a small piece of deterministic code, marked for review, wired behind the existing forecasting patterns. Running it on your real demand data needs your feed connected and the system stood up, which is the pilot setup, not a rebuild. No live customer run has happened yet.
First design partner slot open, a demand-planning or S&OP team that files enough forecasts to feel the "prove it beats naive" pressure. case study to follow
HONEST NOTE
We would rather under-promise. Where the demand data supports a decisive comparison, ForecastProof gives you the Forecast Value Add and stands behind it. Where it does not (too little history, a held-out window that sums to zero demand), it tells you the number is undefined and stops, it does not manufacture a green figure to look finished. That refusal to fake a proof number on something your procurement budget rides on is the entire product.
Book a pilot
*This page is a specification. The capability it describes is not built yet, and nothing here is a claim that it runs today.*