Dispensight Forecasting Engine · v4.5.2
Real-world accuracy report

Forecasting that knows
what it doesn't know.

Most sales forecasts hand you a single number and hope. Dispensight's engine is Bayesian — every prediction comes with a calibrated confidence range, so you can staff and stock against real uncertainty instead of a lucky guess. Here's how it held up over 35 consecutive days on a live retail deployment.

~10%
Avg daily error
MAPE across 35 days
91.4%
Interval coverage
actuals inside the 80% band
+2.4%
Directional bias
near-zero — no systematic drift
98.6%
Cumulative accuracy
5-week total vs actual

Predictions vs. reality

Each day's forecast (line) carries an 80% credible interval (shaded band); the dots are what actually happened. Values are indexed to the period average (= 100) — no store, date, or revenue figure is shown. When the dots land inside the band as often as the band promises, the uncertainty is honest. Here, they land inside even a little more often than the 80% target.

80% credible interval Forecast Actual — inside band Actual — outside band

Why a range beats a number

A point forecast that names one number for Friday is silently wrong almost every time — real demand lands a little above or below. A calibrated interval says “80% chance between here and here,” and then actually keeps that promise. That's what lets you set staffing and reorder points with a known safety margin instead of gut feel.

Calibration is the hard part: many models quote confidence they don't earn. Over this run, an 80% interval captured reality 91.4% of the time — slightly conservative, which is exactly the safe direction to miss.

Calibration score

91.4% / 80% target

32 of 35 days fell inside the 80% band. The three exceptions were single-day demand spikes tied to external shocks (e.g. extreme-weather events), not model drift — the forecast re-centred immediately after each.

How close, day to day

Share of the 35 days whose actual sales landed within a given band of the forecast. Even single-store daily demand — the noisiest horizon there is — clustered tightly around the prediction.

What's under the hood

Dispensight v4.5.2 · HMC-NUTS forecasting engine

Bayesian HMC-NUTS core

A Hamiltonian Monte Carlo sampler with the No-U-Turn Sampler explores the full space of likely outcomes, so every forecast is a probability distribution — not a fragile single guess.

Weather-aware

External regressors let the model anticipate how conditions move foot traffic, and cleanly attribute the demand shocks it can't prevent so they don't poison future forecasts.

Self-diagnosing

Sampling chains are checked for convergence in-flight; any that fail restart automatically, so a bad draw never ships as a forecast.

Continuously back-tested

Every forecast is scored against what actually happened — MAPE, interval coverage, and bias tracked daily — so accuracy is measured, not assumed.

Stock and staff with confidence, not guesswork.

Dispensight turns your point-of-sale history into calibrated demand forecasts your team can actually plan around.

See Dispensight → Request a demo
Validation snapshot from a single, anonymized retail deployment over 35 consecutive days. All values are indexed to the period average; no store-identifying, location, date, or financial data is disclosed. Forecast accuracy varies with sales history, data quality, and local conditions — figures shown are representative of one deployment, not a guarantee.
© Dispensight · HMC-NUTS forecasting engine v4.5.2 · www.dispensight.com