Demand & Forecasting — 4 calculators

CV bands: Very low (<0.25) → smooth, easy to forecast with simple methods. Low–medium (0.25–0.5) → standard, exponential smoothing works. Medium–high (0.5–1.0) → noisy, need safety stock and possibly lead-time reductions. High (>1.0) → lumpy / intermittent, classical statistics fail — switch to Croston's method or service-level based on order history.

The "lumpy" trap. Many B2B and project-driven businesses compute CV on daily demand and get something like 1.4 — then apply Z-score safety-stock math and get nonsensical numbers. That's because σ stops describing the distribution when there are zero days interspersed with big orders. The tool flags this with a "lumpy" classification.

Outliers. One day of 600 in a series otherwise around 40 will inflate CV dramatically. If that day was a one-off (single customer spike, promo), drop it before computing. The tool shows both with-outlier and (if you trimmed) without-outlier metrics in the result.

Limitations. This calculator uses the sample standard deviation, which is fine for n≥10. Below that, treat the CV as directional only.

The formula. Total lead-time variance = (σ_d)² × L + (σ_L)² × d². The first term is what most calculators model. The second term — variance of demand multiplied by variance of lead time — is often larger and almost always ignored.

What it changes. If your supplier quotes 7-day lead time with σ=2, you should plan for lead times that often hit 9–11 days. The safety stock the simple formula gives you is too low.

Where σ_L comes from. Pull 12 months of receiving dates for this supplier. Subtract the PO date from each. Compute σ. Don't use their quoted lead time — their actual.

Who this is for. Anyone whose suppliers transit customs, ports, or have any human approval step in the chain. Domestic pallet deliveries from a major carrier usually have σ_L ≈ 0.5–1 day.

Probability bands: <5% = safe; 5–15% = monitor; 15–30% = consider expediting a PO; >30% = act now (call supplier, find a substitute, communicate to sales).

The model assumes normal demand. If demand is lumpy (see Variability tool), this number understates risk. For lumpy demand, use a Monte Carlo simulation or a simple percentile-based check ("in the worst 10% of days, can we still cover?").

Open POs are assumed certain to arrive. If your PO has historically failed to arrive on the date promised, reduce confidence — the tool's stockout probability is the floor, not the ceiling.

What this won't tell you. A stockout isn't just a probability — it's a customer experience cost. Run the numbers against the value of the lost sale (typically 3–10× margin for B2B; lifetime value for B2C subscriptions).

MAPE (mean absolute percentage error) is the most common accuracy metric. It's intuitive but it punishes over-forecasts more than under-forecasts and ignores sign — a forecast that's off by 50% in either direction has the same MAPE. That's why bias matters separately.

Bias (mean forecast error / mean actual) is positive when you systematically over-forecast, negative when you under-forecast. Anything outside ±5% deserves attention.

Tracking signal = cumulative error / MAD (mean absolute deviation). When its absolute value exceeds ~4–8, your forecast is consistently wrong in one direction — your model has drifted. Time to retrain or reset.

Limitations. If your forecast is for new product launches or items with sparse history, MAPE will look terrible even when the model is doing as well as possible. Always sanity-check against "could a human do better?" before trusting the number.