Demand Forecast — Multi-SKU, from CSV
Uses the 24-month history file from the upload page. Choose a statistical logic, tell us your commercial expectations, and we generate the next 12 months per item with a confidence band you can plan around.
Calculated locally in your browser. No server uploads.
1. Statistical logic
The baseline before any of your commercial input. Pick the one that matches how your demand behaves.
2. Which sales count as real demand
Tick the sales types from your file that represent genuine customer demand. Everything unticked is ignored here and in the Safety Stock calculation.
3. Your inputs
Showing demo history until you upload a 24-month Forecast file or load a sample.
Methodology & Guide
How the 12-month demand forecast, confidence band and accuracy backtest are calculated, with a worked example.
Formula breakdown
Outlier capping (before any forecasting):
Month capped to mean ± k × σ (k = 2, 2.5 or 3; off = raw history)
Statistical baseline — pick one logic:
Moving average: F = average of the last w months Seasonal average: F(m) = overall mean × S(m) Trend + seasonality: F(t) = ( a + b × t ) × S(t mod 12)
Commercial uplift and confidence band (h = months ahead):
Forecast = Baseline × ( 1 + uplift % ) capped at 2.0× Low / High = Forecast ∓ Z × σ × √( h ÷ 3 )
Accuracy backtest (last 6 months hidden, re-forecast with the same logic, no uplift):
WAPE = Σ | actual − forecast | ÷ Σ actual × 100 Accuracy = 100% − WAPE
- F
- Forecast quantity per month before uplift (the statistical baseline).
- w
- Moving-average window in months (default 6).
- S(m)
- Seasonal index for calendar month m = average of that month ÷ overall average, from the 24-month history.
- a, b
- Level and monthly slope of a straight line fitted to the de-seasonalised history (least squares).
- σ
- Standard deviation of the item's cleaned monthly history.
- Z
- Normal-distribution factor for the confidence interval (80% → 1.28, 90% → 1.64, 95% → 1.96).
- WAPE
- Weighted absolute percentage error: total forecast miss as a share of total actual demand.
Worked example: a steady item with a 10% commercial uplift
An item alternates between 90 and 110 units a month for 24 months (average 100, standard deviation about 10.2). You pick the moving average over 6 months, add a 10% commercial uplift, keep 3-σ outlier capping and a 90% confidence band.
- No month is more than 3 σ from the mean (100 ± 30.6), so nothing is capped.
- Baseline = average of the last 6 months (90, 110, 90, 110, 90, 110) = 100 units a month.
- Forecast = 100 × 1.10 = 110 units a month → 1,320 units over the next 12 months.
- Month 1 band = 110 ∓ 1.64 × 10.2 × √(1 ÷ 3) ≈ 110 ∓ 9.7 → about 100 to 120 units.
- By month 12 the band widens to 110 ∓ 1.64 × 10.2 × √4 ≈ 110 ∓ 33.6 → about 76 to 144 units.
Plan on 110 units a month (1,320 a year), with stock decisions sized for the 120-unit high end in the first month.
Quick FAQs
Which forecasting method should I choose?
Moving average suits steady items with no clear season. Seasonal average suits items that peak in the same months every year. Trend + seasonality suits items that are clearly growing or fading. The accuracy card backtests your chosen logic on the last 6 months, so you can compare methods on your own data.
Is this Holt-Winters exponential smoothing?
Not exactly. Trend + seasonality is a classical decomposition: it removes multiplicative seasonal indices, fits a straight trend line and puts the season back. It captures the same level, trend and season as Holt-Winters, but with fixed, transparent weights instead of tuned smoothing parameters, so every number can be checked by hand.
How is forecast accuracy measured?
The tool hides your most recent 6 months, forecasts them from the older 18 months with the same logic and outlier rule (without your uplift), and compares the result with what really happened. Accuracy is 100% minus the weighted absolute percentage error (WAPE) across all items.
Full derivations, data safeguards and the privacy architecture are on the Methodology & Formulas page.