📈 Supply Chain & Planning
📈 Demand Forecasting & Accuracy
Generate a simple forecast — naive, moving average or exponential smoothing — and score it with the standard CPIM/CSCP error metrics: MAD, MSE, MAPE, forecast accuracy %, bias and tracking signal.
1 Inputs
Oldest period first, newest last. Up to ~24 periods. Use commas, spaces or new lines.
Naïve: Fₜ = Aₜ₋₁. MA: mean of last N actuals. SES: Fₜ = Fₜ₋₁ + α(Aₜ₋₁ − Fₜ₋₁).
The next-period forecast uses the chosen method on all available actuals.
2 Accuracy results
📉 Actual vs forecast demand
Actual demand
Forecast
Next-period forecast
🧮 Period-by-period error breakdown
📏 Metrics follow the APICS CPIM / CSCP forecast-error definitions. Figures are for guidance — validate against your own demand history and planning standards.
Frequently asked questions
What is MAD and how is it calculated?
MAD is the Mean Absolute Deviation — the average of the absolute forecast errors, MAD = mean(|Aₜ − Fₜ|). It is in the same units as demand (e.g. units), so a MAD of 12 means the forecast is off by 12 units on average, regardless of direction. It is the basis for the tracking signal.
How is MAPE and forecast accuracy % computed?
MAPE is the Mean Absolute Percentage Error: MAPE = mean(|Aₜ − Fₜ| / Aₜ) × 100%. It expresses error as a percentage of actual demand, so it is comparable across products of different volumes. Forecast accuracy % is simply 100% − MAPE. Periods with zero actual demand are skipped in MAPE because division by zero is undefined.
What does bias tell me?
Bias is the mean error, mean(Aₜ − Fₜ). A positive bias means actuals are consistently higher than the forecast (you are under-forecasting — risk of stockouts); a negative bias means you are over-forecasting (risk of excess inventory). Unlike MAD, errors can cancel out, so a low bias with a high MAD means the forecast is erratic but centred.
What is a tracking signal and when is it out of control?
The tracking signal is the running sum of errors divided by MAD: TS = Σeₜ / MAD. It detects persistent bias. As a rule of thumb a forecast is considered out of control when |TS| > 4 (some planners use ±3.75): the forecast is drifting in one direction and the model should be re-fit or the method changed.
Which method should I pick — naive, moving average or SES?
Naïve is the baseline every forecast should beat. Moving average (N) smooths noise — larger N is smoother but slower to react. Single exponential smoothing weights recent periods more via α (0–1): higher α reacts faster, lower α is smoother. SES suits stationary demand with no strong trend or seasonality; for trend use Holt's method and for seasonality use Holt-Winters.
Why do the first periods have no error?
Each method needs history before it can forecast. The naïve method has no forecast for period 1; a moving average of N needs N actuals before the first forecast; SES is seeded with the first actual. Error metrics are computed only over periods that have both an actual and a forecast.