Forecasting playground
About this app
An interactive demand-forecasting lab: 52 weeks of demand across five trend types (horizontal, seasonal, sporadic, linear-seasonal, nonlinear) with adjustable noise. Four forecasting methods (naive, moving average, exponential and seasonal smoothing) forecast live over the test range; a MAPE/MAD ranking shows instantly which method fits which pattern.
The basics
Demand forecasting derives a forecast for upcoming periods from past demand, and the right method depends on the pattern. The naive forecast repeats the last value, the moving average takes the mean of the last n periods, and exponential smoothing weights recent values more: L_t = α·y_t + (1 − α)·L_t−1. These three produce a flat forecast and follow neither trend nor season. Holt-Winters adds trend and seasonal components. Accuracy is commonly measured with MAD (mean absolute deviation) and MAPE (mean absolute percentage error).
What the app shows
A chart shows 52 weeks of demand: weeks 1 to 40 are the history used for fitting, weeks 41 to 52 the test horizon. You choose one of five patterns, horizontal, seasonal, sporadic, linear-seasonal or nonlinear with a life cycle, and the amount of noise. Four methods can run at once: naive, moving average with a window of 2 to 12 weeks, exponential smoothing with adjustable α, and a simplified additive Holt-Winters with a 13-week season. The accuracy panel computes MAPE and MAD over the test weeks and marks the best method.
Things to try
Pick “seasonal” and compare: the flat methods miss the oscillation, Holt-Winters follows it. With “linear-seasonal” the naive methods fall behind the trend. Raise the noise and watch all errors increase; a longer window or a smaller α smooths more. Sporadic demand has weeks with zero demand, where MAPE is undefined, so the app points you to MAD. Such series call for specialised methods such as Croston’s. “Reroll” creates a new random series with the same pattern.
Frequently asked questions
- Which forecasting method works for seasonal demand?
- Seasonal demand needs a method with a seasonal component, such as Holt-Winters (triple exponential smoothing). The naive forecast, moving average and simple exponential smoothing produce a flat value and cannot capture seasonal swings.
- What is the difference between MAPE and MAD?
- MAD (mean absolute deviation) is the average absolute forecast error in units. MAPE (mean absolute percentage error) divides each error by the actual value and averages the percentages. When some periods have zero demand, MAPE is undefined and MAD is the more useful measure.
- How do you choose alpha in exponential smoothing?
- A large α weights the latest values heavily and reacts quickly to change, but also follows the noise. A small α smooths more and reacts slowly. In practice α is often chosen to minimise the forecast error on historical data.
- What data does the app use?
- The app generates synthetic 52-week demand series from the chosen pattern plus random noise. All methods learn only from weeks 1 to 40 and are compared on weeks 41 to 52.
Subject: Logistics & Supply Chain
More from Logistics & Supply Chain
- AAS & digital product passport
- Autonomous traffic in logistics systems (3D)
- Stock & service level
- Blockchain traceability
- Bullwhip effect
- Circular Economy — R-strategies
- Digital Twin level model
- Bottleneck & throughput (3D)
A VisuApp by heyprof: interactive, free in your browser, no sign-up. What is a VisuApp? · All apps