📉 Forecasting Methods

Two of the most common ways to turn past demand into a forecast, run side by side on the same demand so you can see the difference: a moving average treats the last few periods equally; exponential smoothing leans on the most recent one. Watch how each reacts to a spike, a trend, or a season.

Demand pattern

Average demand

Noise

Moving-avg window (n)

Smoothing α

Moving average

F₍ₜ₎ = average of the last n actual demands

MAPE – MAD –

Exponential smoothing

F₍ₜ₎ = α·(last actual) + (1−α)·(last forecast)

MAPE – MAD –

Actual demand Forecast

The x-axis label P + a number marks the period count — P0 is the first period, P49 the 50th and last (this run simulates 50 periods total). Both charts share the same P-axis so you can compare the two methods at the same moment in time.

How they differ

Moving average = (D₍ₜ₋₁₎ + … + D₍ₜ₋ₙ₎) ⁄ n –

Exponential smoothing = α·D₍ₜ₋₁₎ + (1−α)·F₍ₜ₋₁₎ – Smaller α = smoother & more sluggish; larger α = twitchier & more responsive. It quietly weights every past period, fading geometrically.

Try this: keep the pattern on Step change and hit play. The moving average climbs in a straight ramp over n periods (every dropped old value is worth the same), while exponential smoothing rises in a curve, fast at first then easing in. Now switch to Linear trend: both lag behind, because each only extrapolates the past level — neither projects a trend. That lag is exactly what forecast bias measures, and choosing n or α is the responsiveness-vs-stability dial.

A short window / high α chases noise (low bias, high variance); a long window / low α is calm but slow (higher bias on trends). There's no free lunch — only the trade-off you pick for your demand.