What Is Demand Forecasting?
Demand forecasting is the practice of estimating future demand so that an organization can plan inventory, capacity, staffing, procurement, and finances. Forecasts are never perfectly accurate — the goal is to be as accurate as is useful and to understand the error so plans can buffer against it. Forecasting methods range from qualitative (expert judgment, market research) to quantitative time-series and causal models; this guide focuses on the core quantitative time-series techniques.
Components of a Time Series
Historical demand typically contains several patterns:
- Level — the baseline average around which demand fluctuates.
- Trend — a sustained upward or downward movement over time.
- Seasonality — a regular, repeating pattern tied to the calendar (e.g., higher sales each December).
- Cyclical — longer wave-like movements tied to economic cycles.
- Random (noise) — irregular variation that cannot be predicted.
The right method depends on which components are present.
Moving Averages
A simple moving average (SMA) forecasts the next period as the average of the last n actual periods:
Fₜ₊₁ = (Aₜ + Aₜ₋₁ + … + Aₜ₋ₙ₊₁) / n
A larger n smooths more (more stable, slower to react); a smaller n reacts faster (more responsive, noisier). SMA is simple but lags trends and gives every recent period equal weight.
A weighted moving average (WMA) assigns larger weights to more recent periods (weights summing to 1), making the forecast more responsive to recent demand while still smoothing noise.
Exponential Smoothing
Single exponential smoothing elegantly weights all past data with exponentially decreasing importance using a single smoothing constant α (between 0 and 1):
Fₜ₊₁ = α·Aₜ + (1 − α)·Fₜ
The new forecast is a blend of the most recent actual and the most recent forecast. A high α (say 0.5) reacts quickly to change; a low α (say 0.1) produces a smooth, stable forecast. It requires storing only the last forecast and actual, which made it popular before cheap computing and keeps it popular today.
Example. If α = 0.3, last forecast Fₜ = 100, and actual Aₜ = 120, then Fₜ₊₁ = 0.3·120 + 0.7·100 = 36 + 70 = 106.
Capturing Trend and Seasonality
Single smoothing lags a trend. Double exponential smoothing (Holt's method) adds a second equation that smooths the trend, so the forecast tracks rising or falling demand. Triple exponential smoothing (Holt-Winters) adds seasonal indices to capture recurring seasonal patterns along with level and trend. For strongly seasonal data, demand is often deseasonalized (divided by seasonal indices), forecast, then reseasonalized.
Measuring Forecast Accuracy
You cannot manage forecasting without measuring error. Let the error each period be eₜ = Aₜ − Fₜ (actual minus forecast).
| Metric | Formula | What it tells you |
|---|---|---|
| MAD | (Σ |Aₜ − Fₜ|) / n | Average error size, in original units |
| MSE | (Σ (Aₜ − Fₜ)²) / n | Penalizes large errors heavily |
| MAPE | (Σ |Aₜ − Fₜ| / |Aₜ|) / n × 100% | Average percent error, scale-free |
| Bias (mean error) | (Σ (Aₜ − Fₜ)) / n | Systematic over/under-forecast (signed) |
MAD (Mean Absolute Deviation) averages the absolute errors and is easy to interpret, but its units make it hard to compare across products. MAPE (Mean Absolute Percentage Error) expresses error as a percentage of actual demand, so it is comparable across items of different volumes (though it misbehaves when actuals are near zero). MSE squares errors, penalizing large misses. Bias keeps the sign — it reveals whether you consistently over- or under-forecast, which absolute measures hide.
The Tracking Signal
To detect when a forecast has gone systematically wrong, monitor the tracking signal:
Tracking Signal = (running sum of forecast errors) / MAD
It accumulates signed errors and divides by MAD. A value near zero means errors are balanced; a value drifting beyond roughly ±4 to ±6 signals a biased forecast that needs investigation and recalibration. The tracking signal is a control-chart-style early warning for forecasting.
Choosing a Method
| Data pattern | Suitable method |
|---|---|
| Stable, no trend or season | Moving average or single exponential smoothing |
| Trend, no season | Double exponential smoothing (Holt) |
| Trend and season | Triple exponential smoothing (Holt-Winters) |
| Driven by external variables | Regression / causal models |
You can experiment with moving averages, smoothing, and accuracy metrics on the forecasting calculator.
Best Practices
- Match the method to the pattern — using a level method on trending data guarantees lag.
- Track bias, not just MAD/MAPE — a low absolute error can still hide a systematic skew.
- Tune α (and trend/season constants) on historical data — pick the value that minimizes error out-of-sample, not in-sample.
- Combine judgment with statistics — incorporate promotions, new products, and known events the model cannot see.
- Forecast at the right level and update regularly — re-forecast as new actuals arrive and watch the tracking signal.