Demand Forecasting: Methods, Accuracy, and Why a Bad Forecast Beats No Forecast

Demand forecasting is the process of predicting future demand for a product using historical sales data, trend analysis, and contextual factors — producing a quantitative estimate that drives every downstream procurement decision from order quantity to safety stock sizing.

Quick answers

What is demand forecasting? Demand forecasting uses historical sales patterns to predict how much of each item you will sell in a future period. The forecast feeds directly into reorder point calculations, safety stock sizing, and purchase order timing. Without a forecast, every order is a guess.

What methods work for SMBs? Two cover most cases. Simple moving average — average the last N periods of demand. Exponential smoothing — weight recent periods more heavily than older ones. Both are computable in a spreadsheet. The right method depends on the item's demand pattern, which is where SBC classification matters: smooth demand responds well to exponential smoothing; intermittent or lumpy demand requires bias-corrected methods like the Syntetos-Boylan Approximation.

How accurate does my forecast need to be? Perfect accuracy is impossible. What matters is measuring the error. MAPE (Mean Absolute Percentage Error) under 25% is strong for most SMB categories. Even a forecast with 40% MAPE is better than no forecast, because you can quantify how wrong you are and size your safety stock buffer accordingly.

What about seasonality? If your product has predictable demand cycles (holiday spikes, summer slowdowns), apply a seasonal index: multiply the baseline forecast by a ratio derived from the same period in prior years. Without seasonal adjustment, your forecast will systematically under-order before peaks and over-order before troughs.

The formulas

Simple moving average:

forecast = (D_1 + D_2 + ... + D_n) / n

where D_i is demand in period i and n is the number of periods (typically 4–12 weeks).

Exponential smoothing:

forecast_t = α × D_(t-1) + (1 − α) × forecast_(t-1)

where α (alpha) is the smoothing factor between 0 and 1. Higher α reacts faster to recent changes; lower α produces a more stable forecast. α = 0.2–0.3 is a standard starting point.

Forecast accuracy (MAPE):

MAPE = (1/n) × Σ |actual − forecast| / actual × 100

Forecast method selection by demand pattern

Demand pattern	CV²	ADI	Recommended method
Smooth	≤ 0.49	≤ 1.32	Exponential smoothing or moving average
Intermittent	≤ 0.49	> 1.32	Syntetos-Boylan Approximation
Erratic	> 0.49	≤ 1.32	SBA or damped exponential smoothing
Lumpy	> 0.49	> 1.32	SBA + manual operator review

Applying a moving average to intermittent demand systematically over-forecasts, because the zeros pull the average down while the non-zero periods create spikes the average cannot track. This is the core insight of the Syntetos-Boylan Approximation — it separates demand size from demand frequency and corrects the bias.

Worked example

A coffee shop sells cold brew concentrate. Weekly sales for the past 8 weeks: 24, 28, 32, 30, 26, 34, 38, 36.

4-week moving average: (26 + 34 + 38 + 36) / 4 = 33.5 units

Exponential smoothing (α = 0.3): Starting from week 5 forecast of 28.5:

Week 6: 0.3 × 26 + 0.7 × 28.5 = 27.8
Week 7: 0.3 × 34 + 0.7 × 27.8 = 29.7
Week 8: 0.3 × 38 + 0.7 × 29.7 = 32.2
Week 9 forecast: 0.3 × 36 + 0.7 × 32.2 = 33.3 units

Both methods converge on ~33 units. But the exponential smoothing picked up the upward trend faster. The operator should also note that weeks 6–8 show an accelerating pattern — possibly seasonal (summer approaching). Comparing to the same period last year would confirm whether a seasonal adjustment is warranted.

Why most operators get forecasting wrong

Most SMBs do not forecast at all. They reorder what sold last time, which is a one-period moving average with no smoothing and no trend adjustment. This produces three predictable failures:

No buffer for variability. Reordering last period's sales assumes next period will be identical. It won't be. Without a forecast, you cannot compute the standard deviation of demand — and without σ, your safety stock formula has no input.
Trend blindness. A product growing 5% per week will be systematically under-ordered every cycle if you reorder based on last week's sales. By the time you notice, you have accumulated weeks of stockouts.
Seasonal whiplash. Operators who experienced a slow January will under-order for February, even if February has historically been stronger. Without at least one year of historical comparison, seasonal shifts are invisible until they hit.

The key insight: a forecast with 30% error is still dramatically better than no forecast. The error itself is the input to safety stock — it tells you exactly how much buffer you need. No forecast means no error measurement, which means safety stock is set by gut feel.

How LineNow handles demand forecasting

Classifies every SKU by demand pattern using ADI and CV², then routes each item to the appropriate forecasting method — exponential smoothing for smooth demand, SBA for intermittent and erratic patterns.
Computes forecast error (MAD and MAPE) per item nightly and feeds the error directly into the safety stock calculation, so buffer sizing tracks actual forecast reliability rather than an assumed constant.
Detects seasonal patterns by comparing trailing demand to same-period-prior-year data when 12+ months of history exist, applying a seasonal index automatically.
Surfaces forecast vs. actual in the dashboard so operators can see which items the model struggles with — those are the items that need manual review or event-based overrides.
Recalculates daily from POS data, ensuring that trend shifts, new product ramps, and demand regime changes propagate into reorder points within days rather than waiting for a manual spreadsheet refresh.