From Spreadsheet to Forecast: Exponential Smoothing for Demand Planning

Forecasting projects fail for boringly consistent reasons: the model is more sophisticated than the data supports, nobody agreed on the error metric, and the planner overrides the output every month without leaving a trace. All three are fixable before you touch an algorithm.

Pick the simplest model the data justifies

The exponential smoothing family covers most industrial demand, and the choice among them is driven by two questions: is there a trend, and is there seasonality?

Simple exponential smoothing

F(t+1) = alpha * A(t) + (1 - alpha) * F(t)

alpha near 0.1  -> heavy smoothing, slow to react
alpha near 0.4  -> responsive, noisier

Fit alpha by minimising squared error over a holdout period rather than choosing it by feel. In practice most industrial series land between 0.15 and 0.35; if your optimiser wants 0.8, your series has a structural break in it and you should investigate that instead.

Holt-Winters, multiplicative

Use multiplicative seasonality when the seasonal swing scales with the level, which is the normal case for consumer demand. Use additive when the swing is a roughly constant number of units. Getting this backwards is a common and expensive mistake: an additive model on a growing multiplicative series under-forecasts every peak by a widening margin.

Choose an error metric before you see results

• MAPE is intuitive but explodes on near-zero actuals and punishes under-forecast differently from over-forecast. Avoid it for intermittent items.
• MAE is robust and in original units, which makes it easy to explain.
• RMSE penalises big misses more heavily. Use it when a single large miss is genuinely worse than several small ones.
• Bias (mean error) is the one people forget and the one that costs the most. A model with excellent MAPE and persistent positive bias will fill your warehouse.

Report MAE and bias together, always. One number tells you accuracy, the other tells you direction.

Benchmark against naive

Before accepting any model, compute the naive forecast — next period equals this period, or equals the same period last year for seasonal series. If your sophisticated model cannot beat naive on a holdout sample, it has learned noise. This check takes five minutes and quietly kills a lot of bad projects.

Make overrides visible

Planners will override the forecast, and they should — they know about the promotion, the tender, the customer who is closing a warehouse. The problem is untracked overrides. Store the statistical forecast, the override, and the reason as three separate fields. After two quarters you can measure whether overrides improve or degrade accuracy. In most organisations the honest answer is that a small number of planners add real value and the rest add noise, and you cannot tell which without the data.

Forecast at the right grain

Forecast where the demand signal is strongest and aggregate or disaggregate from there. Item-level weekly forecasts for a long-tail catalogue are mostly noise. Forecast the family monthly, then split by historical share — the aggregate is far more stable, and the split ratio changes slowly enough to be reliable.

When to stop smoothing and start modelling

Move beyond smoothing when you have real causal drivers you can observe in advance: price changes, promotion calendars, weather, a tender pipeline. That is a regression problem, not a time series problem, and it will beat smoothing when the drivers are genuinely leading. If your only candidate variables are lagged versions of demand itself, stay with smoothing.