Sales forecasting methods, explained
Salesdamus doesn't guess. Every forecast is produced by well-understood statistical methods — and the tool tells you which one it used and how accurate it was. Here's what each method does and when it shines.
Linear trend (OLS regression)
Ordinary least squares fits a straight line ŷ = a + b·t through your history, minimizing the squared distance to each point. The slope bis your period-over-period trend. It's the right default when sales move steadily up or down without a repeating pattern.
Seasonality (Holt-Winters)
Most real sales have a rhythm — holiday spikes, summer lulls, quarter-end pushes. Holt-Winters exponential smoothing tracks three things at once: the level, the trend, and a repeating seasonalcomponent. Given enough history (ideally two-plus full cycles), it captures patterns a straight line can't. Salesdamus tests a 12-period and 4-period season and keeps whichever forecasts best.
Bayesian shrinkage
With only a few data points, a raw trend can over-react to noise. Bayesian shrinkage pulls the forecast partway back toward the long-run average — a lot when you have little data, barely at all once you have a year or more. It trades a touch of bias for a big drop in variance.
Confidence intervals: bootstrap & Monte Carlo
A single number is a false promise. Salesdamus reports a 95% interval — the plausible range. Bootstrapresamples the model's residuals thousands of times to gauge how much the forecast could vary. Monte Carlodraws thousands of samples from the forecast's distribution. Both widen further into the future, because uncertainty compounds with distance.
Model selection by backtest
Which method should you trust? The honest answer is: whichever one would have predicted your recent past best. Salesdamus holds out the last few periods, forecasts them with each candidate model, and measures the error. The model with the lowest out-of-sample error is auto-selected — and you see the full comparison.
Accuracy metrics
R² is the share of variation the trend explains. MAPE is the average percentage error. p-value tests whether the trend is statistically real or could be chance. Higher R², lower MAPE and p-value mean a more trustworthy fit. See the glossary for full definitions.