Forecasting methods explained
The forecast tool tries several methods on your numbers and picks the one that would have done best. Here's what each one does and when it tends to win.
Reviewed October 2026
The methods
| Method | What it does | Works best when |
|---|---|---|
| Last value (naive) | Every future period equals the most recent value. | Numbers wander without a clear trend or season. It's a surprisingly strong benchmark. |
| Same period last year (seasonal naive) | Each future month equals the same month last year. | There's strong seasonality and little growth. |
| Average change (drift) | Starts from the last value and keeps adding the average change per period, like drawing a line from your first value to your last and extending it. | Steady growth or decline, and the first and last values are typical. |
| Average of all values | Every future period equals the average of all your data. | Numbers bounce around a stable level. |
| Straight-line trend | Fits the best straight line through all your data and extends it. | Growth has been steady over the whole history. |
| Smoothed trend (Holt, damped) | Tracks the level and trend, giving recent values more weight, and gradually flattens the trend into the future. | There's a trend that changes over time. Flattening the trend avoids forecasts that grow forever. |
| Trend + seasonality (Holt-Winters) | Adds a seasonal pattern to the smoothed trend. The additive version assumes seasonal swings stay the same size; the multiplicative version assumes they grow with sales. | You have at least two full years of data with a repeating seasonal pattern. |
Holt and Holt-Winters are forms of exponential smoothing: they update an estimate each period by blending the newest value with the previous estimate. Weights on older values fade away exponentially, so recent data counts most. The tool picks how fast they adapt by testing settings and keeping the ones that fit your history best.
How the tool picks a method
- It sets aside your most recent values, about 20% of your data, as a test.
- It fits each method to the older values and forecasts the test period.
- It compares those forecasts with what actually happened and picks the method with the smallest average miss.
- It refits that method on all your data to make the forecast.
The same idea is recommended by forecasting researchers: judge a method by how well it forecasts data it hasn't seen, not by how well it fits the past. A method with enough moving parts can fit any history perfectly and still forecast badly.
Measuring accuracy
| Measure | What it means |
|---|---|
| Average miss (MAE) | The average size of the errors, in your own units. "Off by $2,300 a month on average." |
| Percentage miss (MAPE) | The average error as a percentage of the actual value. Easy to compare across products, but it can't be calculated if any actual value is zero, and it's misleading when values are close to zero. |
| Compared with naive | How much better the method did than just repeating the last value. If a method can't beat that, it isn't adding anything. |
The likely range
Next to each forecast, the tool shows a range where about 80% of outcomes should fall if the past pattern holds. It's based on how far off the method's one-step-ahead forecasts were across your history, and it widens the further ahead you look. It's a rough guide: it can't account for surprises the history doesn't contain, so real outcomes fall outside it more often than the 80% suggests.
Sources
- Forecasting: Principles and Practice, 3rd edition, Rob J Hyndman and George Athanasopoulos. See chapters on simple methods, Holt-Winters and forecast accuracy.