Exponential smoothing projection
Weight recent observations more heavily to capture evolving trends.
The exponential smoothing method (Holt's two-parameter method) forecasts future values by fitting an exponentially weighted model that separately tracks the current level and trend of your data. Recent observations count more than older ones.
How it works
Holt's method maintains two smoothed components that update as each data point arrives:
- Level (
l): the baseline value at the current point in time. - Trend (
b): how much the level is changing per period.
Each new observation updates both components using a smoothing factor α (for level) and β (for trend). Values close to 1 mean new data has high influence; values close to 0 mean the model changes slowly.
The forecast for h steps ahead is: l + h × b.
TotalKPI uses fixed smoothing parameters tuned for typical business metrics.
When to use it
- Your data has a trend (rising or falling), but the rate of change is evolving - not perfectly constant.
- Recent values should carry more weight than older history.
- Examples: a metric that started flat and began accelerating recently, or a recovering metric after a dip.
Limitations
- Can over-project if the most recent data contains an unusual spike or dip - that distortion gets baked into both the level and trend.
- Fixed smoothing parameters mean it may not be perfectly tuned to every dataset. A well-calibrated Linear trend may perform better for very stable trends.
- Does not model seasonality (use it on seasonally-adjusted data if cycles are present).