Blog · Forecast and pipeline · Finance and FP&A teams
Adaptive smoothing changes the weight on recent observations in response to forecast errors. Work through an error-responsive example, compare it with fixed smoothing, and see the checks needed before adopting it.
Adaptive smoothing forecasting adjusts how strongly a forecast responds to recent observations by examining its errors. It can react faster when the business repeatedly exceeds or falls below expectations. Assess that response against a fixed-smoothing baseline before using it to change purchasing, capacity, or financial commitments.
Simple exponential smoothing updates a level estimate from the latest actual and the previous forecast. NIST's explanation of single exponential smoothing shows how a larger smoothing weight gives more influence to recent observations and reduces the influence of older ones.
Next forecast = Current forecast + Response weight × (Current actual − Current forecast)
With fixed smoothing, the response weight stays constant. An adaptive approach changes it under a stated rule. ETH Zurich's teaching material on Trigg and Leach adaptive smoothing describes using a tracking signal based on forecast errors to adjust the smoothing constant.
Repeated errors in one direction suggest that the level estimate is behind the business. Errors that cancel suggest less reason to chase the latest fluctuation. Neither pattern establishes the cause: review types of forecasting errors before treating a data problem as a demand change.
Here is one explicit error-responsive implementation. Let F be the forecast, Y the actual, e the error, S the smoothed signed error, and A the smoothed absolute error. The subscript t identifies the period just completed.
e_t = Y_t − F_t
S_t = β × e_t + (1 − β) × S_(t−1)
A_t = β × |e_t| + (1 − β) × A_(t−1)
α_t = |S_t| / A_t
F_(t+1) = F_t + α_t × e_t
The bars mean absolute value. The error-smoothing weight β stays fixed; the response weight α changes. Under consistent initialization, the ratio lies between zero and one. If A is zero, this example uses a fallback response weight of 0.20 rather than dividing by zero.
Timing is part of the specification. Calculate the current error only after the period closes, then use the resulting weight to forecast the next period. Do not use that weight to revise the forecast being scored for the period just closed. Other implementations can time their updates differently, so retain the exact rule with the forecast version.
An illustrative US business compares its next-week forecasts for service orders. Both methods start with a forecast of 100 orders, established before the evaluation period. Fixed smoothing uses α = 0.20. Adaptive smoothing uses β = 0.20, S_0 = 0, and A_0 = 10 orders; these are declared starting assumptions for the example.
Actual demand rises after the first week. The table records forecasts made before each week's actual became available. Errors equal actual minus forecast.
| Week | Actual orders | Fixed forecast | Fixed error | Adaptive forecast | Adaptive error |
|---|---|---|---|---|---|
| 1 | 100 | 100.00 | 0.00 | 100.00 | 0.00 |
| 2 | 120 | 100.00 | 20.00 | 100.00 | 20.00 |
| 3 | 125 | 104.00 | 21.00 | 107.69 | 17.31 |
| 4 | 128 | 108.20 | 19.80 | 117.48 | 10.52 |
| 5 | 130 | 112.16 | 17.84 | 124.26 | 5.74 |
| 6 | 129 | 115.73 | 13.27 | 128.19 | 0.81 |
This is a level-shift illustration, not evidence that adaptive smoothing wins generally. Keep fractional forecasts during calculation even though actual orders are whole numbers. An operating team can apply its capacity rounding rule after the forecast comparison.
The adaptive tracking calculations behind those forecasts are:
| Week just closed | Smoothed signed error S | Smoothed absolute error A | Response weight α | Forecast for next week |
|---|---|---|---|---|
| 1 | 0.0000 | 8.0000 | 0.0000 | 100.00 |
| 2 | 4.0000 | 10.4000 | 0.3846 | 107.69 |
| 3 | 6.6615 | 11.7815 | 0.5654 | 117.48 |
| 4 | 7.4335 | 11.5295 | 0.6447 | 124.26 |
| 5 | 7.0944 | 10.3712 | 0.6840 | 128.19 |
| 6 | 5.8381 | 8.4595 | 0.6901 | 128.75 |
After week 1, S is 0.20 × 0 + 0.80 × 0 = 0, and A is 0.20 × 0 + 0.80 × 10 = 8. The resulting response weight is zero, so week 2's forecast stays at 100.
After week 2, S becomes 0.20 × 20 + 0.80 × 0 = 4, and A becomes 0.20 × 20 + 0.80 × 8 = 10.4. The next forecast is 100 + (4 / 10.4) × 20 = 107.69 orders, rounded to two decimals. Fixed smoothing instead gives 100 + 0.20 × 20 = 104.
All calculations retain full precision. Tracking values and weights display four decimals; forecasts and errors display two.
Mean absolute error = Sum of absolute forecast errors / Number of forecast periods
| Method | Sum of absolute errors, orders | Periods | Mean absolute error, orders |
|---|---|---|---|
| Fixed smoothing | 91.91 | 6 | 15.32 |
| Adaptive smoothing | 54.38 | 6 | 9.06 |
The fixed errors add to 91.912 before rounding; dividing by 6 gives 15.32 orders. Adaptive errors add to 54.3800348884 before display rounding; dividing by 6 gives 9.06. The adaptive calculation responds faster to this sustained increase.
The result concerns one-step forecasts updated as actuals arrive. It does not prove superiority for a forecast fixed several weeks in advance. Forecasting: Principles and Practice recommends evaluating genuine forecasts on observations unavailable when forecasts were made. Use earlier history to choose settings and later periods to assess them, then repeat across stable conditions, spikes, and reversals.
Finance or planning should own the comparison, while operations validates whether the errors affect a real staffing or purchasing decision. Keep versions separate from targets; budget vs forecast explains why they serve different purposes. If aggregate improvement hides a persistently wrong location or item, review SKU and site forecast bias.
Bring a file containing the fixed and adaptive forecast versions, publication dates, forecast periods, and matching actuals. Covirage maps the columns, and its tools compute and check the error comparisons; the external AI model explains the results with every figure traced to its rows. Create a free Covirage account and bring a sample of your own data.
See forecast analysis for the supporting use case.
Adaptive smoothing is a form of exponential smoothing with a changing smoothing weight. Ordinary simple exponential smoothing keeps that weight fixed between updates. The adaptive rule uses recent forecast errors to decide how strongly the next estimate should respond to the latest observation.
No. Faster responses can help after a lasting demand change, but they can also chase temporary spikes or bad records. Compare genuine forecasts over multiple business conditions, using a fixed baseline and the horizon that matters for your purchasing, staffing, or finance decision.
Changing the smoothing weight does not by itself add a seasonal component. Recurring holiday or monthly patterns require an approach that represents them explicitly, or a separately documented seasonal adjustment. Otherwise predictable calendar effects can be mistaken for changes in the underlying demand level.
Record the starting forecast, error-tracking starting values, error-smoothing weight, response-weight formula, bounds, and zero-denominator rule. Also state when each weight becomes available. Two implementations with different timing or initialization can produce different forecasts from the same actuals.