Holt method forecasting updates the level and trend of nonseasonal demand separately. Follow a complete numerical example, check the forecast equations, and compare saved forecasts with actual demand.
Holt method forecasting estimates demand by smoothing its current level and trend separately, then extending that trend over the forecast horizon. It is a trend adjusted forecast for a series without an explicit seasonal component. The useful test is whether its saved predictions improve decisions when compared with later actual demand.
The level estimates where demand currently sits. The trend estimates the change per period. A distribution planner might use them to project weekly case demand for a product whose volume is rising without a repeating calendar pattern.
NIST's explanation of double exponential smoothing describes the separate updates and why adding a trend equation helps with trending observations. Here, double exponential smoothing means Holt's level-and-trend formulation. Always identify the equations because that broader label can be ambiguous.
For the basic idea of giving recent observations more weight, see exponential smoothing. If the problem includes recurring seasonal peaks, Holt-Winters forecasting owns that separate method.
Level: L(t) = α × Y(t) + (1 − α) × [L(t−1) + B(t−1)]
Trend: B(t) = β × [L(t) − L(t−1)] + (1 − β) × B(t−1)
Forecast h periods ahead: F(t+h given t) = L(t) + h × B(t)
Y(t) is actual demand in period t. L(t) is the updated level, and B(t) is the updated trend in demand units per period. Alpha, α, controls the level update; beta, β, controls the trend update. Both are between zero and one, inclusive.
The one-period forecast available before observing Y(t) is L(t−1) + B(t−1). Update the level only after that actual arrives, then update the trend. Comparing the updated level directly with the same period's actual would give a misleading measure of forecasting performance.
These are the equations in Forecasting: Principles and Practice's treatment of Holt's linear trend method. That text labels the trend smoothing parameter β*, which is β here. This article uses ordinary, undamped Holt forecasting; the forecast continues the latest trend linearly.
Use a demand export with period, product, location, actual quantity and unit of measure. Aggregate it to the same regular interval used for planning. Keep returns, canceled orders and stockout periods identifiable, since shipped units may understate demand when stock was unavailable.
Store supplied forecasts separately with their creation date, target period and forecast horizon. Those fields distinguish a prediction made before demand occurred from one revised afterward.
The worked example uses one US distributor's weekly demand in cases. The starting level is 100 cases and starting trend is 10 cases per week. Both smoothing parameters are 0.5. These are explicit teaching assumptions, not optimized values or a claim that this parameter choice suits other products.
| Input | Value | Meaning |
|---|---|---|
| Starting level L(0) | 100 | Demand level before week 1 |
| Starting trend B(0) | 10 | Increase in cases per week |
| Alpha | 0.5 | Level smoothing weight |
| Beta | 0.5 | Trend smoothing weight |
In operational work, document how the starting values were obtained and keep evaluation periods out of that choice. NIST describes several possible initializations; there is no reason to assume every workbook uses the same one.
Actual demand is 112, 118, 135 and 140 cases across the next four weeks. Each row's forecast uses the previous row's level and trend. Calculations retain full precision; this table displays four decimal places where needed.
| Week | Actual cases | Forecast made before week | Updated level | Updated trend |
|---|---|---|---|---|
| 1 | 112 | 110.0000 | 111.0000 | 10.5000 |
| 2 | 118 | 121.5000 | 119.7500 | 9.6250 |
| 3 | 135 | 129.3750 | 132.1875 | 11.0313 |
| 4 | 140 | 143.2188 | 141.6094 | 10.2266 |
The first forecast is 100 + 10 = 110 cases. Then:
Week 1 level = 0.5 × 112 + 0.5 × (100 + 10) = 111
Week 1 trend = 0.5 × (111 − 100) + 0.5 × 10 = 10.5
Week 2 forecast = 111 + 10.5 = 121.5
Demand below that second forecast pulls the updated level to 119.75 and trend to 9.625. The next prediction therefore becomes 119.75 + 9.625 = 129.375 cases. The trend changes because it is smoothed after every actual, rather than held at its starting value.
At the end of week 4, the unrounded level is 141.609375 and trend is 10.2265625. Freeze that forecast version before weeks 5 and 6 occur:
Week 5 forecast = 141.609375 + 1 × 10.2265625 = 151.8359375
Week 6 forecast = 141.609375 + 2 × 10.2265625 = 162.0625
The difference between adjacent forecasts is 10.2265625, exactly the latest trend. That identity checks the forecast extension. Separately, substituting each row into both update equations checks that the level and trend were calculated in the right order. For a worksheet implementation, use double exponential smoothing in Excel. Covirage is not affiliated with Microsoft.
Weeks 5 and 6 later record 146 and 160 cases. Define signed error as actual minus forecast: negative means the forecast was too high. The table again displays four decimals, while totals use unrounded figures.
| Target week | Horizon from week 4 | Actual cases | Frozen forecast | Signed error | Absolute error |
|---|---|---|---|---|---|
| 5 | 1 | 146 | 151.8359 | −5.8359 | 5.8359 |
| 6 | 2 | 160 | 162.0625 | −2.0625 | 2.0625 |
| Total | 306 | 313.8984 | −7.8984 | 7.8984 |
Total signed error must equal total actual minus total forecast: 306 − 313.8984375 = −7.8984375 cases. Mean absolute error across this planning window is 7.8984375 / 2 = 3.94921875 cases, or 3.9492 displayed. The MAD forecast accuracy guide explains that unit-based measure.
A last-observation baseline predicts 140 cases for both weeks. Its absolute misses are 6 and 20 cases, giving a mean of 13 cases. Holt performs better on this tiny window, but two future observations cannot establish a reliable advantage.
This example combines one- and two-week horizons to score a single frozen planning window. For method selection, also compare repeated one-week forecasts with one another and repeated two-week forecasts with one another. Use many forecast origins, exclude future actuals from fitting, and check whether the improvement persists by product and location. The forecast methods comparison explains why methods can disagree.
Bring your saved trend forecasts and actual demand export to Covirage, which maps the uploaded columns. Its tools compute and check period errors and totals; the external AI model explains the results with every figure traced to its rows. Create a free Covirage account, then set up your business and bring a sample of your own data.
See how this supports forecast analysis.
Holt's level-and-trend method is commonly called double exponential smoothing because it updates two components. The label can also refer to other formulations, so state the equations and smoothing parameters when comparing forecasts from different workbooks.
Holt's method updates a level and trend without a seasonal component. Holt-Winters adds seasonality. If demand repeatedly peaks in the same calendar periods, evaluate a seasonal method rather than expecting an ordinary trend adjustment to capture that pattern.
Choose parameters using historical training periods and assess the resulting forecasts on later periods excluded from fitting. Larger values react more quickly to changes, but can also react to noise. Report the initialization and evaluation horizon alongside the parameters.
Yes. A sufficiently negative estimated trend can produce negative forecasts at longer horizons. For physical demand, inspect that outcome before using it operationally. A constrained alternative or explicit planning adjustment needs its own evaluation and documentation.