8.2 Models with trend

Holt’s linear trend method

Holt (1957) extended simple exponential smoothing to allow the forecasting of data with a trend. This method involves a forecast equation and two smoothing equations (one for the level and one for the trend): \[\begin{align*} \text{Forecast equation}&& \hat{y}_{t+h|t} &= \ell_{t} + hb_{t} \\ \text{Level equation} && \ell_{t} &= \alpha y_{t} + (1 - \alpha)(\ell_{t-1} + b_{t-1})\\ \text{Trend equation} && b_{t} &= \beta^*(\ell_{t} - \ell_{t-1}) + (1 -\beta^*)b_{t-1}, \end{align*}\] where \(\ell_t\) denotes an estimate of the level of the series at time \(t\), \(b_t\) denotes an estimate of the trend (slope) of the series at time \(t\), \(\alpha\) is the smoothing parameter for the level, \(0\le\alpha\le1\), and \(\beta^*\) is the smoothing parameter for the trend, \(0\le\beta^*\le1\). (We denote this as \(\beta^*\) instead of \(\beta\) for reasons that will be explained in Section 8.5.)

As with simple exponential smoothing, the level equation here shows that \(\ell_t\) is a weighted average of observation \(y_t\) and the one-step-ahead training forecast for time \(t\), here given by \(\ell_{t-1} + b_{t-1}\). The trend equation shows that \(b_t\) is a weighted average of the estimated trend at time \(t\) based on \(\ell_{t} - \ell_{t-1}\) and \(b_{t-1}\), the previous estimate of the trend.

The forecast function is no longer flat but trending. The \(h\)-step-ahead forecast is equal to the last estimated level plus \(h\) times the last estimated trend value. Hence the forecasts are a linear function of \(h\).

Example: Australian population

In Table 8.2 we demonstrate the application of Holt’s method to annual population totals for Australia. The smoothing parameters, \(\alpha\) and \(\beta^*\), and the initial values \(\ell_0\) and \(b_0\) are estimated by minimising the SSE for the one-step training errors as in Section 8.1.

Table 8.2: Forecasting the goods and services exports for Algeria using simple exponential smoothing.
Year Time Observation Level Forecast
\(t\) \(y_t\) \(\ell_t\) \(\hat{y}_{t+1|t}\)
1959 0 10.05
1960 1 10.28 10.28 10.28
1961 2 10.48 10.48 10.50
1962 3 10.74 10.74 10.70
1963 4 10.95 10.95 10.97
1964 5 11.17 11.17 11.17
1965 6 11.39 11.39 11.39
1966 7 11.65 11.65 11.61
2014 55 23.50 23.50 23.52
2015 56 23.85 23.85 23.87
2016 57 24.21 24.21 24.21
2017 58 24.60 24.60 24.57
\(h\) \(\hat{y}_{T+h|T}\)
2018 1 24.97
2019 2 25.34
2020 3 25.71
2021 4 26.07
2022 5 26.44
2023 6 26.81
2024 7 27.18
2025 8 27.55
2026 9 27.92
2027 10 28.29

The very small value of \(\beta^*\) means that the slope hardly changes over time. In this case, \(\beta^* = 0.3266\) is fairly large, suggesting that the trend changes often (even if the changes are slight).

Damped trend methods

The forecasts generated by Holt’s linear method display a constant trend (increasing or decreasing) indefinitely into the future. Empirical evidence indicates that these methods tend to over-forecast, especially for longer forecast horizons. Motivated by this observation, Gardner & McKenzie (1985) introduced a parameter that “dampens” the trend to a flat line some time in the future. Methods that include a damped trend have proven to be very successful, and are arguably the most popular individual methods when forecasts are required automatically for many series.

In conjunction with the smoothing parameters \(\alpha\) and \(\beta^*\) (with values between 0 and 1 as in Holt’s method), this method also includes a damping parameter \(0<\phi<1\): \[\begin{align*} \hat{y}_{t+h|t} &= \ell_{t} + (\phi+\phi^2 + \dots + \phi^{h})b_{t} \\ \ell_{t} &= \alpha y_{t} + (1 - \alpha)(\ell_{t-1} + \phi b_{t-1})\\ b_{t} &= \beta^*(\ell_{t} - \ell_{t-1}) + (1 -\beta^*)\phi b_{t-1}. \end{align*}\] If \(\phi=1\), the method is identical to Holt’s linear method. For values between \(0\) and \(1\), \(\phi\) dampens the trend so that it approaches a constant some time in the future. In fact, the forecasts converge to \(\ell_T+\phi b_T/(1-\phi)\) as \(h\rightarrow\infty\) for any value \(0<\phi<1\). This means that short-run forecasts are trended while long-run forecasts are constant.

In practice, \(\phi\) is rarely less than 0.8 as the damping has a very strong effect for smaller values. Values of \(\phi\) close to 1 will mean that a damped model is not able to be distinguished from a non-damped model. For these reasons, we usually restrict \(\phi\) to a minimum of 0.8 and a maximum of 0.98.

Example: Australian Population (continued)

Figure 8.3 shows the forecasts for years 2018–2032 generated from Holt’s linear trend method and the damped trend method.

Forecasting total annual passengers of air carriers registered in Australia (millions of passengers, 1990--2016). For the damped trend method, $\phi=0.90$.

Figure 8.3: Forecasting total annual passengers of air carriers registered in Australia (millions of passengers, 1990–2016). For the damped trend method, \(\phi=0.90\).

We have set the damping parameter to a relatively low number \((\phi=0.90)\) to exaggerate the effect of damping for comparison. Usually, we would estimate \(\phi\) along with the other parameters. We have also used a rather large forecast horizon (\(h=15\)) to highlight the difference between a damped trend and a linear trend.

Example: Internet usage

In this example, we compare the forecasting performance of the three exponential smoothing methods that we have considered so far in forecasting the number of users connected to the internet via a server. The data is observed over 100 minutes and is shown in 8.4.

Users connected to the internet through a server

Figure 8.4: Users connected to the internet through a server

We will use time series cross-validation to compare the one-step forecast accuracy of the three methods.

Damped Holt’s method is best whether you compare MAE or RMSE values. So we will proceed with using the damped Holt’s method and apply it to the whole data set to get forecasts for future years.

The smoothing parameter for the slope is estimated to be almost one, indicating that the trend changes to mostly reflect the slope between the last two minutes of internet usage. The value of \(\alpha\) is very close to one, showing that the level reacts strongly to each new observation.

Forecasting internet usage: comparing forecasting performance of non-seasonal methods.

Figure 8.5: Forecasting internet usage: comparing forecasting performance of non-seasonal methods.

The resulting forecasts look sensible with decreasing trend, and relatively wide prediction intervals reflecting the variation in the historical data. The prediction intervals are calculated using the methods described in Section 8.5.

In this example, the process of selecting a method was relatively easy as both MSE and MAE comparisons suggested the same method (damped Holt’s). However, sometimes different accuracy measures will suggest different forecasting methods, and then a decision is required as to which forecasting method we prefer to use. As forecasting tasks can vary by many dimensions (length of forecast horizon, size of test set, forecast error measures, frequency of data, etc.), it is unlikely that one method will be better than all others for all forecasting scenarios. What we require from a forecasting method are consistently sensible forecasts, and these should be frequently evaluated against the task at hand.

Bibliography

Gardner, E. S., & McKenzie, E. (1985). Forecasting trends in time series. Management Science, 31(10), 1237–1246. https://doi.org/10.1287/mnsc.31.10.1237

Holt, C. E. (1957). Forecasting seasonals and trends by exponentially weighted averages (O.N.R. Memorandum No. 52). Carnegie Institute of Technology, Pittsburgh USA. https://doi.org/10.1016/j.ijforecast.2003.09.015