Exponential smoothing This is a widely used forecasting technique in retailing, even though it has not proven to be especially accurate.
Why is exponential smoothing so popular? It's easy—the exotic term notwithstanding. Data storage requirements are minimal (even though this is not the problem it once was due to plunging memory prices). It is very cost effective when forecasts must be made for a large number of items--hence it has extensive use in retailing.
The basic algorithm (1) Where: Lt is the forecast for the current period; Xt is the most recent observation of the time series variable—such as, for example, sales last month of part #000897 Lt-1 is the most recent forecast; and is the smoothing constant, where 0 < < 1
Equation (1) can be written as follows: New Forecast = (New Data) + (1 - )Most Recent Forecast
Exponential smoothing is weighted moving average process To demonstrate, let (2) Substitute (2) into (1): (3)
But notice that: Substitute (4) into (3) to obtain: If we continue to substitute recursively, we get:
Notice that are the weights attached to past values of X. Since < 1, the weights attached to earlier or remoter observations of X are diminishing.
You don’t have to go through this recursive process each time you do a forecast. The process is summarized in the most recent forecast.
Selecting the smoothing constant () ?alpha? The range of possible values is zero and one. If you select a value of close to 1, that means you are attaching a large weight to the most recent observation. This is not indicated if your series is very choppy. For example, suppose you were forecasting the demand for part #56 in month t. If you attached too much weight to the observation for t-1, you will have a large forecast error for month t. Sales of part #56 t-2 t-1 t Month
Application We will now forecast sales of liquor and floor covering using this technique. We have monthly data for each variable beginning in January 1995 and running through July of 2000.
Summary statistics for monthly sales of floor covering and liquor sales, 1995:1 to 2000:7 (in millions of dollars) Floor Covering Liquor Mean 1106.687 2077.642 Median 1079.000 2043.000 Maximum 1420.000 3487.000 Minimum 823.0000 1502.000 Std. Dev. 140.4320 350.4939 Observations 67
Liquor = 0.169 Floor covering = 0.127 The ratio of the standard deviation to the mean gives us a nice measure of the amplitude or volatility of a series month-to-month (or day-to-day, quarter-to-quarter, as the case may be).
Selecting the smoothing constant Pricey time series forecasting software, such as EViews, use an algorithm to select the value of the smoothing constant that minimizes mean square error for in-sample forecasts. If you lack this software, you can use a trial and error process.
The first set of estimates for monthly floor covering and liquor were produced by using the algorithm that selects the best performing value of the smoothing constant () for in-sample forecasts. The second set of estimates is based on values of alpha () arbitrarily selected by the instructor.
Computer algorithm selects alpha to minimize MSE
Actual and smoothed values of floor covering, 1997:7 to 2000:7 (all data in millions of dollars) Alpha = 0.706
Alpha selected arbitrarily
Sum of Squared Residuals Statistics for the floor covering estimates Alpha Sum of Squared Residuals Root Mean Square Error 0.30 600529.2 94.67 0.7060 563741.2 91.73 Data is for 1995:1 to 2000:7
Computer algorithm selects alpha to minimize MSE
Actual and smoothed values of liquor sales, 1997:7 to 2000:7 (all data in millions of dollars) Alpha = 0.122
Alpha selected arbitrarily
Sum of Squared Residuals Statistics for the liquor estimates Alpha Sum of Squared Residuals Root Mean Square Error 0.45 7888362 343.1 0.122 6540216 312.4 Data is for 1995:1 to 2000:7
Forecasts for August, 2000 Remember our basic algorithm Hence to forecast floor covering sales for August, 2000: Floor CoveringAUG=(0.706)(1420) + [(1 - .706)(1375)] = $1406.77 To forecast liquor sales LiquorAUG=(0.122)(2560) + [(1 - .122)(2349)] = $2374.72