I have read that it can be used in forecasting. so what is the difference of this method and the usual method in forecasting? And also, I don't understand the concept that this method assigns geometrically decreasing weights? what's the meaning of that? Please help!
Hello Jane,
First of all, you can't find a unique "usual method" in forecasting, since it depends on what you are trying to predict, so you have to consider different parameters and peculiarities of the data you are examining. For example, the SES is particularly useful in case you want to forecast the trend of material with a short life-cycle.
This happens because of the specific structure of the SES itself: thanks to the α parameter you can have a forecast more or less reactive (that is more or less similar to the original data: the higher the parameter is, the smaller noise you find in your forecast).
Still because of the specific properties of the model, we say it assigns decreasing weights. In fact, the model acts like a weighted moving average where the weights are α and (1- α): so the newer data are coupled with a higher weight and they're considered more influential ("more important") on the final forecast.
Best regards,
DC
AS DC says, the choice of forecasting method should be influenced by your beliefs about how the variable behaves / will behave. Simple exponential smoothing is suited to forecasting a variable that has some slowly changing and unknown expected value and partially adjusts towards the current expected value at each observation date. The geometrically declining weights then make sense because the more recent observations are more informative about the current expected value.
Exponential smoothing is an endogenous prevision method wich is based on the underground historical progress of the observed data.
The choice of the exponential method depends on the degree of stability of the time series (Holt exponential smoothing for series with trend component, Winters exponential smoothing for series with exponential and saisonal components)...
Exponential smoothing does come in different degrees of complexity. For example, you can decide whether or not you want to account for seasonality. And it does let you put more emphasis on more recent data, which can help when you may be having some changes over time. But it is still a forecasting method that takes a time series and anticipates the next value (or values) in the series. Thus, if there is an event that causes a break in a time series, and that happens in the period forecasted, it will not work well. "Prediction" sounds like forecasting, but it is not, because prediction looks at a current sample, relates it to one or more regressors, and estimates ("predicts") for the missing data. Model parameters would actually be 'up to date.' Exponential smoothing cannot overcome a current event/change, that is sudden. But it can somewhat avoid problems from previous changes by emphasizing somewhat newer data. - Exponential smoothing is used for data from a single 'respondent' over instances in time, whereas prediction is used for a category/group of respondents, sampled at a given time, with regressor data that may be from current or past time. I have seen both exponential smoothing and prediction used for imputation for missing data. An advantage of prediction is that it is consistent with the current data, and variance estimation in the application I have in mind, is consistently defined.
The question necessitates and answer that is far to complex/long to explain here. There is a very good paper by Gooijer and Hyndman (2006) in the Intnl. Jnl. Forecasting which provides explanations
http://www.sciencedirect.com/science/article/pii/S0169207006000021
- Badges
- Science topic
- Similar topics
- Medicine
- Physiology
- Biosignals
- Biomedical Signal Processing
- Statistical Analysis