A Regression analysis and Time Series decomposition

[iCloud]

If we can use Regression analysis to forecast, why do we use “Time Series Decomposition”?
What's the difference between the 2?

Thanks

 

[BvU]

Hi Cloudi,

Maybe you want to be a bit more specific: forecasting is difficult, especially when it concerns the future. From a brief search (I can't really pinpoint time series decomposition -- it seems to be pretty general) I suspect regression analysis when applied to temporal data is one form of time series deco.

 

[iCloud]

Hi Cloudi,

Maybe you want to be a bit more specific: forecasting is difficult, especially when it concerns the future. From a brief search (I can't really pinpoint time series decomposition -- it seems to be pretty general) I suspect regression analysis when applied to temporal data is one form of time series deco.

Well I just started with regression, and there was this general question without specifics. I am just at the basics of the regression and multiple-regression now.

 

[MarneMath]

Well, your question is a bit to broad to have a meaningful answer. There are many different methods for regression that each have their own limits and assumptions. For example, if you try a linear regression and your data has auto-correlation, this may skew your results. Decomposition of a time series is a useful way to discover and use seasonality and periodic information into your model. It's a fairly well known result that trying to add seasonality into your data via a dummy variable may actual detrend your results, so in that case a linear regression may not be the best method.

 

[FactChecker]

Time series analysis addresses timing issues directly. Suppose that the Y value of one step is partially (or wholly) dependent on the prior step value of Y. The current values of input variables X_i will not explain all of the current Y value. In fact, you might have to go back a very long way to see where those prior Y influences got started. They may even be initial conditions with no explanation at all in terms of X values. And the influence of a prior Y value might skip several steps before they enter in. The same is true of prior X values, which might not take effect till several steps later. Time series analysis addresses those timing concerns directly and quantifies them. Addressing these issues with regression would be very difficult.

 

[iCloud]

Thank you both, that helps.

 

[chiro]

Hi iCloud.

Time series looks at a variety of correlation values in a very structured way while regression typically [and I say typically] doesn't include that information.

When you do statistics you do inference on parameters and in time series, the parameters you use typically involve some sort of correlation vector where-as regression looks at estimating coefficients for some model in terms of a population expectation [so a mean as a function of the values of different variables].