In the previous chapters, they mentioned the assumptions of a simple and multiple regression. When it comes to a time series, do these assumptions still need to be checked? Thanks
For serial correlation & such? I believe so
Is serial correlation the only condition we need to check? We didnt check for heteroskedasticity or multicollinearity?
I noticed that no t-test were conducted to check for the significance of the coefficients
If you’re running a linear regression, even though it’s time series data, you still need to check the other assumptions. It’s a very similar process. You can still conduct significance tests on the coefficients.
Also, (absence of) multicollinearity is not an assumption. The assumption you might be referring to is a lack of perfect collinearity (if this is case, you can’t estimate the model).
Hope this helps!
To echo tickersu’s point—yes you need to use the DW test for auto/serialcorrelation of residuals and the BP test for heteroskedasticity of your residuals. Generally speaking, both of these will underestimate your standard errors and thus overestimate your t-statistics for the hypothesis tests that your regression coefficients are useful.
With a time series, where your x-axis dependent output is related to prior y-axis independent input values of time, I think it’s fair to say we expect certain particular regression violations to occur more frequently.
To try to put this into perspective… With financial data, your violations are almost always going to be positive serialcorrelation and conditional heteroskedasticity. If you’ve ever pulled up a chart, serialcorrelation is why some people follow trends of stock prices and use a moving average. Also, I believe conditional heteroskedasticity is related to the Level I concept of leptokurtic distributions and how you get pockets of volatility that even out over time.
If the model has autocorrelated error terms, you want to consider using an autoregressive model. Or maybe a moving average model. Or combine them and use an ARMA model. But, remember, here you need to test for unit root (random walks) or unit root with a drift.
If the model has conditional heteroskedasticity, you can use an ARCH model, or some variant of it, to make use of it.
Sorry, I didnt mean only serial correlation, it was the only one that popped into my head at the moment. Thats why I said & such, as to imply you should still use the regular tests.
You may have to maniuplate the data & as such you can run log & log difference time series regressions as well
Thanks gents
Just general curiosity. Which topic is all this from ?
Quantitative Methods - Time Series Analysis.
Thanks.