1

I am trying to determine the factors driving bitcoin prices. I have a time series set of data with my dependent variable as bitcoin prices (denominated in USD). I have a set of explanatory variables including bitcoin-related metrics as well as non-bitcoin metrics. following some research on time series and data transformation, I understood I needed to transform my dependent variable with the Box-Cox method, which I did.

I am now confused on which transformation I need to perform on my independent variables. Do I also need to perform Box-Cox transformations?

Or maybe did I get it completely wrong ?

For info, I am using XL stat.

kjetil b halvorsen
  • 77,844
Tomque
  • 11
  • Without knowing why you are considering transforming this is going to be rather hard to answer. – mdewey Oct 26 '17 at 14:46
  • Sorry for the lack of precision. I am trying to make the data stationary and normal to run a linear regression – Tomque Oct 26 '17 at 14:54
  • For an indication of one way to approach this problem, see https://stats.stackexchange.com/a/35717/919. – whuber Oct 26 '17 at 16:23
  • @Tomque Why would making the data appear normal be required for regression? – Glen_b Oct 27 '17 at 04:36
  • @Glen_b, I am not that knowledgeable in statistics and the only regressions I made using time series were linear ones which required normality in the data – Tomque Oct 27 '17 at 09:38
  • The usual linear regression doesn't assume normality of the data for anything. There is a normality assumption for particular things (e.g. the standard hypothesis tests) but the assumption isn't that either the DV nor the IVs are normal. Do you have a text that says otherwise? Or is the regression not the usual linear regression? – Glen_b Oct 27 '17 at 09:53
  • To elaborate, I transformed my dependent variable with Box Cox and differencing to make the data stationary (i.e. removing trends and seasonality, correct me if I am wrong) because, otherwise, running an OLS regression would have made no sense in the interpretation of the outputs. After this transformation, the Jarque Bera test indicates that my DV follows a normal distribution. Now that my dependent variable has been transformed, it makes sense that I transform my independent variables to some extent doesn't it ? – Tomque Oct 27 '17 at 10:07
  • DIfferencing for stationarity is (usually) important (though stationarity alone may not be sufficient if you're doing it to avoid spurious regression effects). If you're primarily transforming your IVs simply to achieve normality I suggest that you reconsider as there's no need for that (however there are a number of other reasons to do it). You might want to clarify your question with additional details (some of which are here in comments) – Glen_b Oct 28 '17 at 05:06
  • Thanks a lot @Glen_b, if I can restate this from the beginning. I have daily data for bitcoin prices (ranging from 2010 to 2017) as DV and about 13 other unmodified independent variables which I want to assess the influence on the DV. As i am looking at time series data, what analysis would you recommend so I can measure the impact of the IVs on my DV ? – Tomque Oct 28 '17 at 07:15
  • You're probably getting closer to a good question to be asking there, though I'm not the best person to answer that (lots of people with better grounding in econometrics than me who will have smart things to say). If the IVs are also time series I would first be investigating for cointegration. If the IVs are not time series I would not necessarily be looking to transform them (unless it was needed to get a suitable relationship describing the conditional mean. – Glen_b Oct 28 '17 at 07:24
  • Thanks @Glen_b, all my data is time based (and all my data points are on the same days). I have around 1600 observations which will probably decrease due to outliers and data modification. Do you know where I could find a good tutorial or paper to move forward? I can't seem to find practical examples. Thanks a lot for your help – Tomque Oct 28 '17 at 07:34
  • I'd try to reorient your question toward what you're trying to ask me there (or ask a new one with tag references); if I had a good basic tutorial to offer I'd also probably also be in a good place to answer the question. For basic time series some of the advice at Forecasting Principles and Practice (https://www.otexts.org/fpp/ -- scroll down for the links) may be useful but it probably won't get you to the point you need to be at for this task, I think. It may help get you started though. – Glen_b Oct 28 '17 at 07:48

0 Answers0