Timeline for time series Diebold-Mariano Test n-ahead forecast in R
Current License: CC BY-SA 4.0
11 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| May 17, 2021 at 9:08 | comment | added | Richard Hardy | @Ivan, that sounds correct. | |
| May 17, 2021 at 8:09 | comment | added | Ivan | Thank you, Professor, for your detailed explanation, I have benefited a lot. I think I probably understand that h-step-forecast will be a rolling forecast in dm.test, not a fixed from time t. | |
| May 16, 2021 at 10:29 | comment | added | Richard Hardy | @Ivan, yes, your logic up to the question mark makes sense. I have not checked the article, but plenty of online materials contain mistakes or imprecisions, so consider this a possibility (not to say this actually is the case -- just a possibility). The statement It is generally sufficient to use the value $g = n^{1/3} + 1$ (I use $g$ to distinguish from the forecast horizon $h$) suggests an approximation to the naive estimate of long-run variance with $h$ lags when $h>n^{1/3} + 1$, and it may or may not be a good one depending on the details of the problem. | |
| May 16, 2021 at 9:42 | comment | added | Ivan | Hello Professor! But I'm thinking , is it because of n-step-forecast with fixed t, the uncertainty of out of sample prediction will increase with the increase of time t, result d mean(loss-differential) and VaR (d) presenting non-stationary, so that can't be used? In article, the excel example seems doesn't mention fixed n-step-forecast or one-step-forecast, and set h=n^(1/3)+1 even have 20 sample of forecast. | |
| May 16, 2021 at 8:58 | vote | accept | Ivan | ||
| May 16, 2021 at 6:24 | history | edited | Richard Hardy | CC BY-SA 4.0 |
added 60 characters in body
|
| May 16, 2021 at 6:24 | comment | added | Richard Hardy | @Ivan, now I have noticed you have edited your question. I see what you are doing there. I covered this in (This is in contrast to predicting $1,\dots,,h$ steps ahead from a fixed origin at time $t$.) To reiterate, you cannot apply the Diebold-Mariano test in such a situation. | |
| May 16, 2021 at 6:19 | comment | added | Richard Hardy | @Ivan, I do not understand your question, could you try to state it even more clearly? I have covered the most common types of applications of the DM test in my answer. I do not think the DM test is suited for other types of applications, so if yours is such, then you probably cannot apply the test. | |
| May 15, 2021 at 22:16 | comment | added | Ivan | Thanks for your reply, but I still have some questions. I read the dm.test example from forecast package, in the example select dm.test( ,h=1) becasuse of using one-ahead-forecast to predict ? In this case, is it reasonable if I select dm.test( ,h = 2、3)? Then If predict is n.ahead-forecast which dm.test forecast horizon should use?(is dm.test(h=1) acceptable?) | |
| May 15, 2021 at 20:03 | history | edited | Richard Hardy | CC BY-SA 4.0 |
added 167 characters in body
|
| May 15, 2021 at 19:57 | history | answered | Richard Hardy | CC BY-SA 4.0 |