ARIMA - Add support for missing observations and padding #4058
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
Removing |
|
I solved a bug of numerical issues that could result in NaNs, changed one of the new test cases to a more stable ARIMA order, and relaxed the tolerance of confidence intervals tests. The PR should be ready to merge if CI passes. |
|
rerun tests |
Nyrio
added
4 - Waiting on Reviewer
|
I have fixed some C++ tests that were failing due to a change in the Jones transform. |
|
@dantegd Can you please update your review? |
|
rerun tests |
Codecov Report
@@ Coverage Diff @@
## branch-21.10 #4058 +/- ##
===============================================
Coverage ? 86.20%
===============================================
Files ? 231
Lines ? 19072
Branches ? 0
===============================================
Hits ? 16441
Misses ? 2631
Partials ? 0
Flags with carried forward coverage won't be shown. Click here to find out more. Continue to review full report at Codecov.
|
dantegd
approved these changes
Sep 22, 2021
|
@gpucibot merge |
raydouglass
approved these changes
Sep 22, 2021
rapids-bot bot
pushed a commit
that referenced
this pull request
Oct 11, 2021
The formula that we have been using since I added support for confidence intervals in ARIMA is slightly different than the one used in statsmodels. The difference is in particular quite pronounced when datasets have missing observations, which pushed me to raise tolerance for the intervals unit tests when I added test cases in the recent PR #4058. In this PR, I change our calculation to match statsmodels, and decrease the corresponding test tolerance, as we now have a strict match with statsmodels. Previous formula: ```python lower_t = fc_t - sqrt(2) * erfinv(level) * sqrt(F_t * mean(v_i**2 / F_i)) upper_t = fc_t + sqrt(2) * erfinv(level) * sqrt(F_t * mean(v_i**2 / F_i)) ``` New formula: ```python lower_t = fc_t - sqrt(2) * erfinv(level) * sqrt(F_t) upper_t = fc_t + sqrt(2) * erfinv(level) * sqrt(F_t) ``` Authors: - Louis Sugy (https://github.com/Nyrio) Approvers: - Tamas Bela Feher (https://github.com/tfeher) - Dante Gama Dessavre (https://github.com/dantegd) URL: #4248
vimarsh6739
pushed a commit
to vimarsh6739/cuml
that referenced
this pull request
Oct 9, 2023
This PR allows support for missing observations and padding at the start for variable-length batch. Example:  Note: I had to change ARIMA tests because I used a different method than statsmodels (which is used as a reference in tests) to compute the initial parameter estimation. They cut all missing observations for their initial least-square estimation, and I decided to fill them with naive replacements instead, so I keep the temporal relationships in the data and have a much better initial estimate and often a better fit in the end, according to some MASE measurements I made. So I updated the integration test to use the MASE and pass if we are approximately the same _or better_ than statsmodels. Authors: - Louis Sugy (https://github.com/Nyrio) Approvers: - Robert Maynard (https://github.com/robertmaynard) - Tamas Bela Feher (https://github.com/tfeher) - Dante Gama Dessavre (https://github.com/dantegd) - Ray Douglass (https://github.com/raydouglass) URL: rapidsai#4058
vimarsh6739
pushed a commit
to vimarsh6739/cuml
that referenced
this pull request
Oct 9, 2023
The formula that we have been using since I added support for confidence intervals in ARIMA is slightly different than the one used in statsmodels. The difference is in particular quite pronounced when datasets have missing observations, which pushed me to raise tolerance for the intervals unit tests when I added test cases in the recent PR rapidsai#4058. In this PR, I change our calculation to match statsmodels, and decrease the corresponding test tolerance, as we now have a strict match with statsmodels. Previous formula: ```python lower_t = fc_t - sqrt(2) * erfinv(level) * sqrt(F_t * mean(v_i**2 / F_i)) upper_t = fc_t + sqrt(2) * erfinv(level) * sqrt(F_t * mean(v_i**2 / F_i)) ``` New formula: ```python lower_t = fc_t - sqrt(2) * erfinv(level) * sqrt(F_t) upper_t = fc_t + sqrt(2) * erfinv(level) * sqrt(F_t) ``` Authors: - Louis Sugy (https://github.com/Nyrio) Approvers: - Tamas Bela Feher (https://github.com/tfeher) - Dante Gama Dessavre (https://github.com/dantegd) URL: rapidsai#4248
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR allows support for missing observations and padding at the start for variable-length batch. Example:
Note: I had to change ARIMA tests because I used a different method than statsmodels (which is used as a reference in tests) to compute the initial parameter estimation. They cut all missing observations for their initial least-square estimation, and I decided to fill them with naive replacements instead, so I keep the temporal relationships in the data and have a much better initial estimate and often a better fit in the end, according to some MASE measurements I made. So I updated the integration test to use the MASE and pass if we are approximately the same or better than statsmodels.