Issue 275: Prototype VARIMA (Analysis - do not merge) #291
Conversation
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## main #291 +/- ##
==========================================
- Coverage 16.54% 15.87% -0.67%
==========================================
Files 21 21
Lines 1481 1537 +56
==========================================
- Hits 245 244 -1
- Misses 1236 1293 +57
Flags with carried forward coverage won't be shown. Click here to find out more. ☔ View full report in Codecov by Sentry. |
|
Adding a bit of takeaway: In the short term I this analysis plan doesn't give a compelling reason to move away from the current time series baseline (linear pool ensemble of auto-selected ETS + SARIMA univariate ts models) because one of those models is always in the mix of best performing and therefore I don't think gives an "erroneous" baseline to compare the main product However, there are some interesting angles e.g. stacking models that are good in different situation ala |
This draft PR explores the VARIMA baseline model as per #275 and commits an analysis of different VARIMA structures across a few chosen dates with a few chosen ARIMA models aimed at covid and flu NHSN.
This is committed as an analysis here.
Whilst this is not an exhaustive analysis, it shows that the "best" model, as in what time series joint modelling configs have the best out-of-sample CRPS, varies over time and pathogen. This suggests that for multi-signal approaches we probably want to go towards some kind of ensemble over structures to have more consistent forecast performance.
From the pov of a baseline against pyrenew; SARIMA still seems reasonable because it was never the worst time series model.
Linking @seabbs because he was interested in VARIMA.