Feat/statsforecasts #893
Feat/statsforecasts #893
Conversation
hrzn
requested review from
tomasvanpottelbergh,
dennisbader and
brunnedu
as code owners
Codecov Report
@@ Coverage Diff @@
## master #893 +/- ##
==========================================
- Coverage 91.57% 91.55% -0.02%
==========================================
Files 71 73 +2
Lines 7165 7246 +81
==========================================
+ Hits 6561 6634 +73
- Misses 604 612 +8
Continue to review full report at Codecov.
|
|
|
||
| mu = forecast_df["mean"].values | ||
| if num_samples > 1: | ||
| std = 2 * (forecast_df["hi_68%"].values - mu) |
There was a problem hiding this comment.
Why is there a factor 2? Isn't hi_68 - mu equal to 1 std?
There was a problem hiding this comment.
Unless I missed something, With a 68% forecast interval, it means that "hi - mu" represents 34% of the mass. (and "mu - low" represents the other 34%). So I multiply by 2 to get back one std (2x 34%). Makes sense?
There was a problem hiding this comment.
OK my bad that was right ;) Now fixed.
| from darts.timeseries import TimeSeries | ||
|
|
||
|
|
||
| class Croston(ForecastingModel): |
There was a problem hiding this comment.
Apparently TSB and croston_sba are improved version of Croston (https://www.lancaster.ac.uk/pg/waller/pdfs/Intermittent_Demand_Forecasting.pdf), might be worth proposing each of the 3 options in the same function as they already exist in statsforecast?
I've implemented a couple of models coming from Statsforecasts. The most interesting is probably their implementation of AutoARIMA, which uses Numba (and jit compilation) and can in some cases be much faster than pmdarima (after the first call to
fit()). I would propose that we add it first as a separate model, and if everything goes well and it proves superior to the current AutoARIMA version, we could even replace it entirely at some point (it is not quite the same model, though).