Drop ARIMA when month_start freq is detected

[ParthivNaresh]

Fixes #2631

[codecov]

Codecov Report

Merging #2632 (fcedb3f) into main (06c21cf) will increase coverage by 0.1%.
The diff coverage is 100.0%.

@@           Coverage Diff           @@
##            main   #2632     +/-   ##
=======================================
+ Coverage   99.9%   99.9%   +0.1%     
=======================================
  Files        297     297             
  Lines      27037   27044      +7     
=======================================
+ Hits       26993   27000      +7     
  Misses        44      44             

Impacted Files	Coverage Δ
evalml/automl/automl_search.py	99.9% <100.0%> (+0.1%)	⬆️
...omponents/estimators/regressors/arima_regressor.py	100.0% <100.0%> (ø)
evalml/tests/automl_tests/test_automl.py	99.7% <100.0%> (+0.1%)	⬆️
...ests/automl_tests/test_automl_search_regression.py	100.0% <100.0%> (ø)
...alml/tests/component_tests/test_arima_regressor.py	100.0% <100.0%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 06c21cf...fcedb3f. Read the comment docs.

[ParthivNaresh]

Sneaking this in here because it didn't go through in my last ARIMA PR (just meant to save time from 20s to 10s)

[bchen1116]

LGTM! Just left a question about one of the tests

[bchen1116]

Why did we need to add this to the test here?

[ParthivNaresh]

Originally I had added it because the code I included in AutoMLSearch would throw an error if the data didn't have a date_index column specified, which X_y_binary didn't have. Decided to just clean up the test and replace the data with what I added, thanks!

[freddyaboulton]

@ParthivNaresh When I run this repro for Arima with MS frequency on main I get a value error but it's coming from our own arima implementation.

from evalml.pipelines.components import ARIMARegressor
from evalml.demos import load_diabetes
import pandas as pd

X, y = load_diabetes()
X.ww["Date"] = pd.Series(pd.date_range(start="1/1/2018", periods=X.shape[0], freq="MS"))
arima = ARIMARegressor(date_index="Date")
arima.fit(X, y)

When I delete the line that sets the frequency to M for MS, i.e. freq = pd.infer_freq(dates), the above repro doesn't throw an error anymore and the arima is able to fit:

Unit tests also pass. So I'm wondering whether we need to make the changes to AutoMLSearch?

[ParthivNaresh]

@freddyaboulton If that line is changed to freq = pd.infer_freq(dates) the error just gets pushed back to predict.

from evalml.pipelines.components import ARIMARegressor
from evalml.demos import load_diabetes
import pandas as pd
import numpy as np
X = pd.DataFrame()
y = pd.Series(np.random.randint(1, 10, 10))
X["Date"] = pd.Series(pd.date_range(start="1/1/2018", periods=10, freq="MS"))

arima = ARIMARegressor(date_index="Date")
arima.fit(X, y)
arima.predict(X)
------------------------------------------
   481         msg = str(e)
    482         if "Invalid frequency" in msg or "_period_dtype_code" in msg:
--> 483             raise ValueError(
    484                 "Invalid frequency. Please select a frequency that can "
    485                 "be converted to a regular `pd.PeriodIndex`. For other "

ValueError: Invalid frequency. Please select a frequency that can be converted to a regular `pd.PeriodIndex`. For other frequencies, basic arithmetic operation to compute durations currently do not work reliably.

[freddyaboulton]

Thank you @ParthivNaresh !

[freddyaboulton]

I'm on the fence as to whether this should live here or in get_estimators. I'm ok keeping this here. Is this a bug that sktime can fix in the future or is it a limitation of the MS frequency?

[ParthivNaresh]

The only reason I chose to put it here instead of get_estimators was to avoid passing problem_configuration to get_estimators just for this check
It's a limitation with MS unfortunately, due to the unreliability of forecasting over intervals that randomly change in values (different days in a month). This doesn't seem to be an issue with month end

[freddyaboulton]

Yea let's not change the api of get_estimators yet! Ok, I think there will be other frequencies with that limitation so I will file an issue for following up about that. Might make sense to add yet another data check for this hehe