Install AutoTs on Azure ML kernel and Use it in the Note Book for Time Series

[RCML2016]

Hello All, I am trying to Use Autots for forecasting in the Azure ML Note Book.

Used below :
!pip3 install auto-ts

The Output is : /bin/bash: pip3: command not found.

Any Suggestions ?

[winedarksea]

This is nothing to do with AutoTS but rather with Python. You can do some more research, there are plenty of answers on Google.
For Linux, for example, you might run: sudo apt install python3-pip

[winedarksea]

also this is pip install autots, auto-ts with the hypen is actually a different package

[RCML2016]

also this is pip install autots, auto-ts with the hypen is actually a different package

The !pip install autots worked on Azure ML Note Book.

[RCML2016]

Hi, How to use the Holidays Package ? Is there any Pre-requisites and Pre installs for Holidays Package. I am using AutoTs and Its Even Predicting for Weekends and Holidays. I would like the Models to be Not to future forecasting for Weekends/Holidays.

[winedarksea]

!pip install holidays from here
that's for the holiday_country usage. The HolidayDetector/HolidayTransformer for automatic holiday detection do not require additional packages.

[winedarksea]

I don't entirely understand your question. Did you want to not predict holidays?

[RCML2016]

I don't entirely understand your question. Did you want to not predict holidays?

Yes , I don't want to NO predict for Holidays. Example Jan-01-2023 or future May-29-2023 are Holidays and some Stores will be Not Opened. So NO Prediction will help in consolidating the Over All Sales at Daily level.

[winedarksea]

constraint is probably your best bet for that.
For the example dataset of df = load_daily(long=False)

forecast_index = pd.date_range(start=df.index[-1], periods=forecast_length + 1, freq=df.index.freq)[1:]
# sets an extremely high value for the cap, one that should never actually be reached by the data normally
upper_constraint = pd.DataFrame(9999999999, index=forecast_index, columns=df.columns)
# in this case also assuming negatives won't happen so setting a lower constraint of 0
lower_constraint = pd.DataFrame(0, index=forecast_index, columns=df.columns)
# add in your dates you want as definitely 0
upper_constraint.loc["2022-10-31"] = 0  # can do this to specific columns if some are opened, some closed
constraint = {
    "constraint_method": "absolute",
    "upper_constraint": upper_constraint,
    "lower_constraint": lower_constraint,
    "bounds": True,
}
model = AutoTS(
    _your args here_
)
prediction = model.predict(...)
prediction.apply_constraints(**constraint)

constraints can be passed to AutoTS as part of training, however the need to fit a 0 often skews the error metrics and results in a poor model choice. Better to do it after in the case of extreme constraints.

another solution is to pass in an external (future_regressor) which contains a flag (0 or 1 feature) for historical holidays (in training) and for future holidays for the predict.

[RCML2016]

Hello Experts, When i use the ****model_list='probabilistic' ** and Sort the fit with sMAPE, MAE, RMSE..etc ,

I am seeing something Model Name as "Ensemble" . Which is some thing different other than FBProphet or ETS.. What is Ensemble and Why i am seeing it in Models List ??

Model	Runs	smape	mae	rmse	Score	TotalRuntimeSeconds

650 FBProphet 1 12.533976 288191.856204 497702.808894 24.256076 0.3352
777 Ensemble 1 12.567421 288779.309279 499049.020462 24.079424 2.0422
648 FBProphet 1 12.576267 289133.565814 499448.183847 24.175259 0.3896
775 Ensemble 1 12.607595 289769.620251 500594.426445 24.212839 3.2829
647 Ensemble 1 12.610875 289852.744161 500679.228767 24.283677 1.0805
774 Ensemble 1 12.642774 290549.878459 501678.295230 24.229837 2.5530
507 FBProphet 1 12.678483 292753.181623 506508.080901 24.246685 0.3811
406 FBProphet 1 12.762771 294555.926160 508768.922959 24.347063 0.3465
769 Ensemble 1 12.767177 294642.949796 508857.163158 24.104104 2.0248
646 Ensemble 1 12.773827 293428.072058 505590.046716 23.814324

[winedarksea]

An Ensemble is a combination of other models. It will only use the models you have listed, so in this case it will be some combination of Prophet and ETS models (you can see which in the ModelParameters).
Ensembling usually improves accuracy, but is easy enough to turn off. Set ensemble=None in the AutoTS() args.

[RCML2016]

An Ensemble is a combination of other models. It will only use the models you have listed, so in this case it will be some combination of Prophet and ETS models (you can see which in the ModelParameters). Ensembling usually improves accuracy, but is easy enough to turn off. Set ensemble=None in the AutoTS() args.

@winedarksea , Thank You.

[RCML2016]

Getting Error :
TypeError: infer_freq() got an unexpected keyword argument 'warn' on the Fit.

Below are my args for the Fit.

Model.fit(DF, date_col='Date', value_col='ValueCol', id_col=None)

Anyone Experiencing , I am using this on the Azure Note books. Thnaks.

[winedarksea]

This is from pandas infer_freq function. It looks like the "warn" argument has been removed in the latest release of Pandas. You must have a new environment using Pandas 2.0.0, downgrading/installing to pandas version 1.x should work