nixtlacommunity #general

Channels

random

linen

statsforecast

Valeriy

08/15/2022, 12:42 PM

The link doesn't seem to work?

Max (Nixtla)

08/15/2022, 1:45 PM

Hi Valery! We have been updating a lot of things in preparation for the next release. Some of the things we changed in the new version is a sklearn like syntax. Here you can find the getting started: https://nixtla.github.io/statsforecast/examples/Getting_Started_with_Auto_Arima_and_ETS.html

Max (Nixtla)

08/15/2022, 1:45 PM

If you want to use the older version, we can also share the old getting started.

👍 1

Nelson Elías Troncoso Martínez

08/16/2022, 4:31 PM

Hi everyone! Is there any way to use AutoARIMAProphet inside the class StatsForecast?

Max (Nixtla)

08/16/2022, 5:28 PM

Hi! Currently, that feature is not supported. The reason is that StatsForecast uses numpy arrays - that are far more efficient than data frames. Prophet, on the other side, needs data frames.

😢 2

👍 1

Max (Nixtla)

08/17/2022, 2:48 PM

Our ETS is 30% faster and 100 times faster than NeuralProphet: https://www.linkedin.com/posts/mergenthaler_timeseries-forecasting-statistics-activit[…]yNtW?utm_source=linkedin_share&utm_medium=member_desktop_web

👏 2

👍 2

🙌 4

🔥 3

Kin Gtz. Olivares

08/17/2022, 7:14 PM

Hi channel has anyone had experience with Neural Prophet? Would you be able to comment on it in this reddit post? https://www.reddit.com/r/MachineLearning/comments/wqrw8x/d_fool_me_once_shame_on_you_fool_me_twice_shame/

Valeriy

08/17/2022, 9:24 PM

How it started

🙌 1

Valeriy

08/17/2022, 9:24 PM

How it ended. Lol

🤣 2

Kin Gtz. Olivares

08/17/2022, 9:25 PM

it is fun hahaha

Max (Nixtla)

08/17/2022, 11:04 PM

We made it to the front page of Hacker News. Show some love by upvoting our post: “Show HN: Exp. Smoothing is 32% more accurate and 100x faster than Neural-Prophet”. https://news.ycombinator.com/

👍 1

🚀 1

luca

08/18/2022, 9:51 AM

Hi, as a confirmation: you develop statsforecast via jupyter notebooks that are converted to python scripts via Quarto?

Jason

08/18/2022, 2:14 PM

Hello! Hope all is well.

Jason

08/18/2022, 2:15 PM

Checking out https://github.com/Nixtla/statsforecast. I noticed a couple of the colab links don’t work. Like this one: “Complete pipeline and comparison: 20x faster than pmdarima and 500x faster than Prophet.”

Ginger Holt

08/18/2022, 6:06 PM

Has anyone gotten this error: "ImportError: cannot import name 'naive' from 'statsforecast.models' (/local_disk0/.ephemeral_nfs/envs/pythonEnv-230a87f5-7353-4abc-a244-a806abd865a8/lib/python3.8/site-packages/statsforecast/models.py)"

👀 3

AlexH

08/19/2022, 9:45 AM

Hi All, currently I am working on project where I am forecasting demand for the next two weeks. We do have ~1000 Outlets with ~1000 products that are separted in around 50 groups. Three years of data are around 200M rows. Currently, we are using two different strategies (1) gradient boosting with all data and (2) gradient boosting per group. Both approaches work reasonably well. However, now we would also like to explore if we can gain anything by using deep learning models and thats where I found your neuralforecast library. Would you have any pointers in which directions we should explore? One challenge for us is, that not all articles are offered all the time (some only for like 3-4 weeks per year. However, there might be similar articles that were offered in the other times) and also not in all outlets. Therefore we have many gaps in the data and consequently different lenghts of the timeseries. Additionally, we do have much additional data about promotions, weather, prices etc. Thanks a lot for any input that you might be able to share!

👀 1

Taha

08/23/2022, 11:28 AM

Hello everyone, If I were to use

tspreprocess

and

tsfeatures

through the CLI, does that mean I can only predict using

tsforecast

? Or is it possible to use those 2 preprocessing steps and then use your

statsforecast

hierarchicalforecast

library? Thanks.

Tom Merritt Smith

08/31/2022, 11:35 AM

Hi! Enjoying using

statsforecast

, but I wonder if there's plans for more verbose logging from the models? I'm often getting a 'could not fit model' error, but unclear as to why that is

👀 1

Taha

09/03/2022, 5:33 AM

Hello, Im trying to run the example on the

mlforecast

page but when I ran the line:

Copy code

from mlforecast import Forecast

I got the error:

Copy code

ImportError: cannot import name 'Forecast' from 'mlforecast'

I was able to import it using:

Copy code

from mlforecast.forecast import Forecast

but then the

Forecast

instantiation takes in params:

model

and

ts

(I realized this is an older version of mlforecast, namely, 0.1.0) When I run

!pip show mlforecast

, I see I have the latest version:

0.2.0

but when I run

mlforecast.__version__

I get

0.1.0

Am I doing something wrong? Thanks.

Phi Nguyen

09/08/2022, 4:02 PM

I am getting an error when trying to run statsforecast (ETS) on the M5 data.

Phi Nguyen

09/08/2022, 4:02 PM

Copy code

ValueError                                Traceback (most recent call last)
/home/ray/workspace-project-nixtla-m5/nixtla-m5-v2.ipynb Cell 10' in <cell line: 2>()
      1 init = time()
----> 2 Y_hat = fcst.forecast(28)
      3 end = time()
      4 print(f'Minutes taken by StatsForecast using: {(end - init) / 60}')

File ~/anaconda3/lib/python3.8/site-packages/statsforecast/core.py:405, in StatsForecast.forecast(self, h, df, X_df, level, fitted, sort_df)
    403 X, level = self._parse_X_level(h=h, X=X_df, level=level)
    404 if self.n_jobs == 1:
--> 405     res_fcsts = self.ga.forecast(models=self.models, h=h, fitted=fitted, X=X, level=level)
    406 else:
    407     res_fcsts = self._forecast_parallel(h=h, fitted=fitted, X=X, level=level)

File ~/anaconda3/lib/python3.8/site-packages/statsforecast/core.py:148, in GroupedArray.forecast(self, models, h, fitted, X, level)
    146 if has_level:
    147     kwargs['level'] = level
--> 148 res_i = model.forecast(h=h, y=y_train, X=X_train, X_future=X_f, fitted=fitted, **kwargs)
    149 cols_m = [key for key in res_i.keys() if any(key.startswith(m) for m in matches)]
    150 fcsts_i = np.vstack([res_i[key] for key in cols_m]).T

File ~/anaconda3/lib/python3.8/site-packages/statsforecast/models.py:245, in ETS.forecast(self, y, h, X, X_future, fitted)
    237 def forecast(
    238         self, 
...
--> 950 data_positive = min(y) > 0
    951 if (not data_positive) and errortype == 'M':
    952     raise ValueError('Inappropriate model for data with negative or zero values')

ValueError: min() arg is an empty sequence

Phi Nguyen

09/08/2022, 4:03 PM

I have filtered down my dataset to just one unique ID

Phi Nguyen

09/08/2022, 4:03 PM

Copy code

Y_df.describe()

Phi Nguyen

09/08/2022, 4:03 PM

Copy code

y
count	1941.000000
mean	10.786193
std	1.259658
min	10.000000
25%	10.000000
50%	10.000000
75%	11.000000
max	25.000000

Phi Nguyen

09/08/2022, 4:03 PM

and running this code

Phi Nguyen

09/08/2022, 4:03 PM

Copy code

fcst = StatsForecast(
    df=Y_df, 
    models=[ETS(season_length=7, model='ZNA')], 
    freq='D'
)

Phi Nguyen

09/08/2022, 4:04 PM

Y_hat = fcst.forecast(28)

Phi Nguyen

09/08/2022, 4:04 PM

I have drop the na rows in my dataset but it doesn’t look I have zeros of negative value

Max (Nixtla)

09/08/2022, 4:28 PM

Hi @Phi Nguyen, could you share a head of your DF?

Max (Nixtla)

09/08/2022, 4:31 PM

I have filtered down my dataset to just one unique ID

Probably you droped the id column and that is causing the error. Just add a

unique_id

column with any value (e.g. “series_1”) so you have a DF with the following columns:`unique_id:string` ,

ds:datestamp

y:float