Wide Format about autots

[winedarksea]

If you want to learn how to shape your data, take a look at the examples with long=False. I'd rename your columns storeA_product1, storeA_product2, storeB_product1 and so on, and the just pass as many columns as you need. Behind the scenes all data is made into the "wide" format before AutoTS runs the models.
For regressors try following the example here:
https://github.com/winedarksea/AutoTS/blob/dev/extended_tutorial.md#adding-regressors-and-other-information
Some models only take one variable but I handle the dimensionality reduction for you, so pass in as many variables as you want for a regressor.
@catchlui

[shuaiwang88]

@winedarksea Do I create every autots for each store_product, by using a loop?
like:
for p in store_prodct:
model = autots(subset, x, x, ,x )

[winedarksea]

You can do it that way, however autots accepts multivariate data so either passing wide style data or using long format data with the series_id distinct for each store_product will work.
For large scale it often makes sense to combine the two approaches. Loop AutoTS through high level groupings like product category or store regions, then in each of those subsets run a single AutoTS run on all store_products

[shuaiwang88]

@winedarksea
I added a column company in the data you mentioned.

For the model.fit, I just passed id_col = 'company':
model = model.fit(
df,
id_col='company',
)
Is this correct?

For the forecasting:
prediction = model.predict(forecast_length = 30, verbose=0)
forecasts_df = prediction.forecast
It only returns company B:

[winedarksea]

So your data format in that example data is wide style data. In this case, each column is a target series. So instead of CSUSHPISA you would have "companyB_product1" and for EMVOVERALLEMV you might have "companyA_product1" or whatever as appropriate
If using long style data your approach would work if you had three target series, and company was the series_id. It looks like you are using a load_monthly example so you can switch to long example data by load_monthly(long=True)

[shuaiwang88]

I see. What I am really want to do is using exogenous variables, with series id: forecast sales for different country, with exogenous variables. Any suggestions?

[winedarksea]

future_regressor is for those features which you know about the future. For series you don't know in the future, pass them in as additional series to forecast and some multivariate models will use them as features. I haven't personally found much actual value in a single categorical feature that is constant so there isn't direct support for that all though you could pass it in as a future regressor. Study the production_example.py in this repo for one general approach