Hi, I've got some good results with stataforecast ...
# generalm
Miro Lavi
10/20/2024, 5:52 AMHi, I've got some good results with stataforecast AutoARIMA together with exogenous features. I'm trying to reduce the number of dummy predictors as I have (approx 50) and I've created a custom sklearn transformer for that. It works as intended when I run on MLForecast sklearn's LinearRegression but when I try to replace it with statsforecast AutoARIMA or ARIMA I get "AttributeError: 'DataFrame' object has no attribute 'dtype'". It looks to be running on numpy, and when I edit my transformer to return numpy I get "ValueError: auto_arima can only handle univariate time series".
Is there a way to use custom transformers in statsforecast side too or is there a way to use AutoARIMA in the MLForecast pipeline context?
Miro Lavi
10/20/2024, 6:15 AMlinear_regression = LinearRegression()
arima = AutoARIMA()
pipeline = Pipeline(
[
("PromoAverageTransformer", promotion_avg_transformer),
# ("LinearRegression", linear_regression),
("ARIMAX", arima),
]
)
ml = MLForecast(
models={"promo_lr": pipeline},
freq="1d",
)
cv_results = ml.cross_validation(
df=features,
h=21,
step_size=21,
n_windows=18,
time_col="Date",
id_col="Market",
target_col="Visitors",
static_features=[],
)
o
Olivier
10/21/2024, 11:23 AMWhat happens when you use statsforecast's cross-validation?
m
Miro Lavi
10/21/2024, 5:05 PMGood idea, didnt even think about it, but what I get is "ValueError: could not convert string to float: 'Season Sale...'". I think this means that the sklearn transformer is not working in the StatsForecast context.
I only changed the MLForecast to StatsForecast, changed model from LR to AutoARIMA In the pipeline and then commented out the static_features param which is required in the MLForecast cv.. other than that it should be the same.
Miro Lavi
10/21/2024, 5:13 PMIf just trying to pass AutoARIMA in the sklearn pipeline without any transformers it also doesnt work and gets AttibuteError: 'str' object has no attribute 'prediction_intervals'.
o
Olivier
10/21/2024, 5:40 PMDoes it work without the pipeline, i.e. just perform cross-validation with AutoArima using statsforecast?
m
Miro Lavi
10/21/2024, 7:44 PMYeh works as normal without the pipeline
o
Olivier
10/22/2024, 2:15 PMOk - if the issue is purely in the pipeline you could consider adding the features first, and then running the cv. Is that a possibility?
m
Miro Lavi
10/22/2024, 8:44 PMIn my case the transformer is adapting based on the new data, so cant really create features prior using the cross validation, but I'll try if I can reach similar results with LinearRegression + LightGBM combo.
👍 1
3 Views