Please help with multivariate time series analysis in gluonts

[pratikgehlott]

I need urgent help in this part, please can anyone help?

[mbohlkeschneider]

Hi @pratik325,

can you be more precise? What exactly is the issue?

Best regards.

[mbohlkeschneider]

Hi @pratik325,

Here is a snipped on how to setup a dummy dataset with dynamic features. Basically, you have to convert your data into this format. Let me know if there are more questions. For more information, please refer to the extended tutorial: https://ts.gluon.ai/tutorials/forecasting/extended_tutorial.html

import pprint

from gluonts.dataset.repository.datasets import get_dataset, dataset_recipes
from gluonts.evaluation import Evaluator
from gluonts.evaluation.backtest import make_evaluation_predictions
from gluonts.model.deepar import DeepAREstimator
from gluonts.mx.trainer import Trainer
from gluonts.dataset.common import ListDataset
import numpy as np

data_entry_train = {"target": np.linspace(0,1,100), "feat_dynamic_real": [np.linspace(0,1,100)], "start":"2020-01-01"}

dataset_train = ListDataset([data_entry_train]*100, freq='H')

data_entry_test = {"target": np.linspace(0,1,124), "feat_dynamic_real": [np.linspace(0,1,124)], "start":"2020-01-01"}

dataset_test = ListDataset([data_entry_test]*100, freq='H')


estimator = DeepAREstimator(
    prediction_length=24,
    freq="H",
    trainer=Trainer(epochs=5, num_batches_per_epoch=10),
    use_feat_dynamic_real=True
)

predictor = estimator.train(dataset_train)

forecast_it, ts_it = make_evaluation_predictions(
    dataset_test, predictor=predictor, num_samples=100
)

agg_metrics, item_metrics = Evaluator()(
    ts_it, forecast_it, num_series=len(dataset_test)
)

pprint.pprint(agg_metrics)

[MAYUKH97]

Hi @mbohlkeschneider I have a follow up question So, I have a data in the form of hierarchy and want to forecast for the granular level. For example data = ["food" ,"food_category", "revenue"], So in food we have different types of items and we segregate on basis of chinese, thai, india and then in chinese we may have like noodles, pasta, etc. so as every segment is connected so how do we forecast revenue the granular level for each individual items.

[mbohlkeschneider]

Hi @MAYUKH97,

you can encode them as static cat features using standard label encoding. We currently don't have a "real" hierarchical forecasting algorithm in GluonTS currently.

[pratikgehlott]

hey @mbohlkeschneider can we handle multiple row having same dates in gluonts?

[mbohlkeschneider]

Hi @pratik325,

as mentioned above, this depends on your forecasting scenario and how your input data is stored. Note that GluonTS does not provide any tooling to convert data into GluonTS format - this has to be done by the user. If you have specific questions about the GluonTS data format, I'll be happy to answer them.