Skip to main content

This library uses nbeats-pytorch as base and accomplishes univariate time series forecasting using N-BEATS.

Project description


N-BEATS: Neural basis expansion analysis for interpretable time series forecasting

NBEATS is a pytorch based library for deep learning based time series forecasting ( and utilises nbeats-pytorch.

Dependencies: Python >=3.6


$ pip install NBEATS


from NBEATS import NeuralBeats

Mandatory Parameters:

  • data
  • forecast_length

Basic model with only mandatory parameters can be used to get forecasted values as shown below:

import pandas as pd
from NBEATS import NeuralBeats

data = pd.read_csv('test.csv')   
data = data.values        # (nx1 array)

model = NeuralBeats(data=data, forecast_length=5)
forecast = model.predict()

Optional parameters to the model object

Parameter Default Value
backcast_length 3* forecast_length
path ' ' (path to save intermediate training checkpoint)
checkpoint_name ''
mode 'cpu'
batch_size len(data)/10
thetas_dims [4, 8]
nb_blocks_per_stack 3
share_weights_in_stack False
train_percent 0.8
save_model False
hidden_layer_units 128
stack [1,1] (As per the paper- Mapping is as follows -- 1: GENERIC_BLOCK, 2: TREND_BLOCK , 3: SEASONALITY_BLOCK)



This is used for training the model. The default value of parameters passed are epoch=25, optimiser=Adam, plot=False, verbose=True

ex:,optimiser=torch.optim.AdamW(model.parameters, lr=0.001, betas=(0.9, 0.999), eps=1e-07, weight_decay=0.01, amsgrad=False),plot=True, verbose=True)
predict_data ()

The argument to the method could be empty or a numpy array of length backcast_length x 1 which means if no argument is passed and training data is till Dec 2019 the prediction will begin from Jan 2020 but if forcasting for 3 months ahead(forecast_length=3)from March 2020 then numpy array of backcast_length (3 x forecast_length -This is by default) i.e 9(3 x 3) previous months (June 2019 to Feb 2020) needs to be provided to predict for March,Apr,May 2020.

Important Note : Backcast length can be provided as a model argument along with forecast_length eg backcast_length=6,backcast_length=9,backcast_length=12......till backcast_length=21 for forecast_length=3 ,as the paper suggests values between 2 x forecast_length to 7 x forecast_length .The default is 3 x forecast_length .

Returns forecasted values.

save(file) & load(file,optimizer):

Save and load the model after training respectively.

Example:'') or model.load('')


1: GENERIC_BLOCK and 3: SEASONALITY_BLOCK stacks are used below (stack=[1,3]).Go through th paper for more details.Playing around with the 3 blocks(GENERIC,SEASONALITY and TREND) might improve accuracy.

import pandas as pd
from NBEATS import NeuralBeats
from torch import optim

data = pd.read_csv('test.csv')   
data = data.values # nx1(numpy array)


#or use prebuilt models

#use customised optimiser with parameters,optimiser=optim.AdamW(model.parameters, lr=0.001, betas=(0.9, 0.999), eps=1e-07, weight_decay=0.01, amsgrad=False)) 

#model.predict(predict_data=pred_data) where pred_data is numpy array of size backcast_length*1

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

NBEATS-1.3.8.tar.gz (5.4 kB view hashes)

Uploaded Source

Built Distribution

NBEATS-1.3.8-py3-none-any.whl (6.4 kB view hashes)

Uploaded Python 3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page