NBeats_Time_Series

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

N-BEATS is a type of neural network that was first described in a paper published in the 2019 ICLR conference by Oreshkin et al. The authors reported that N-BEATS outperformed the M4 forecast competition winner by 3%. The M4 winner was a hybrid between a recurrent neural network and Holt-Winters exponential smoothing — whereas N-BEATS implements a “pure” deep neural architecture.

In this repo, we will use nbeats_forecast which is an end to end library for univariate time series forecasting using N-BEATS. This library uses nbeats-pytorch as base and simplifies the task of forecasting using N-BEATS by providing a interface similar to scikit-learn and keras.

Requires: Python >=3.6

Installation

$ pip install nbeats_forecast

Import

from nbeats_forecast import NBeats

Input

numpy array of size nx1

Output

Forecasted values as numpy array of size mx1

Mandatory Parameters for the model:

• data
• period_to_forecast

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

import pandas as pd
from nbeats_forecast import NBeats

data = pd.read_csv('data.csv')   
data = data.values        #univariate time series data of shape nx1 (numpy array)

model = NBeats(data=data, period_to_forecast=12)
model.fit()
forecast = model.predict()

The other optional parameters for the object of the NBeats model (as described in the paper) can be tweaked for better performance. If these parameters are not passed, default values as mentioned in the table below are considered.

Parameter	Type	Default Value	Description
backcast_length	integer	3* period_to_forecast	Explained in the paper
path	string	' '	path to save intermediate training checkpoint
checkpoint_file_name	string	'nbeats-training-checkpoint.th'	name for checkpoint file ending in format .th
mode	string	'cpu'	Any of the torch.device modes
batch_size	integer	len(data)/15	size of batch
thetas_dims	list of integers	[7, 8]	Explained in the paper
nb_blocks_per_stack	integer	3	Explained in the paper
share_weights_in_stack	boolean	False	Explained in the paper
train_percent	float(below 1)	0.8	Percentage of data to be used for training
save_checkpoint	boolean	False	save intermediate checkpoint files
hidden_layer_units	integer	128	hissen layer units
stack	list of integers	[1,1]	adding stacks in the model as per the paper passed in list as integer. Mapping is as follows -- 1: GENERIC_BLOCK, 2: TREND_BLOCK , 3: SEASONALITY_BLOCK

Repository Structure

Model

PyTorch implementation of N-BEATS can be found in models/nbeats.py

Implementation

The notebooks directory contains a notebook with univariate time series analysis using the one of the time series in the provided dataset. The notebook consists of steps such as loading time series data, time series visualization, implementing the NBeats model and training it with 75% of the data. It also includes test on the remaining 25% data and evaluation metrics on the prediction along with a forecast plots as well.

Iteration

The iteration folder consists of a simple code that iterates over all the 50 datasets in the directory and produces NBeats model for each indivdual time series along with its prediction plot.

How to use the repo

In order to use this basic repo for implementing NBeats model you can follow these basic steps:

1. Fork or download this repository
2. Install nbeats_forecast library on your system.
3. Open the notebooks/nbeats.ipynb file and modify the file path as per your dataset.
4. Run the notebook to get the NBeats model results for your dataset.
5. Open the iteration/nbeats_iter.py file and modify the dataset path according to your directory.
6. Run the iteration file to create NBeats models for all the time series in your directory.

What works and what doesn't

The given Machine Learning Challenge was truly inspiring and challenged my concepts and knowledge in the field of Machine Learning. Out of the given tasks, I was able to implement the NBeats model using Pytorch implementation and apply it for the enitre 50 time series data files.

The second task regarding REvin normalization was a little beyond my current knowledge base and although I was able to understand how Revin works by normalizing and then denormalizing the time series data to improve performance, I was not able to implement the same in my NBeats pipeline.

All the tasks were performed to the best of my abilities and potential and I hope to higly scale my knowledge and capabilities in the field of Machine Learning as a part of Greendeck ML team.