Nixtla

Nixtla is an open-source time series forecasting library.

We are helping data scientists and developers to have access to open source state-of-the-art forecasting pipelines. For that purpose, we built a complete pipeline that can be deployed in the cloud using AWS and consumed via APIs or consumed as a service. If you want to set up your own infrastructure, follow the instructions in the repository (Azure coming soon).

You can use our fully hosted version as a service through our python SDK (autotimeseries). To consume the APIs on our own infrastructure just request tokens by sending an email to [email protected] or opening a GitHub issue. We currently have free resources available for anyone interested.

We built a fully open-source time-series pipeline capable of achieving 1% of the performance in the M5 competition. Our open-source solution has a 25% better accuracy than Amazon Forecast and is 20% more accurate than fbprophet. It also performs 4x faster than Amazon Forecast and is less expensive.

To reproduce the results: or you can read this Medium Post.

At Nixtla we strongly believe in open-source, so we have released all the necessary code to set up your own time-series processing service in the cloud (using AWS, Azure is WIP). This repository uses continuous integration and deployment to deploy the APIs on our infrastructure.

Python SDK Basic Usage

Install

PyPI

pip install autotimeseries

How to use

Check the following examples for a full pipeline:

• M5 state-of-the-art reproduction.
• M5 state-of-the-art reproduction in Colab

Basic usage

import os

from autotimeseries.core import AutoTS

autotimeseries = AutoTS(bucket_name=os.environ['BUCKET_NAME'],
                        api_id=os.environ['API_ID'],
                        api_key=os.environ['API_KEY'],
                        aws_access_key_id=os.environ['AWS_ACCESS_KEY_ID'],
                        aws_secret_access_key=os.environ['AWS_SECRET_ACCESS_KEY'])

Upload dataset to S3

train_dir = '../data/m5/parquet/train'
# File with target variables
filename_target = autotimeseries.upload_to_s3(f'{train_dir}/target.parquet')
# File with static variables
filename_static = autotimeseries.upload_to_s3(f'{train_dir}/static.parquet')
# File with temporal variables
filename_temporal = autotimeseries.upload_to_s3(f'{train_dir}/temporal.parquet')

Each time series of the uploaded datasets is defined by the column item_id. Meanwhile the time column is defined by timestamp and the target column by demand. We need to pass this arguments to each call.

columns = dict(unique_id_column='item_id',
               ds_column='timestamp',
               y_column='demand')

Send the job to make forecasts

response_forecast = autotimeseries.tsforecast(filename_target=filename_target,
                                              freq='D',
                                              horizon=28,
                                              filename_static=filename_static,
                                              filename_temporal=filename_temporal,
                                              objective='tweedie',
                                              metric='rmse',
                                              n_estimators=170,
                                              **columns)

Download forecasts

autotimeseries.download_from_s3(filename='forecasts_2021-10-12_19-04-32.csv', filename_output='../data/forecasts.csv')

Forecasting Pipeline as a Service

Our forecasting pipeline is modular and built upon simple APIs:

tspreprocess

Time series usually contain missing values. This is the case for sales data where only the events that happened are recorded. In these cases it is convenient to balance the panel, i.e., to include the missing values to correctly determine the value of future sales.

The tspreprocess API allows you to do this quickly and easily. In addition, it allows one-hot encoding of static variables (specific to each time series, such as the product family in case of sales) automatically.

tsfeatures

It is usually good practice to create features of the target variable so that they can be consumed by machine learning models. This API allows users to create features at the time series level (or static features) and also at the temporal level.

The tsfeatures API is based on the tsfeatures library also developed by the Nixtla team (inspired by the R package tsfeatures) and the tsfresh library.

With this API the user can also generate holiday variables. Just enter the country of the special dates or a file with the specific dates and the API will return dummy variables of those dates for each observation in the dataset.

tsforecast

The tsforecast API is responsible for generating the time series forecasts. It receives as input the target data and can also receive static variables and time variables. At the moment, the API uses the mlforecast library developed by the Nixtla team using LightGBM as a model.

In future iterations, the user will be able to choose different Deep Learning models based on the nixtlats library developed by the Nixtla team.

tsbenchmarks

The tsbenchmarks API is designed to easily compare the performance of models based on time series competition datasets. In particular, the API offers the possibility to evaluate forecasts of any frequency of the M4 competition and also of the M5 competition.

These APIs, written in Python and can be consumed through an SDK also written in Python. The following diagram summarizes the structure of our pipeline:

Build your own time-series processing service using AWS

Why ?

We want to contribute to open source and help data scientists and developers to achieve great forecasting results without the need to implement complex pipelines.

How?

If you want to use our hosted version send us an email or open a github issue and ask for API Keys.

If you want to deploy Nixtla on your own AWS Cloud you will need:

• API Gateway (to handle API calls).
• Lambda (or some computational unit).
• SageMaker (or some bigger computational unit).
• ECR (to store Docker images).
• S3 (for inputs and outputs).

You will end with an architecture that looks like the following diagram

Each call to the API executes a particular Lambda function depending on the endpoint. That particular lambda function instantiates a SageMaker job using a predefined type of instance. Finally, SageMaker reads the input data from S3 and writes the processed data to S3, using a predefined Docker image stored in ECR.

Run the API locally

1. Create the environment using make init.
2. Launch the app using make app.

Create AWS resources

Create S3 buckets

For each service:

1. Create an S3 bucket. The code of each lambda function will be uploaded here.

Create ECR repositorires

For each service:

1. Create a private repository for each service.

Lambda Function

For each service:

1. Create a lambda function with Python 3.7 runtime.
2. Modify the runtime setting and enter main.handler in the handler.
3. Go to the configuration:
   • Edit the general configuration and add a timeout of 9:59.
   • Add an existing role capable of reading/writing from/to S3 and running Sagemaker services.
4. Add the following environment variables:
   • PROCESSING_REPOSITORY_URI: ECR URI of the docker image corresponding to the service.
   • ROLE: A role capable of reading/writing from/to S3 and also running Sagemaker services.
   • INSTANCE_COUNT
   • INSTANCE_TYPE

API Gateway

1. Create a public REST API (Regional).
2. For each endpoint in api/main.py… add a resource.
3. For each created method add an ANY method:
   • Select lambda function.
   • Select Use Lambda Proxy Integration.
   • Introduce the name of the lambda function linked to that resource.
   • Once the method is created select Method Request and set API key required to true.
4. Deploy the API.

Usage plan

1. Create a usage plan based on your needs.
2. Add your API stage.

API Keys

1. Generate API keys as needed.

Deployment

GitHub secrets

1. Set the following secrets in your repo:
   • AWS_ACCESS_KEY_ID
   • AWS_SECRET_ACCESS_KEY
   • AWS_DEFAULT_REGION