Time Series Analysis and Forcasting with Facebook Kats

Example code files and referenced papers presented on YAZSUM 2021 AI Event

Kats Info

Kats is a a lightweight, easy-to-use, and generalizable time series analysis framework from Facebook. It performs various time series analysis taks, including detection, forecasting, feature extraction.

Kats is available for download on PyPI with just "pip install kats".

Important links

• Kats Homepage: https://facebookresearch.github.io/Kats/
• Kats Python package: https://pypi.org/project/kats/0.1.0/
• Source code repository: https://github.com/facebookresearch/kats
• Official Tutorials: https://github.com/facebookresearch/Kats/tree/master/tutorials
• Video Tutorial: https://youtu.be/GzP2xdoVqJE (in Turkish with English subtitles)

Contents:

Forecasting

• sarima.py Basic Forecasting example with SARIMA Model
• prophet.py Basic Forecasting example with FB Prophet Model
• theta.py Basic Forecasting example with Theta Model
• holtwinters.py Basic Forecasting example with Holt-Winters Model
• ensemble.py Basic Forecasting example with an Ensemble Model of ARIMA, Prophet, Theta, and Linear Models

Hyperparameter Tuning and Backtesting

• arima.py Hyper parameter tuning an ARIMA model with Grid Search method
• backtest.py Backtesting an ARIMA and a Prophet Model using metrics mape, smape, mae, mase, mse, rmse

Changepoint, Outlier, and Trend Detection

• cusum.py Changepoint detection with CUSUM method
• bocpd.py Changepoint detection with Bayesian Online Change Point Detection (BOCPD) method
• outlier.py Outlier detection and removal with and without interpolation
• trend.py Trend detection

Time Series Features

• features.py Extracting statistical time series features

Papers

• Bayesian Online Changepoint Detection
• Self-supervised learning for fast and scalable time series hyper-parameter tuning