Orbit

​

Object Oriented Bayesian Time Series Forecasting

​

Orbit Development Team

Jul 2021 Uber Meetup

github: https://github.com/uber/orbit

doc: https://orbit-ml.readthedocs.io/en/stable/

blog post: https://eng.uber.com/orbit/

previous ISF talk: https://youtu.be/LXDpq_iwcWY

tutorial notebook: link

Agenda

Motivation

Introduction

Package Design

Performance

Hands-on

Motivation

Bayesian methods provide an intuitive way to connect prior belief (information) with new data

The Bayesian approach in general and for times series in particular can

• Parsimoniously resolve complex structures
• Account for multiple sources of variation
• Combine multiple types of information
• Be easily interpreted
• Recently published “Bayesian Time Varying Coefficient Model with Applications to Marketing Mix Modeling” https://arxiv.org/abs/2106.03322

Motivation

Orbit aims to simplify Bayesian time-series workflow with powerful APIs

Your Business

Problem

​

​

Stan

Pyro

Motivation

An overview of Bayesian time-series workflow

Motivation

Compare to other solutions in Bayesian time-series

ABILITY TO GENERALIZE

MODEL IMPLEMENTATION

prophet

bsts

orbit

pymc3

pyro

tensorflow

probability

Introduction

Why Orbit?

• Ease of use
• Flexibility
• Interpretability

• Object-Oriented Design
• Workflow support
• High-performance

Introduction

Time series modeling with ease

Intuitive initialize-fit-predict interface while Orbit doing the heavy lifting under the hood.

Initialize

Fit and Predict

Introduction

Allows user choice of

• Prior distributions (e.g., uniform between bounds, or informative)
• Algorithm for sampling the Posterior distribution
• Some models support multiple APIs e.g. Stan vs. Pyro
• Returning a point estimate (with choice of aggregation; e.g., mean or map) or the whole posterior distribution

Allow prediction decomposition

Introduction

Orbit predictions can be decomposed in a structural way

Package Design

Use composition and inheritance to build customized model

Package Design

Follow initialize-fit-predict convention along with the object-oriented design

from estimation template

​

from base model

Package Design

Diagnose the fitted models with built-in backtest modules

Performance

Benchmark against other models with built-in backtest modules

Performance

More robust check behind the scene

Meet Michelangelo: Uber’s Machine Learning Platform https://eng.uber.com/michelangelo-machine-learning-platform

A Peek into the Math for Damped Local Trend (DLT)

Hands-on

Forecasting Equations

Update Equations

Hands-on

A couple lines from data to forecast and diagnostics

Initialize

Fit and Predict

Plot

Hands-on

Consistent interface for all models

Exponential Smoothing (ETS)

Local Global Trend (LGT)

Damped Local Trend (DLT)

Kernel Time-based Regression (KTR-Lite)

Hands-on

Support for different types of regression penalties

Regular Regression

Fixed/Auto Ridge

Lasso

Hands-on

Posterior distributions and model convergence check

Integrated with arviz

Available on arxiv and github

Code on Github

https://github.com/uber/orbit

Paper on arxiv

https://arxiv.org/abs/2004.08492

​

Paper with the code

https://stat.paperswithcode.com

Post on Uber Engineering Blog

https://eng.uber.com/orbit/

The Team and Alumni

Team at Uber

​

Alumni

​

Roadmap

• Explore more sampling API (e.g. PyStan3, Numpyro)
• Allow configuring distribution of some parameters (e.g. noise, regression coefficients)
• Allow configuring link function (identity, logit, poisson …)
• Explore multivariate time-series

Q&A

Thanks for joining!

Take a minute and fill out the survey for feedbacks!

Survey Link & QR Code: https://forms.gle/ZXqSqXahRS7G3W157

​