Sep 23–26, 2019

Lightning-fast time series modeling and prediction: (S)ARIMA on steroids

Meir TOLEDANO (Anodot)
11:20am12:00pm Wednesday, September 25, 2019
Location: 1A 06/07
Average rating: ***..
(3.00, 3 ratings)

Who is this presentation for?

  • Chief data scientists, and algorithms experts

Level

Intermediate

Description

For decades, researchers and practitioners have used different statistical techniques for modeling time-indexed data (i.e., time series). One model among many is still under the spotlight: ARIMA.

The ARIMA family model is the best time series model in some circumstances. In practice, it performs very well at predicting the future of a time series and even against highly sophisticated super-new models like long short-term memory (LSTM), according to benchmarks. Moreover, a complete theoretical analysis was conducted in the last decades. So it’s a rare machine learning technique where a mathematical analysis is tractable (but not simple, unfortunately). For all those reasons, ARIMA is a central element in the time series analysis field. With the growing user base, a subproblem arose that remained unsolved for decades: the efficient estimation of seasonal ARIMA models a.k.a. (S)ARIMA.

Meir Toledano explains how Anodot solved this problem, because the seasonal problem is extremely important in practice. Most of the gathered data that monitors human activity is seasonal. For example, the number of taxi rides sampled hourly exhibits a weekly pattern: more rides in the day than in the night; more rides during the week than during the weekend. What’s true for taxi rides is true for electricity consumption, hotel reservations, car traffic, ad spending, ecommerce revenue, etc.

Using (S)ARIMA has been challenging. Standard implementations take around 15 minutes to learn the model, mainly because the algorithm has to “see” very far in past in order to “understand” what’s happening. This limitation makes (S)ARIMA impractical to use in the big data era. Anodot estimates it can now use (S)ARIMA for more than 100 million metrics every day. On average it takes 80 ms to estimate the model, a 5 orders of magnitude improvement! This was made possible from breakthroughs in theory and code implementation.

Meir demonstrates how to use the model to make predictions and for real-time anomaly detection, where Anodot is a specialist.

Prerequisite knowledge

  • A basic understanding of on time series modeling

What you'll learn

  • Understand why seasonal time series modeling is challenging, how Anodot solved the problem at the scale of millions of time series, and how to use the new method for time series prediction and real-time anomaly detection
Photo of Meir TOLEDANO

Meir TOLEDANO

Anodot

Meir Toledano is a data scientist and algorithm engineer at Anodot. He’s an engineer and entrepreneur, having studied and started his career in Paris, France. Previously, he was an aeronautic engineer, developed trading algorithms and risk models in the financial industries, and worked to the internet and high-tech industry.

Comments on this page are now closed.

Comments

Picture of Meir TOLEDANO
Meir TOLEDANO | Data scientist and algorithm engineer
10/05/2019 6:28am EDT

Hi,

I uploaded the slides to slideshare :

http://bit.ly/359s0YG

Best,

Meir

Picture of Margarida Campos
Margarida Campos | Associate, Data Scientist
10/03/2019 9:33am EDT

Hello,
I really liked your presentation.
Can you please provide the slides?
Thank you so much.
Margarida Campos

    Contact us

    confreg@oreilly.com

    For conference registration information and customer service

    partners@oreilly.com

    For more information on community discounts and trade opportunities with O’Reilly conferences

    strataconf@oreilly.com

    For information on exhibiting or sponsoring a conference

    pr@oreilly.com

    For media/analyst press inquires