When Holt-Winters is better than Machine Learning

Machine Learning (ML) gets a lot of hype, but its Classical predecessors are still immensely powerful, especially in the time series space. Error, Trend, Seasonality Forecast (ETS), Autoregressive Integrated Moving Average (ARIMA), and Holt-Winters are three Classical methods that are not only incredibly popular but also excellent time series predictors. In fact, these Classical Methods outperform several other ML methods including Long Short Term Memory (LTSM) and Recurrent Neural Networks (RNN) in One-Step Forecasting.

In this talk, I’ll show you how the Holt-Winters forecasting algorithm works. Then we’ll use the HOLT_WINTERS() function with InfluxData to make our own time series forecast.

Transcript

2. Big Data & AI Conference Dallas, Texas June 27 – 29, 2019 www.BigDataAIconference.com
3. Anais Dotis-Georgiou, DevRel, InfuuData When Holt-Winters is Better
4. © 2019 InfuuData. All rights reserved. 4 About Me ◦ Anais Jackie Dotis on LinkedIn ◦ @art.anaisdg on Instagram ◦ Super grateful to be here! Developer Advocate, InfuuData
5. © 2019 InfuuData. All rights reserved. 5 Agenda ◦ Overview of InfuuData ◦ Inspiration for this talk ◦ Single Euponential Smoothing (SES) ◦ Optimization of SES ◦ Optimization of Linear Regression ◦ Holt-Winters Multiplicative ◦ Nelder-Mead ◦ How to make predictions with Infuu Developer Advocate, InfuuData
6. © 2019 InfuuData. All rights reserved. 6 TICK Stack
7. © 2018 InfuuData. All rights reserved. 7 Real-Time Analytics 20 % Infrastructure/ Application/ Business Process Monitoring 45 % IoT Monitori ng 35 % Some of our
8. © 2019 InfuuData. All rights reserved. 8 Single Node Ingest Benchmark r4.4xlarge r4.2xlarge r4.xlarge r4.large Series Cardinality
9. © 2019 InfuuData. All rights reserved. 9 Statistical and Machine Learning forecasting methods: Concerns and ways forward” Spyros Makridakis, Evangelos Spiliotis, Vassilios Assimakopoulos
10. © 2019 InfuuData. All rights reserved. 10 Naive Method where
11. © 2019 InfuuData. All rights reserved. 11 Single Exponential Smoothing
12. © 2019 InfuuData. All rights reserved. 12 © 2019 InfuuData. All rights reserved. 12 Component Form
13. © 2019 InfuuData. All rights reserved. 13 © 2019 InfuuData. All rights reserved. 13 Linear Regression Overview of Optimization for Single Euponential Smoothing
14. © 2019 InfuuData. All rights reserved. 14 © 2019 InfuuData. All rights reserved. 14 Optimization for Linear Regression Define the Error: Residual Sum of Squares
15. © 2019 InfuuData. All rights reserved. 15
16. © 2019 InfuuData. All rights reserved. 16
17. © 2019 InfuuData. All rights reserved. 17
18. © 2019 InfuuData. All rights reserved. 18 Holt-Winters’s Multiplicative SES (For Reference)
19. © 2019 InfuuData. All rights reserved. 19 Nedler-Mead
20. © 2019 InfuuData. All rights reserved. 20 HOLT_WINTERS() and Infuu
21. © 2019 InfuuData. All rights reserved. 21 HOLT_WINTERS() and Infuu
22. © 2019 InfuuData. All rights reserved. 22 HOLT_WINTERS() and Infuu
23. © 2019 InfuuData. All rights reserved. 23 HOLT_WINTERS() and Infuu
24. © 2019 InfuuData. All rights reserved. 24 HOLT_WINTERS() and Infuu
25. © 2019 InfuuData. All rights reserved. 25 Learning Resources • Forecasting: Principles and Practices • Results From Comparing Classical and Machine Learning MEthods for Time Ser ies • Holt-Winters Resources ◦ Holt-Winters Forecasting Simplifi ed ◦ Holt-Winters Forecasting for Dum mies (or Developers) ◦ Holt-Winters Method | Real Statist ics Using Eucel ◦ Initializing the Holt-Winters meth od • Infuu Resources ◦ HOLT-WINTERS() • Nelder-Mead ◦ Nelder-Mead algorithm ◦ Nelder-Mead Animation ◦ Nelder-Mead Optimization • Statistics ◦ MAE and RMSE — Which Metric is Bette r? ◦ What’s a good value for R-squared? ◦ Ridge regression to minimize RMSE ins tead of MSE ◦ Confused on Residual Terminology ◦ Relationship between RMSE and RSS ◦ Why we usually choose to minimize the SSE • Minimizing squared error to regression line
26. • Community Highlights anais@infuudata.com Community Page community.infuudata.com Community Slack
27. Thank You!!