11-03, 11:00–11:40 (America/New_York), Music Box (Room 5411)
Traditional time series models such as ARIMA and exponential smoothing have typically been used to forecast time series data, but the use of machine learning methods have been able to set new benchmarks for accuracy in high profile forecasting competitions such as M4 and M5.
However, the use of machine learning models can easily lead to inferior results under common conditions. This talk is a discussion of how each of these methods can be used to model time series data, and demonstrate how SKTime provides a unified framework for implementing both families of techniques.
Popular forecasting competitions such as M4 and M5 have demonstrated that machine learning and deep learning models have the capability to deliver powerful and accurate results for a variety of time series problems.
However, without proper treatment of data and a suitable feature engineering pipeline, results are likely to disappoint practitioners expecting breakthrough results.
This talk is meant to give a brief but comprehensive overview of the different issues people ought to consider when deciding how to decide what method is appropriate for a particular problem.
Issues discussed will include:
- The current SOTA found on different types of problems in the forecasting literatures
- Successful use cases for tabular machine learning models such as LightGBM
- Global vs Local forecasting
- Data processing techniques that can make ML models more effective time series forecasters
A common problem with using machine learning models in forecasting is that the chain of transformations you typically need to perform on your data can be very time consuming and difficult to perform across training and test sets.
We'll also discuss how SKTime provides a unified interface for handling many problems that arise due to the inoperability of different steps in the forecasting process. The presentation will provide examples of how different composition pipelines can be used for both sets of models to provide a simplified workflow for many forecasting problems.
Previous knowledge expected
Jonathan is the principal data scientist for the Data Science and Machine Learning Research Group, which is a consultancy that allows ambitious academics to do transformative work in the commercial sector. He's also a contributor and community council member for SKTime.
In the past he's worked with organizations such as General Assembly, NYPD, Amber Capital and Advent International to assist them with their data science needs.