11-09, 14:45–15:30 (America/New_York), Radio City (6th floor)
We will explore how statistics can be combined with new deep learning methods to increase usability, accuracy and interpretability in the forecasting field.
We will outline latest deep learning innovations and introduce the NixtlaVerse: a group of open-source Python libraries written in Pytorch and NumBa that facilitate the use of statistical and neural forecasting models to data scientists and developers.
Attendees will gain theoretical and practical knowledge on how to build robust and scalable forecasting pipelines.
Time-series modeling – analysis, and prediction of trends and seasonalities for data collected over time – is a rapidly growing category of software applications.
Businesses, ranging from finance to healthcare analytics, collect time-series data daily to predict patterns and build better data-driven product experiences. For example, temperature and humidity prediction is used in manufacturing to prevent defects, streaming metrics predictions help identify music's popular artists, and sales forecasting for thousands of SKUs across different locations is used to optimize inventory costs. As data generation increases, the forecasting necessities have evolved from modeling a few time series to predicting millions. Explainable and scalable forecasting remains a challenging task for strategic decision-making.
During the last decades, the forecasting field was dominated by statistical techniques like Auto-Regressive Integrated Moving Average (ARIMA) and Exponential Smoothing (ETS). However, with the explosions in datasets size, neural networks, and machine learning techniques regained their popularity in forecasting and showed to be quite effective, simple, and scalable. Although neural Networks have proven powerful and flexible, they are not easily interpretable, which constitutes a barrier to their wider adoption. Interpretability is critical because it encourages trust and widens the knowledge of the modeled process.
In this talk, we will explore how we can combine econometrics and statistics with neural network innovations to advance usability, usefulness, and interpretability in forecasting.
We will deep dive into different SoTA models like the ES-RNN, N-BEATS, N-HiTS, and large-scale benchmarking. We will also introduce the NixtlaVerse: a group of open-source python libraries that facilitate the use of these competition-winning models for data scientists and developers. We wrote Nixtla in PyTorch and NumBa, focusing on usability, speed, and reproducibility.
The talk is intended as an intermediary introduction to the field. It aims to introduce different theoretical elements and practical tips to help the attendees implement robust and accurate forecasting pipelines with a better understanding of the intricacy of the time series field. Basic knowledge of Python and high school math is expected.
The talk will be outlined as followed:
- Introduction to time series
-- Lightning Fast Statistical Forecating: StatsForecast
-- AutoARIMA
-- AutoETS
-- Large Scale Benchmarking
- Interpretable Deep learning Forecasting: NeuralForecast
-- N-BEATSx -> univariate point forecasting with exogenous variables
-- N-HiTS -> Long-Horizon Forecasting
-- ES-RNN -> winner of the M4 competition
- Practical examples and conclusions
No previous knowledge expected
CEO and Co-Founder of Nixtla, a time-series forecasting startup. Previously he was CTO and Co-Founder of Levo (YC S21). Max has worked in the ML industry for the last decade, where he has built and led ML teams. He has co-authored different papers on forecasting algorithms and decision theory. He is a co-maintainer of different open source libraries in the python ecosystem. His passion is the intersection between business and technology.