(1/2) Google released a new foundation model for time series forecasting 🚀
The TimeFM (Time Series Foundation Model) is a foundation model for time series forecasting applications. This pre-trained model was developed by the Google Research team. It joins the recent trend of leveraging foundation models for time series forecasting, which includes Salesforce's Moirai and Amazon's Chronos.
Version 0.12.0 of the skforecast Python library for time series forecasting with regression models was released this week. The release includes new features, updates for existing ones, and bug fixes. 🧵👇🏼
Version 1.7.1 of the NeuralForecast #Python library was released last month by Nixtla. The NeuralForecast library, as the name implies, provides a neural network framework for time series forecasting. 🧵👇🏼
The skforecast Python 🐍 library provides ML applications for time series forecasting using different regression models from the scikit-learn library. Here is a tutorial by Joaquín Amat Rodrigo and Javier Escobar Ortiz for time series forecasting with the skforecast using XGBoost, LightGBM, Scikit-learn, and CatBoost models 🚀.
Introduction to forecasting with ARIMA and seasonal ARIMA with Python 🚀👇🏼
This short tutorial by Joaquín Amat Rodrigo and Javier Escobar Ortiz provides an introduction to forecasting with ARIMA models using Python 🐍. This includes using different flavors of ARIMA methods from the statsmodels, pmdarima, skforecast, and statsForecast libraries.
(1/2) Moirai - Salesforce's Foundation Forecasting Model 🚀
Salesforce recently released Moirari - a new #Python 🐍 library with a foundation model for time series forecasting applications. According to the release blog - the model comes with universal forecasting capabilities and can handle multiple scenarios and different frequencies.
Yesterday, Amazon released a new open-source project, Chronos - a family of pre-trained time series forecasting models based on language model architectures.
If you are looking to learn the foundation of time series analysis and forecasting, I recommend checking the Applied Time Series Analysis course from the University of Washington. The course by Prof Eli Holmes, Prof. Eric Ward, and Prof. Mark Scheuerell focuses on the foundation of time series analysis with applications for environmental science.
(2/2) The course covers the following topics:
✅ Time series decomposition
✅ Covariance and correlation
✅ Autoregressive (AR) models
✅ Moving average (MA) models
✅ Forecasting with ARIMA models
✅ Univariate state-space models
✅ Dynamic linear models
✅ Hidden Markov models
(1/2) Microsoft Forecasting Framework for Finance Time Series 🚀
The finnts is an R package from Microsoft for automated forecasting for financial forecast. This includes the following features:
✅ Automated feature engineering and feature selection
✅ Training models with backtesting, scoring, and ranking
✅ Supports more than 25 models for univariate and multivariate time series
✅ Supports daily, weekly, monthly, quarterly, and yearly frequency #RStats#DataScience#forecasting#MachineLearning
Thanks to Forecasting for Social Good for inviting me to deliver a workshop on "Forecasting with Generalised Additive Models (GAMs) in R" today! We talked about:
❓ What are GAMs?
❓ How do we fit GAMs in R?
❓ Model evaluation and forecasting
❓ When GAMs go wrong!
(1/3)Modeling Short Time Series with Prior Knowledge in PyMC 🚀
Yesterday, I shared an article by Tim Radtke about forecasting insufficient time series data with a Bayesian approach using R. Here is the Python version 🧵👇🏼
(2/3) The TLDR is when you need to model a short time series (less than one seasonal cycle) and have some knowledge or assumption about the expected behavior of the series - either from a similar series (i.e., similar products or geos) you can translate those assumptions to the model's prior distributions and use it to build a forecasting model.
(1/2) Modeling Short Time Series with Prior Knowledge
When modeling time series data, you may find yourself with insufficient data. Insufficient data in time series would typically be defined as less than one seasonal cycle. This would challenge us to understand whether some events are driven by seasonality or other reasons, such as one-time events, outliers, etc.
(2/2) If you have some prior knowledge about the series (e.g., learning from similar products or goes, etc.), you should consider using the Bayesian approach.
The article below by Tim Radtke provides an example of how to incorporate prior assumptions into a time series forecasting model when having insufficient data.
Weird day. We were expected to get 6-10 inches of snow starting at 2 am and lasting through this afternoon. Schools closed. City declared a snow emergency. But it just started snowing at 8:30 and it looks like it'll only be a dusting.
Hudson Valley #Weather: “This is one of the more severe changes [to a forecast] we’ve seen in recent memory. For the data to be unanimous in the result… and then to all begin shifting dramatically SE in about 12 to 18 hours, is a rare event.” #forecasting
(2/3) The talk focuses on methods to identify patterns in time series using cluster analysis inspired by Prof Rob Hyndman's paper about Feature-based time series analysis 👇🏼
Seminar 🔗: https://robjhyndman.com/seminars/fbtsa-ssc/
Radar at 12/21 1804 currently shows the storm is parked offshore from Los Angeles (dumping quite a bit there). Still raining in Oxnard. #rain#CAwx#LosAngeles#SanDiego
So, the ghost in the HRRR machine puts totals near Santa Barbara (on 12/22/23 at 9am) at less than an Inch. But at 12/22/23 at 10am then says it thinks total accumulation will be 5 inches. It's likely a glitch (the HRRR accumulation algorithm), but I do wonder what's going to be correct. #mystery#HRRR#wx#weather#forecasting