Перейти к содержимому

Hierarchical Time Series Forecasting | Intermittent Demand (M5 Comp)

Matt Dancho (Business Science)

0:00 / 0:00

Hierarchical Time Series Forecasting | Intermittent Demand (M5 Comp)

23 063 просмотра · 5 л. назад
Matt Dancho (Business Science)
29,8 тыс. подписчиков
23 063 просмотра · 5 л. назад
📖 Learning Labs PRO (get code & #shiny app): https://university.business-science.i... 😀 ABOUT: In Learning Labs PRO Episode 50, Matt tackles an in-depth tutorial on Hierarchical Forecasting using the M5 #Forecasting Competition. This is a challenging forecasting problem that includes intermittent demand, when demand becomes very granular with lots of zeros. This is also a hierarchical dataset, where there are 50 lower-level time series that are aggregated by the organization's departments and product types. We'll use #Modeltime along with #Tidymodels and XGBoost, LightGBM, and CatBoost Machine Learning Algorithms. 📋 INTRODUCTION: Agenda - M5 Forecasting Competition | Tidymodels, Treesnip, Modeltime | XGBoost, LightGBM, CatBoost - 00:00 Introducing the Shiny Hierarchical Forecaster App - 3:46 Business Problem - What is Hierarchical Demand Forecasting & Why Do I Care? - 7:38 Why Learn Tidymodels? 11:00 📖 FULL CODE TUTORIAL Project Setup - 11:55 Part 1 - XGBoost vs LightGBM vs CatBoost - 14:00 LightGBM Basic Usage (without Tidymodels ☹️) - 15:31 Classification: XGBoost, LightGBM, & CatBoost (with Tidymodels😎) - Agaricus - 17:37 Regression CV: XGBoost, LightGBM, & CatBoost (with Tidymodels😎) - Diamonds - 22:59 Part 2 - FULL HIERARCHICAL FORECASTING TUTORIAL - 25:46 Load the Data, Reshape & Join - 27:19 Quick EDA: Skim Data & Visualize Sales Trends for 6 Product Items - 30:51 FEATURE ENGINEERING: Making the "Full Dataset" - 33:30 Discussion: Hierarchical Forecasting Strategies & Alternatives - 40:01 Splitting Full Data - Data Prepared / Future Data - 44:55 Time Splitting - Train/Test Sets - 46:03 Preprocessing Pipeline (Time Series Features & One-Hot Features) - 46:58 MACHINE LEARNING - 49:37 MODELTIME - Model Comparison & Selection - 53:21 ENSEMBLE LEARNING - Combine Your Best Models into a Super Model - 1:03:12 CONCLUSIONS - 380 Lines of Code for a High-Performance Forecast is GOOD, but can IMPROVE - 1:07:01 LLPRO BONUS - Shiny App Code - Hierarchical Forecaster - 1:09:30 🧙‍♂️ LEARNING RECOMMENDATIONS How do I learn what Matt just taught? - 1:11:30 👉Is Learning Labs PRO for me? - 1:13:00 - https://university.business-science.i.... What if I'm just starting & learning R shiny much deeper? - 1:14:20 Is the R-Track right for me? - 1:15:00 👉15% OFF R-Track: https://university.business-science.i...