AFRICAST 2026 – Day 5: Residual diagnostics and foundation models
Forecasting for Social Good
0:00 / 0:00
AFRICAST 2026 – Day 5: Residual diagnostics and foundation models
3 просмотра · 2 часа назад
Forecasting for Social Good
377 подписчиков
3 просмотра · 2 часа назад
This is the Day 5 recording of AFRICAST 2026, a free online course in tidy time series analysis and forecasting with R. It was run by Forecasting for Social Good (F4SG) from 19 to 23 October 2026.
🔁 Session 1: Residual diagnostics and cross-validation (Bahman Rostami-Tabar, Cardiff University)
More reliable ways to evaluate and check forecasting models:
• Problems with a single train/test split
• Time series cross-validation with rolling training sets
• Residual diagnostics: residual plots, the ACF of residuals and combined diagnostic graphs
• The Ljung–Box test
🤖 Session 2: Time series foundation models (Harsha Halgamuwe Hewage, Cardiff University)
What pretrained foundation models are, and how to use them for forecasting:
• What a foundation model is, and how attention combines information from the history
• Models built for time series vs language models adapted for forecasting
• Transfer learning: pretraining, zero-shot forecasting and fine-tuning
• An overview of TimeGPT and Chronos-2
• A hands-on lab forecasting Australian tourism with zero-shot and fine-tuned models (including LoRA), compared with a seasonal naïve benchmark
• Evidence from forecasting health supplies in Lao PDR
📚 Materials
• Course website (slides, exercises and recordings): https://workshop.f4sg.org/africast/
• Day 5 slides: https://workshop.f4sg.org/africast/se... and https://workshop.f4sg.org/africast/se...
• Day 5 exercises: https://workshop.f4sg.org/africast/se...
• Get the exercises project in R: `usethis::use_course("https://workshop.f4sg.org/africast/ex...")`
• Textbook: Forecasting: Principles and Practice (3rd ed.), free at https://otexts.com/fpp3/
🌍 About AFRICAST
AFRICAST is part of the Forecasting for Social Good (F4SG) initiative. It builds forecasting skills among researchers in Africa so that forecasts can inform planning and resource allocation. The project is led by Bahman Rostami-Tabar (Cardiff University). The instructors are Mitchell O'Hara-Wild, Abiodun Egwuenu, Harsha Halgamuwe Hewage and Mustafa Aslan, and Henry Kissinger Ochieng leads coordination. Learn more at https://www.f4sg.org/
#RStats #Forecasting #TimeSeries #F4SG #Africa #tidyverts #FoundationModels