Notebooks
Eight hands-on Colab notebooks covering AutoGluon, tabular foundation models, and time series — from first contact to pretraining your own TFM.
Browse notebooks → · GitHub →Tutorial Paper
The accepted KDD’26 tutorial paper describing the motivation, outline, and prior presentations.
Coming soonAutoGluon
Open-source AutoML framework powering this tutorial. State-of-the-art on tabular and time-series benchmarks.
auto.gluon.ai →Chronos-2
New multivariate zero-shot time-series foundation model from AWS. Extends Chronos to universal forecasting.
chronos-ts.ai →About This Tutorial
Structured data drives enterprise decision-making, yet building predictive pipelines for time-series and tabular modalities requires intensive feature engineering, model selection, and other “tricks of the trade”. AutoGluon, an open-source AutoML system, automates this through multi-layer stack ensembling, providing a unified API that reliably achieves state-of-the-art accuracy across time series and tabular data.
Concurrently, Foundation Models for Structured Data (FM4SD) have emerged to push the boundaries of predictive performance, enabling both powerful zero-shot inference and efficient fine-tuning. Specifically, Chronos-2 delivers state-of-the-art multivariate forecasting for time series, while a rapidly growing ecosystem of tabular foundation models (e.g., TabPFN, Mitra, TabICL) transforms how practitioners execute classification and regression tasks.
This system-focused tutorial teaches attendees how to operationalize FM4SDs using AutoML across temporal and tabular modalities. We guide participants through a technical progression: establishing AutoGluon as the orchestration framework, detailing the internal mechanics of Chronos-2 and various tabular FMs, and unifying them into production-ready pipelines.
Key Topics
AutoGluonChronos-2TabPFNMitraTabICLTime SeriesTabular DataFM4SDAutoMLStack EnsemblingZero-shot Forecasting
Requirements
- Laptop with a modern web browser
- Google Colab (free tier with CPU/T4 GPU)
- Basic Python and ML familiarity
No GPU requiredFree tier sufficient
Tutorial Outline (3 hours)
The tutorial guides attendees from foundational AutoML principles to the deployment of cutting-edge FM4SD engines.
Overview of structured data challenges (tabular & time series). Introduction to AutoGluon as the unified AutoML framework. Quick setup check to ensure all attendees can initialize the libraries in Colab.
Tabular dataTime seriesAutoGluon introSetup
The philosophy of multi-layer stacked ensembling vs. hyperparameter optimization. Coding Exercise: Attendees build base TabularPredictor and TimeSeriesPredictor pipelines, experiencing the high-level API for different modalities.
Stack EnsemblingTabularPredictorTimeSeriesPredictorCoding Exercise
Chronos-2 (Time Series): Understanding time series tokenization, group attention, and multivariate zero-shot forecasting. Tabular Foundation Models: Understanding TabPFN, Mitra, and TabICL. Handling complex covariates, defining custom evaluation metrics, and optimizing inference speeds for AutoML systems of tabular foundation models.
Chronos-2TokenizationGroup AttentionTabPFNMitraTabICL
Best practices for deploying AutoGluon and FM4SDs effectively. Common pitfalls in data formatting and evaluation. “Bring Your Own Data” (BYOD) challenge: Attendees upload a small CSV from their own domain for 1-on-1 architecture advice from tutors.
BYOD ChallengeBest Practices1-on-1 Advice
Hands-On Colab Notebooks
Each technical section culminates in a guided API coding exercise. Open any notebook directly in Google Colab using the links below, or browse the GitHub repository — every notebook ships with saved outputs so you can read along without running.
Notebook 01 — First Contact
Naive XGBoost vs AutoGluon-bagged XGBoost vs TabICLv2 vs TabFM on one real dataset. Opens with the TFM at-a-glance table (constraints & licenses).
Open in Colab →Notebook 02 — AutoGluon Essentials
Fit, predict, evaluate, and inspect a TabularPredictor; pick the right metric first, finish with the extreme preset.
Notebook 03 — The TFM Zoo
Many foundation models behind one predictor: TabICLv2, TabDPT-Turbo, TabPFN-3, bagged and ensembled, with the classical toolkit one uncomment away.
Open in Colab →Notebook 04 — Inside the Prior
Generate the synthetic datasets TFMs are pretrained on, live, and see what the learnability filter rejects. CPU only.
Open in Colab →Notebook 05 — Pretrain Your Own
Pretrain a working TabPFN-style transformer in about a minute with nanoTabPFN, then apply it zero-shot to real data.
Open in Colab →Notebook 06 — Thinking Mode
TabPFN-3 thinking mode conquers the high-cardinality-categorical dataset where every TFM loses to CatBoost. Needs a free Prior Labs token.
Open in Colab →Notebook 07 — Interpretability & Serving
Shapley explanations for TabPFN-3 predictions via shapiq — which financial ratios flagged this company — plus the KV-cache fit_mode that makes repeated inference fast.
Notebook 08 — Non-IID Validation
A naive split vs validation_structure on grouped data: the naive estimate is off 25× and picks the wrong model. CPU only.
Notebook 09 — Time Series Forecasting
Forecasting with AutoGluon’s TimeSeriesPredictor and time-series foundation models.
What Participants Learn
This tutorial provides a complete technical journey from foundational AutoML principles to cutting-edge foundation model deployment for structured data.
SETUPENSEMBLINGFORECASTINGFINE-TUNINGDEPLOYMENTBYOD
- Understand the structured data challenge landscape
- Use AutoGluon’s unified API for tabular and time-series tasks
- Apply Chronos-2 for zero-shot multivariate forecasting
- Leverage TabPFN, Mitra, and TabICL for tabular tasks
- Fine-tune FM4SDs and fuse with classical baselines
- Evaluate with custom metrics and uncertainty quantification
init, fit, predict), we dedicate the final 15 minutes to a “Bring Your Own Data” challenge where tutors provide 1-on-1 architecture advice on real participant data.
Tutorial Organizers
Nick Erickson
Prior Labs
Lead developer and maintainer of AutoGluon and TabArena. Research scientist focused on AutoML and foundation models for structured data.
Yuyang (Bernie) Wang
NTT Data AIVista
Member of Technical Staff at NTT Data AIVista. In his previous role as a Principal Scientist at AWS, he worked to democratize advanced AI/ML capabilities, making them accessible to practitioners across diverse domains.
Boran Han
Amazon Web Services
Senior applied scientist working on AutoML, foundation models for structured data, and agentic frameworks. Core contributor to AutoGluon and MLZero.
Abdul Fatir Ansari
Amazon Web Services
Lead author of the Chronos series. Researcher working on deep learning for time series, probabilistic forecasting, and foundation models.
Oleksandr Shchur
Amazon Web Services
Lead author of AutoGluon-TimeSeries. Researcher specializing in probabilistic time series forecasting and AutoML for temporal data.
Xiyuan Zhang
Amazon Web Services
Lead author of Mitra and co-author of the Chronos series. Applied scientist working on foundation models for tabular and time series data.
Shuai Zhang
Amazon Web Services
Senior Applied Scientist at AWS. His research focuses on LLMs, retrieval-augmented generation, reinforcement learning, and machine learning systems.
Haoyang Fang
Amazon Web Services
Applied scientist working on agentic frameworks to enhance RL performance. Develops AutoGluon Multimodal and leads MLZero, a multi-agent system for AutoML.
Danielle Maddix
Amazon Web Services
Senior Applied Scientist at AWS AI. Danielle has worked on statistical and deep learning foundation models for structured data. She also leads the sciML research initiative.
Michael Bohlke-Schneider
Amazon Web Services
Applied Science Manager at AWS. At AWS, Michael works on machine learning and forecasting, with a focus on foundation models for structured data and AutoML.
Prior Presentations
AutoGluon-Tabular: Robust and Accurate AutoML for Structured Data
Virtual — August 2020
Introduced AutoGluon’s core ensembling architecture for tabular data. Focused on robust, high-accuracy AutoML without manual tuning.
Multimodal AutoML for Image, Text and Tabular Data
Washington, DC — August 2022 — ~200 attendees
Extended AutoGluon to multi-modal settings including text and image data alongside tabular modalities.
Key Papers & Resources
| Paper | Venue | Link |
|---|---|---|
| AutoGluon-Tabular: Robust and Accurate AutoML for Structured Data — Erickson et al. | AutoML Workshop @ ICML 2020 | arXiv |
| AutoGluon-TimeSeries: AutoML for Probabilistic Time Series Forecasting — Shchur et al. | AutoML Conference 2023 | arXiv |
| Chronos: Learning the Language of Time Series — Ansari et al. | TMLR 2024 | arXiv |
| Chronos-2: From Univariate to Universal Forecasting — Ansari et al. | arXiv 2025 | arXiv |
| Accurate predictions on small data with a tabular foundation model (TabPFN) — Hollmann et al. | Nature 2025 | Nature |
| Mitra: Mixed Synthetic Priors for Enhancing Tabular Foundation Models — Zhang et al. | NeurIPS 2025 | arXiv |
| TabICL: A Tabular Foundation Model for In-Context Learning on Large Data — Qu et al. | ICML 2025 | arXiv |
Societal Impact
Improved tabular and time series modeling impacts critical resource allocation in healthcare, finance, and logistics. While mastering tools like AutoGluon and its engines democratizes powerful predictive capabilities, it carries the risk of automation bias—where users blindly trust ML outputs.
This tutorial addresses this by emphasizing rigorous data science practices such as probabilistic forecasting and uncertainty quantification, ensuring attendees learn to evaluate model uncertainty and apply appropriate skepticism to automated predictions.