Presentations

Jun 26, 2025 09:05 AM
Speaker: Rob Hyndman

25 years of open source forecasting software

Open source forecasting software has evolved from a few basic R functions to a rich ecosystem of R, Python, Julia, and C/C++ implementations of established models. In this session, the historical development of these tools will be reviewed and key gaps and areas requiring further attention will be highlighted.

Jun 26, 2025 09:45 AM
Speaker: Mitchell O'Hara-Wild
Discussant: Anthony Bagnall

Designing extensible forecasting frameworks

A well-designed interface is crucial for making statistical forecasting software easy to learn and use, particularly when defining a cohesive design language for the forecasting workflow. In this session, the design decisions behind the fable forecasting framework will be examined, and the extensibility of workflow elements such as models, accuracy metrics, and forecasting techniques will be discussed.

Jun 26, 2025 10:55 AM
Speaker: Daniele Girolimetto
Discussant: Charupriya Sharma

FoReco: From foundations to frontiers

Forecast reconciliation is presented as a solution to the inconsistency problems that arise when independently generating multiple time series forecasts across different aggregation levels. The session topics will be focused on how the FoReco R package implements classical and modern reconciliation methods (including hierarchical, grouped, temporal, cross-temporal, non-negative, and probabilistic reconciliation under various constraints), along with future extensions such as standardized interfaces, machine learning-based reconciliation, coherent forecast combination, and a forthcoming Python implementation.

Jun 26, 2025 11:35 AM
Speaker: Lorenzo Zambon & Anubhab Biswas
Discussant: Olivier Sprangers

Probabilistic forecast reconciliation with bayesRecon

Forecasts in hierarchical time series often violate aggregation constraints, motivating probabilistic reconciliation methods and software solutions. In this session, the bayesRecon (R) and bayesreconpy (Python) packages are introduced and demonstrated, and their probabilistic reconciliation approach, support for multiple forecast types (Gaussian, non-Gaussian, discrete) and mixed hierarchies, unified cross-language interface, and licensing are discussed.

Jun 26, 2025 01:45 PM
Speaker: Kandrika Pritularga
Discussant: Adam Wang

Forecasting smoothly with smooth() package

ADAM (Augmented Dynamic Adaptive Model) is a flexible, R-implemented forecasting framework that unifies and extends ARIMA and ETS methods to handle multiple data structures, distributions, and modelling scenarios. In this session, key topics will be covered including the use of diverse distributions and estimation strategies, the production of flexible forecasts and prediction intervals, and the application of ADAM's tools for model tuning, selection, diagnostics, outlier treatment, and handling of missing data.

Jun 26, 2025 02:25 PM
Speaker: Filotas Theodosiou
Discussant: Pablo Montero-Manso

Integrating Smooth Capabilities into Python

The smooth package's Augmented Dynamic Adaptive Model (ADAM) provides a unified state-space framework that extends classical time series models like ETS, ARIMA, and regression, originally developed in R and now motivating a native Python implementation. In this session, the engineering details of the Python implementation will be presented, including the C++-backed architecture, the refactoring into an idiomatic object-oriented Python API, and the key engineering lessons for cross-language development and use of Large Language Models to accelerate such efforts.

Jun 26, 2025 03:35 PM
Speaker: Tomasz Wojciech Wozniak
Discussant: Rob Hyndman

bsvars.org design concept: R packages for Predictive Analyses using Bayesian Structural Vector Autoregressions

bsvars.org is a family of R packages that provide a fast, user-friendly framework for Bayesian Structural Vector Autoregression analysis in macroeconomics and finance. In the session, key topics will be covered including the use of bsvars and bsvarSIGNs for structural and predictive analysis, the implementation of Bayesian estimation and related tools (forecasting, variance decompositions, impulse responses, historical decompositions), and the challenges and design choices in developing C++-backed R packages for diverse users.

Jun 26, 2025 04:15 PM
Speaker: Anirban Ray and Franz Király
Discussant: Christoph Bergmeir

Open Source Forecasting in Python: A Survey of Tools, Trends, and Trade-offs

Time series forecasting has become a core component of data-driven decision-making, and the Python ecosystem now offers a rapidly expanding range of specialized tools to support this need. In this session, a comprehensive review will be presented of open-source Python forecasting packages and their technical capabilities, governance models, openness vs. gated designs (including foundation models), and the practical trade-offs that influence tool selection and adoption.

Jun 27, 2025 09:00 AM
Speaker: Pablo Montero-Manso
Discussant: Kandrika Pritularga

Forecasting 2.0: A Framework for Near-Optimal Time Series Forecasting and Inference via Pre-Trained Models

A new open-source framework for time series modeling is introduced that uses neural networks trained on simulated data to deliver near-optimal forecasting and inference, outperforming traditional models and complex neural architectures while remaining accessible in R and Python. In the session, the underlying “grammar” for optimal time series model design, the use of pre-trained neural estimators as drop-in replacements for classical methods, and the extensibility of the framework to non-Gaussian, nonlinear, censored, and time-varying scenarios will be discussed.

Jun 27, 2025 09:40 AM
Speaker: Anthony Bagnall
Discussant: Mitchell O'Hara-Wild

The aeon toolkit for time series machine learning

aeon is an open source Python toolkit for time series machine learning, developed by a diverse research-focused team and now expanding its capabilities in forecasting. In this session, an overview of aeon and its newly redesigned forecasting module will be presented, and future development directions and community feedback on the forecasting roadmap will be invited.

Jun 27, 2025 10:50 AM
Speaker: Azul Garza
Discussant: Filotas Theodosiou

LLMs Meet Time Series Foundation Models: Are We Ready for Forecasting Agents?

AI agents are rapidly reshaping many AI applications, but their potential for time series forecasting is still underexplored. In this session, the core concepts of AI agents and their architectures will be revisited, their evolution with foundation models will be traced, and the unique challenges and opportunities of applying agent-based design to forecasting within the open-source ecosystem will be examined.

Jun 27, 2025 11:30 AM
Speaker: Haixu Wu
Discussant: Arjun Ashok

Deep Time Series Forecasting: Tools and Open Challenges

Deep learning has recently shown promise for time series forecasting, raising questions about when it truly outperforms traditional statistical methods. In this session, recent advances in deep time series models will be reviewed, the open-source libraries TSLib and OpenLTM will be introduced, and future research directions in deep learning-based time series forecasting will be discussed.

Jun 27, 2025 01:40 PM
Speaker: Charupriya Sharma
Discussant: Daniele Girolimetto

Hierarchical Forecast Reconciliation on Networks: A Network Flow Optimization Formulation

Hierarchical forecasting often fails to guarantee coherence across aggregation levels (such as state, district, and city), motivating reconciliation methods like FlowRec that reformulate the problem as a network flow optimization to handle general network structures efficiently. In this session, rigorous complexity results for reconciliation under various loss functions, the FlowRec algorithm's computational and structural advantages over MinT (including dynamic, localized, and approximate updates), and empirical evidence of its runtime, memory, and accuracy improvements will be presented.

Jun 27, 2025 02:20 PM
Speaker: Arjun Ashok
Discussant: Joaquin Amat Rodrigo

Context-Aided Forecasting: Progress So Far and Next Big Challenges

Classical numerical forecasting models are being extended by incorporating rich natural-language context to better capture the information human forecasters routinely use. In this session, key topics will be the CiK benchmark for context-aided forecasting, the Region-of-Interest CRPS scoring rule, the evaluation and prompting of LLM-based forecasters (including bootstrapping and in-context learning), community efforts and hybrid TSF/LLM methods, and future challenges around interactive, agentic forecasting systems.

Jun 27, 2025 03:30 PM
Speaker: Resul Akay
Discussant: Yangzhuoran Yang

Durbyn.jl: A Julia Framework for Time Series Forecasting

Durbyn.jl is a lightweight, fully Julia-native framework for time series forecasting that balances clarity with computational power for researchers, educators, and practitioners. In the session, its modular architecture, rich suite of classical models, and built-in statistical tools will be discussed and demonstrated.

Jun 27, 2025 04:10 PM
Speaker: Mariana Menchero García
Discussant: Resul Akay

A Look Under the Hood of StatsForecast

StatsForecast is a widely used open-source Python library for statistical and econometric forecasting, originally developed as a Python implementation of models from the forecast R package and now recognized as a key tool in the field. In this session, the architecture and performance of StatsForecast will be examined, the implementations of major model families (including ARIMA, ETS, Theta, MSTL, baseline, and sparse/intermittent series models) will be explored, and planned future features and models will be discussed.
No matching items