The Evolution of Time Series Forecasting: From ARIMA to Foundation Models

Time series forecasting has come a long way from classical statistical models like ARIMA and ETS, to machine learning methods such as Random Forest and XGBoost, and most recently to advanced deep learning architectures like NBEATSx, TFT, and TiDE. Each stage has expanded our ability to capture seasonality, trends, and external drivers at scale. In this session, we’ll explore this evolution in a structured way, highlighting when to use statistics, machine learning, and deep learning for forecasting. Finally, we’ll introduce Moirai, a new class of foundation models for time series that aim to generalize across domains, transfer knowledge, and reduce the need for task-specific retraining, a potential GPT moment for forecasting. These models bring scalability, adaptability, and transfer learning into time series, enabling organizations to quickly apply insights from one domain to another, handle diverse data with minimal fine-tuning, and accelerate the deployment of high-accuracy forecasts across industries like supply chain, finance, and healthcare.

Key Takeaways:

• Foundation Model for Time Series: Moirai is trained on massive, diverse time series datasets, enabling broad generalization across domains similar to how GPT works for text.
• Masked Encoder Architecture: Uses masked pretraining to learn universal temporal representations, capturing both local patterns (seasonality, spikes) and global structures (long-term trends).
• Transferability: Once pretrained, Moirai can be fine-tuned with minimal data for specific tasks (forecasting, anomaly detection, classification).
• Versatility: Supports univariate and multivariate series, and handles heterogeneous covariates, making it suitable for industries like finance, healthcare, and supply chain.
• Scalability & Efficiency: Eliminates the need to train a separate model per dataset. Organizations can adapt a single pretrained model for multiple use cases.
• The GPT Moment for Forecasting: Moves time series from bespoke, task-specific models (Stats, ML, DL) toward general-purpose, reusable models, accelerating deployment and reducing costs.
• GitHub: https://github.com/redoules/moirai

Prerequisites:

• Basic knowledge of Python
• Familiarity with Pandas/NumPy
• Some exposure to ML/forecasting concepts is helpful but not required

Data Scientist who enjoys connecting the dots, be it ideas from different disciplines, people from different teams, or applications from different industries. Ankit's passion for coding started during high school and has only grown over time, particularly in the areas of different domains of Data Science. His experience working in the technology divisions of various corporations such as Ascentt Systems Inc, Radial Inc, Native Agtech Inc, and Grad Dreams Education Consulting Pvt Ltd has provided him with invaluable industry insights. As the integration of businesses and information technology continues to grow, he is committed to leveraging his expertise to bridge the gap between technology and business effectively. His passion lies in solving business problems with tailored data and algorithms while communicating complex ideas to non-technical stakeholders. He is able to jump across verticals to deliver high-performing AI solutions. You can reach him on LinkedIn.