Data Machina #252

Data Machina · 2 years ago

Multiple new foundation models and techniques for time-series forecasting are emerging, including diffusion models, transformers, state space models, and hybrid approaches like MambaForer and Google's TimesFM trained on 100 billion data points. IBM Research released TinyTimeMixers with under 1 million parameters, demonstrating that pre-trained models can be made compact while addressing key issues in deep neural network time-series models such as training complexity and inference costs. These innovations are producing competitive results against traditional statistical methods and enabling zero-shot performance across different domains.