Data Machina #261

Substack · 2 years ago · 38

Recent research demonstrates integration of generative AI and language models with time-series forecasting through methods like vision-language models, specialized VAEs, and tokenization approaches. Datadog's Toto foundation model trained on one trillion time series data points achieved state-of-the-art zero-shot performance on multiple benchmarks, while JPMorgan's LETS-C method achieved better time-series classification accuracy using only 14.5% of the trainable parameters compared to existing solutions. These advances enable organizations to deploy more efficient and accurate forecasting systems across domains including weather, electricity, stock prices, and retail optimization.