DEVOURED
Toto 2.0: Time series forecasting enters the scaling era
Datadog has released Toto 2.0, an open-weights family of time series forecasting models up to 2.5B parameters on Hugging Face, demonstrating improved performance and scalability without saturation.
What: Datadog launched Toto 2.0, a family of open-weights time series forecasting models ranging from 4 million to 2.5 billion parameters, now available on Hugging Face under the Apache 2.0 license. Toto 2.0 sets new state-of-the-art results on benchmarks like BOOM, GIFT-Eval, and TIME, showing monotonic improvement with scale and 7x greater parameter efficiency than Toto 1.0. It features contiguous patch masking (CPM) for faster inference.
Why it matters: Toto 2.0 signifies a significant advancement in time series forecasting, proving that scaling models reliably improves performance in this domain, similar to how it has in NLP and vision. This makes powerful, general-purpose time series foundation models more accessible and practical for a broader range of applications, especially in observability.
Takeaway: Developers working with time series data can experiment with Toto 2.0 models on Hugging Face to improve forecasting accuracy and inference speed for their applications.
Decoder
- Time series forecasting: The process of predicting future values of a variable based on its past values, often used for data that changes over time, like stock prices, sensor readings, or website traffic.
- Open-weights: Refers to a model where the trained parameters (weights) are publicly released, allowing anyone to download, use, and further develop the model.
- Hugging Face: A popular platform and community for machine learning, providing tools, datasets, and pretrained models, often for natural language processing and computer vision.
- CRPS (Continuous Ranked Probability Score): A metric used to evaluate the quality of probabilistic forecasts, measuring how well the predicted distribution of outcomes aligns with the actual observed outcomes.
- MASE (Mean Absolute Scaled Error): A measure of point forecast accuracy, normalized against a naive seasonal baseline, making it suitable for comparing forecasts across different time series.
Original article
Full article content is not available for inline reading.
Explore all 57 articles in today's digest
Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi