TimesFM
A large-scale time-series forecasting foundation model with parameter counts in the billions range. Its scale results in relatively high computational resource requirements for both training and inference.
Timeline (last 90 days)
TimesFM models can only learn the mapping between covariates and the target that appeared in the training data, and cannot infer the elasticity of variables that never changed in history; traditional machine learning models are similarly limited.
The TimesFM model cuts historical data into patches of 32 time steps, treating each patch as a token, similar to how a language model predicts the next word to forecast the next segment of the trend.
Time series foundation models (TSFMs) refer to models pretrained on large-scale multi-domain time series data that can be transferred to downstream forecasting tasks via zero-shot or lightweight fine-tuning; representative works include Chronos, TimesFM, Moirai, and others.
Representative work on time-series foundation models includes TimeGPT, Lag-Llama and TimesFM.
Google's TimesFM adopts a decoder-only patch-based architecture, splitting continuous time segments into patches that are fed into a Transformer, with the training objective directly regressing numeric values.
TimesFM is a decoder-only transformer whose input tokens are patches of 32 consecutive numbers, with a prediction task of the next 128 numbers.
The code and weights of TimesFM 2.5 and earlier versions are open-sourced under the Apache-2.0 license.
TimesFM originally comes from the ICML 2024 paper "A decoder-only foundation model for time-series forecasting."
Lumina uses Google's TimesFM for time series forecasting, outputting a 5-day probabilistic fan chart with an 80% confidence interval and a crash probability estimate.
Time-Series Foundation Models are general-purpose forecasting models pretrained on large-scale, multi-domain time-series data with cross-task generalization capabilities, with representative works including TimesFM, Chronos, and Moirai.
16 more timeline events
All Facts (20)
Google officially disclosed that TimesFM was pre-trained on 200B time points of data, covering multiple domains including finance, retail, energy, and healthcare.
70%UnverifiedTimesFM adopts patch-based tokenization, packing several consecutive time steps into a patch and treating it as a single token, similar to the image patch strategy used by ViT.
70%UnverifiedTimesFM models can only learn the mapping between covariates and the target that appeared in the training data, and cannot infer the elasticity of variables that never changed in history; traditional machine learning models are similarly limited.
50%UnverifiedThe TimesFM model cuts historical data into patches of 32 time steps, treating each patch as a token, similar to how a language model predicts the next word to forecast the next segment of the trend.
50%UnverifiedTime series foundation models (TSFMs) refer to models pretrained on large-scale multi-domain time series data that can be transferred to downstream forecasting tasks via zero-shot or lightweight fine-tuning; representative works include Chronos, TimesFM, Moirai, and others.
50%UnverifiedRepresentative work on time-series foundation models includes TimeGPT, Lag-Llama and TimesFM.
50%UnverifiedGoogle's TimesFM adopts a decoder-only patch-based architecture, splitting continuous time segments into patches that are fed into a Transformer, with the training objective directly regressing numeric values.
50%UnverifiedTimesFM originally comes from the ICML 2024 paper "A decoder-only foundation model for time-series forecasting."
50%UnverifiedThe code and weights of TimesFM 2.5 and earlier versions are open-sourced under the Apache-2.0 license.
50%UnverifiedTimesFM is a decoder-only transformer whose input tokens are patches of 32 consecutive numbers, with a prediction task of the next 128 numbers.
50%UnverifiedLumina uses Google's TimesFM for time series forecasting, outputting a 5-day probabilistic fan chart with an 80% confidence interval and a crash probability estimate.
50%UnverifiedTime-Series Foundation Models are general-purpose forecasting models pretrained on large-scale, multi-domain time-series data with cross-task generalization capabilities, with representative works including TimesFM, Chronos, and Moirai.
50%UnverifiedGoogle TimesFM is a foundation model dedicated to time series forecasting, capable of zero-shot prediction without requiring specialized training on specific business data.
50%UnverifiedTimesFM 3.0 is a foundation model dedicated to time series forecasting, with zero-shot forecasting capability.
50%UnverifiedRepresentative time-series foundation models include Google's TimesFM, Amazon's Chronos, and Salesforce's Moirai.
50%UnverifiedTimesFM draws on the pre-training + zero-shot inference paradigm of large language models, enabling direct forecasting of unseen time series after pre-training without targeted training.
50%UnverifiedTimesFM adopts a decoder-only Transformer architecture, consistent with the design philosophy of the GPT series of models.
50%UnverifiedThe patch-based input approach is inspired by the Vision Transformer's idea of dividing images into patches.
50%UnverifiedTimesFM provides implementations based on both the JAX and PyTorch deep learning frameworks.
50%UnverifiedTimesFM is implemented in Python and released as open source on GitHub, providing both source code and pre-trained weights.
50%