What is zero-shot forecasting?

Zero-shot forecasting allows the model to predict future values for a time series it has never seen before, leveraging patterns learned during pretraining without requiring a specific training phase for your data.

Can I use this skill for anomaly detection?

Yes. By using the quantile forecasts as prediction intervals, any actual data points falling outside the 90% confidence interval (q10–q90) can be flagged as statistically unusual anomalies.

What are the hardware requirements for TimesFM 2.5?

TimesFM 2.5 requires at least 4 GB of RAM (CPU) or 2 GB of VRAM (GPU), along with approximately 800 MB of free disk space for the model weights.

Does it support multivariate time series?

TimesFM is a univariate model, meaning it forecasts one variable at a time. However, it supports exogenous variables (covariates) like weather or holidays that influence the primary variable.

Is a GPU required to run this skill?

No, it can run on a CPU, though inference will be slower. The included preflight checker will automatically detect your hardware and configure the model for optimal performance.

TimesFM Time Series Forecasting

Name: TimesFM Time Series Forecasting
Author: henriquescastilho

byhenriquescastilho

•

数据科学与机器学习

Performs zero-shot, univariate time series forecasting using Google’s TimesFM foundation model to generate point forecasts and prediction intervals without custom training.

TimesFM Forecasting integrates Google's pretrained decoder-only foundation model into the Claude environment, enabling high-accuracy predictions for any univariate time series data. This skill eliminates the need for manual model training or complex parameter tuning by leveraging a foundation model approach that handles up to 16,384 context points. It provides both point forecasts and calibrated quantile intervals (10th–90th percentiles), making it ideal for demand planning, sensor monitoring, and financial trend analysis. To ensure stability, the skill includes a mandatory preflight system checker that verifies hardware resources (RAM, GPU, and Disk) before downloading or loading the model weights.

主要功能

011 GitHub stars

02Preflight system checker to prevent machine crashes by verifying RAM and GPU availability

03Zero-shot forecasting requiring no custom model training or fine-tuning

04Support for exogenous variables (covariates) to enhance prediction accuracy

05Flexible data ingestion from CSV files, Pandas DataFrames, and NumPy arrays

06Probabilistic predictions with 10 calibrated quantile intervals for uncertainty estimation

使用场景

01Forecasting energy consumption or weather patterns using long-context temporal data

02Predicting retail sales demand and inventory requirements without historical training cycles

03Identifying anomalies in sensor or infrastructure data using prediction interval boundaries

主要功能

011 GitHub stars

02Preflight system checker to prevent machine crashes by verifying RAM and GPU availability

03Zero-shot forecasting requiring no custom model training or fine-tuning

04Support for exogenous variables (covariates) to enhance prediction accuracy

05Flexible data ingestion from CSV files, Pandas DataFrames, and NumPy arrays

06Probabilistic predictions with 10 calibrated quantile intervals for uncertainty estimation

使用场景

01Forecasting energy consumption or weather patterns using long-context temporal data

02Predicting retail sales demand and inventory requirements without historical training cycles

03Identifying anomalies in sensor or infrastructure data using prediction interval boundaries