timesfm-forecasting

TimesFM Forecasting - Zero-shot Time Series Forecasting

Overview of Skills

TimesFM Forecasting is a pre-trained time series foundation model developed by Google Research. It can forecast any univariate time series without training. It supports many data types such as sales, sensors, energy consumption, vital signs, and weather, and provides point forecasts and quantile-based prediction intervals.

Use Cases

1. Sales and Demand Forecasting

Forecast business metrics such as product sales, inventory demand, and order quantities. Without having to collect large amounts of historical data to retrain the model, you can input sales data directly to obtain future trend predictions. It supports batch forecasting for the time series of hundreds of products or stores.

2. Sensor and IoT Data Analysis

Handle industrial sensor data such as temperature, pressure, and flow rate, as well as IoT scenarios like energy usage monitoring and equipment status monitoring. The model supports context windows up to 16,384 data points, can process high-frequency sampled sensor time series, and provides prediction intervals for anomaly detection.

3. Rapid Prototyping and Benchmarking

As a zero-shot baseline model for time series forecasting, it can quickly assess the feasibility of a forecasting task. Compared with traditional statistical models like ARIMA/ETS that require parameter tuning, TimesFM is ready to use out of the box, making it suitable for data exploration, model comparison, and rapid prototype development.

Core Features

1. Zero-shot Forecasting

TimesFM uses a 200M-parameter pre-trained model and does not require training or fine-tuning for a specific dataset. It supports multiple input formats such as CSV, DataFrame, and NumPy arrays. By simply providing historical time series data, the model outputs forecasts and prediction intervals based on 10 quantiles.

2. System Resource Checks

A built-in forced precheck system runs the script check_system.py to automatically verify system requirements—available memory (≥4GB), GPU availability and VRAM, disk space (≥2GB), Python version, and more—before the model is loaded for the first time, preventing model load failures due to insufficient resources.

3. Batch Forecasting and Covariate Support

Supports simultaneous forecasting of hundreds of univariate time series; internal batching improves efficiency. TimesFM 2.5+ supports external variables (price, promotions, holidays, etc.) via the forecast_with_covariates() API. Combining baseline model forecasts with covariate adjustments improves prediction accuracy.

Frequently Asked Questions

Does TimesFM require training?

No. TimesFM is a pre-trained foundation model that uses zero-shot forecasting. You just need to provide historical time series data, and the model can output forecast results directly without training or fine-tuning for your dataset. This is the main advantage of TimesFM over traditional time series models.

What are TimesFM’s system requirements?

TimesFM 2.5 (recommended) uses a 200M-parameter model and requires at least 4GB RAM (CPU mode) or 2GB VRAM (GPU mode), plus about 800MB of disk space to store model weights. On the first run, the model automatically downloads weights from HuggingFace and caches them. It is recommended to run python scripts/check_system.py before loading the model to check system configuration.

Can TimesFM be used for anomaly detection?

TimesFM itself does not include built-in anomaly detection, but anomaly detection can be performed using prediction intervals. The model’s quantile-based forecasts include 10 quantiles (q10–q90). If the actual value falls outside the 90% prediction interval (q10–q90), it can be considered a statistical outlier. The skill includes a complete anomaly detection example, demonstrating how to implement two-stage anomaly detection by combining linear trend removal from historical data and Z-score computation.