Time-series forecasting

   

⏱️ Time to complete: 20 min

Get the code

git clone https://github.com/anyscale/templates && cd templates/templates/e2e-timeseries-forecasting

These tutorials implement an end-to-end time-series application including:

• Distributed data preprocessing and model training: Ingest and preprocess data at scale using Ray Data. Then, train a distributed DLinear model using Ray Train.

• Model validation using offline inference: Evaluate the model using Ray Data offline batch inference.

• Online model serving: Deploy the model as a scalable online service using Ray Serve.

• Production deployment: Create production batch Jobs for offline workloads including data prep, training, batch prediction, and potentially online Services.

Setup

Run the following:

uv pip install -r python_depset.lock --system --no-deps --no-cache-dir --index-strategy unsafe-best-match && uv pip install -e . --system

Acknowledgements

This repository is based on the official DLinear implementations:

• DLinear

• LTSF-Linear

And the original publication:

• “Are Transformers Effective for Time Series Forecasting?”