Time Series Forecasting with Amazon SageMaker & Chronos

Production-ready pipeline for time series forecasting using Amazon SageMaker and Chronos-bolt-tiny.

Original Chronos repository: amazon-science/chronos-forecasting

Overview

End-to-end ML pipeline for time series prediction:

1. Data preparation and upload to S3
2. Model fine-tuning on SageMaker (AutoGluon + Chronos)
3. Model deployment as REST API endpoint
4. Local Docker-based inference (FastAPI + Uvicorn)
5. Model comparison and evaluation

Model

amazon/chronos-bolt-tiny - Lightweight transformer model for time series forecasting.

Features:

• Multivariate forecasting
• Missing value handling
• Temporal dependency capture
• Fast CPU inference

Documentation

Document	Description
Usage Guide	Step-by-step workflow
API Reference	Endpoints, request/response formats
AWS Setup	IAM, ECR, S3 configuration
Local Development	Running locally with Docker

Architecture

Configuration

All settings in config.yaml:

s3:
  bucket: chronos-presmanes

sagemaker:
  training:
    limit_time: 300
    instance_type: ml.g4dn.xlarge
  inference:
    instance_type: ml.t2.medium
  endpoint_name: chronos-endpoint-prod

TODO

• EDA for testing base model locally
• Training script with AutoGluon locally
• Training job in SageMaker
• Deploy model to SageMaker endpoint
• Inference endpoint testing
• Streamlit app for user interaction