TimeRSI

Time-Series Foundation Model Deployment Needs a Self-Improving Harness

TimeRSI pairs frozen TimesFM 3.0 with a code agent that develops task-specific forecasting pipelines. The agent revises features, model inputs and forecast corrections using historical backtests; forecast-time selection and composition use available task context and feedback from completed windows. TimesFM weights remain frozen.

Environment

Linux, Python 3.12, a CUDA-capable NVIDIA GPU, Docker, Claude Code 2.1.245, and access to Fable 5 through Vertex AI.

git clone https://github.com/haoxin1998/TimeRSI.git
cd TimeRSI
python3.12 -m venv .venv
source .venv/bin/activate
python -m pip install -e .
python -m pip install 'timesfm[torch] @ git+https://github.com/google-research/timesfm.git@45e0a3bc7fc4acef17b7ba7910488be2159bae5f'
docker build -t timersi-isolation:1.1 -f docker/Dockerfile docker
export TIMERSI_VERTEX_PROJECT=YOUR_PROJECT
gcloud auth application-default login

Install Claude Code and make claude available on PATH. Use google/timesfm-3.0-pytorch or a local checkpoint. See the usage guide for running a task or benchmark.

Code

• Task runner: code research followed by chronological forecast composition.
• Code research and forecast composition.
• Execution environment: The public runner uses Docker to isolate code-agent sessions and generated candidates. Implementation.

Benchmarks

FEV-Bench results, evaluation instructions and mini20 inspection trajectories · TIME result files.

TimesFM weights have their own license; see the checkpoint license.

Updated at 2026-10-14 02:02 PDT.