(EMNLP2026) MetaCaster: Meta-Harness-Optimized Agent for End-to-End Few-Shot Learning of Lightweight Time Series Forecasters

English | 简体中文

🔍 About · 🧩 Framework · 🚀 Quick Start · 📊 Results · 🔗 Citation

🔍 About

MetaCaster is a multi-agent framework for few-shot time-series forecasting. Instead of using a large language model as the forecaster, MetaCaster uses agents to generate task-adaptive training data and train compact, task-specific Forecasters for downstream deployment.

🔑 Key Features

• Few-shot learning for lightweight Forecasters: MetaCaster addresses lightweight Forecaster learning under data scarcity, training a task-specific, deployable forecasting model end to end from only a few time-series examples and textual context.
• Meta-harness-optimized multi-agent framework: Agents act as intermediary engineers rather than forecasters. HPAgent optimizes MGAgent's Harness against downstream forecasting quality, MGAgent generates task-adaptive training data, and FTAgent trains and selects the best lightweight Forecaster.
• Unified lightweight forecasting library: LT-Lib brings together 23 state-of-the-art lightweight Forecasters from 2022–2026 behind unified configuration, training, and evaluation interfaces for model development and selection across architectures.
• High-quality and efficient forecasting: Experiments across 18 datasets, 23 lightweight Forecasters, and 14 baselines validate MetaCaster's effectiveness. After deployment, only the selected lightweight Forecaster remains, with no Agent or LLM on the inference path.

🧩 Framework

Figure 1. MetaCaster's Harness Optimization framework.

MetaCaster provides a unified entry point that runs MGAgent and FTAgent sequentially:

python -m agents.metacaster

It contains three agents:

Component	Role	Entry point
MGAgent	Generates task-adaptive training windows with the trained Harness	python -m agents.mgagent.agent
FTAgent	Trains LT-Lib Forecasters in parallel and selects Top-1 by validation MSE	python -m agents.ftagent.agent
HPAgent	Optimizes the MGAgent Harness against LT-Lib Forecasters	python -m agents.hpagent.agent

The release already includes the final trained Harness. Run HPAgent only when training a new Harness from scratch.

🚀 Quick Start

Requirements: Python 3.12, uv, an LLM provider key, and a CUDA-capable GPU for Forecaster training. First prepare the environments and data from the release root:

uv sync
uv sync --project lt_lib
cp .env.example .env  # Fill in the models, API key, and data paths
uv run python scripts/download_gift_eval.py --all
uv run python scripts/prepare_gift_eval.py --all

DATASET="your-dataset-name"
INPUT_DIR="data/GIFT-Eval/test/${DATASET}/agent_view_k30"
TEST_FILE="data/GIFT-Eval/test/${DATASET}/test.npy"

1. Complete workflow: generate data and train a downstream Forecaster

uv run python -m agents.metacaster \
  --input-dir "$INPUT_DIR" \
  --test "$TEST_FILE" \
  --gpus 0,1,2,3 \
  --tune \
  --output work_dir/metacaster

MGAgent output is saved under work_dir/metacaster/generated/. The selected model, configuration, and training results are saved under work_dir/metacaster/forecaster/. By default, FTAgent trains the 20-model main pool; pass --all-models to train all 23 Forecasters or --models to choose specific models.

2. Generate data only

uv run python -m agents.mgagent.agent \
  --input-dir "$INPUT_DIR" \
  --output work_dir/generated \
  "Generate task-adaptive training windows"

The generated data is saved to work_dir/generated/dataset.npy, with its validation report at work_dir/generated/validation_report.json.

3. Train a downstream Forecaster only

uv run --project lt_lib python -m agents.ftagent.agent \
  --synthetic /path/to/dataset.npy \
  --input-dir "$INPUT_DIR" \
  --test "$TEST_FILE" \
  --gpus 0,1,2,3 \
  --main-only \
  --tune \
  --output work_dir/forecaster

Omit --main-only to train all 23 Forecasters, or use --models to select specific models. Use agents.ftagent.predict to run the selected model directly; see lt_lib/README.md.

4. Train a New Harness from Scratch

uv run python -m agents.hpagent.agent \
  --gpus 0,1,2,3

HPAgent iteratively optimizes MGAgent's Harness and writes run results under work_dir/runs/.

📁 Repository Structure

agents/                Unified MetaCaster entry point and individual Agent entry points
generation/            MGAgent runtime
harness/               Final trained router and generation skill
optimizer/             Harness Optimization runtime
lt_lib/    LT-Lib: 23 Forecasters and training runtime
scripts/               GIFT-Eval data preparation
eval/                  Generation-quality evaluation
docs/assets/           Framework and results figures

See lt_lib/README.md for LT-Lib installation and model validation.

📊 Results

Figure 2. Forecasting MSE on IND and OOD tasks with $K=10, 30, 50$; MetaCaster achieves the best result in 19 of the 30 dataset–shot settings.

🔗 Citation

@inproceedings{shen2026metacaster,
  title     = {MetaCaster: Meta-Harness-Optimized Agent for End-to-End Few-Shot Learning of Lightweight Time Series Forecasters},
  author    = {Shen, ChengAo and Yu, Wenchao and Wu, Fangyu and Song, Dongjin and Tong, Hanghang and Luo, Dongsheng and Cheng, Wei and Chen, Haifeng and Ni, Jingchao},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing},
  year      = {2026}
}

📧 Contact

If you have any questions or concerns, please contact us: cshen9 [at] uh [dot] edu or submit an issue.