fin-jepa

KO | EN

gitlite — search

#docker #proxy #artificial-intelligence #github #mcp #cloud-native #ai-agents #sql #macos #bash

(★ 9)

Fin-JEPA: Joint-Embedding Predictive Representation Learning for Financial Time Series

파일 탐색기

ZIP 다운로드

- - fig1_training_loss.png
  - fig2_ablation.png
  - fig3_downstream_auc.png
  - fig4_tsne.png
  - fig_architecture.png
- jepa_full_draft.md
- jepa_paper.pdf
- jepa_paper.tex
- phase1_train.py
- phase2_train.py
- prepare_dataset.py
- submit_hf.sh
- sweep.py
- train_jepa.py
- train_lewm.py
- train_lewm_v2.py
- - meta.json
- - meta.json
- - meta.json
- - meta.json
- - meta.json
- - meta.json
- - meta.json
- - meta.json
- - meta.json
- - meta.json
- - meta.json
- - meta.json
- - meta.json
- - meta.json
- - meta.json
.gitignore
analyze.py
benchmark.py
compare_arch.py
compare_arch2.py
experiment_e.py
ijepa.py
leworldmodel.py
model.py
model_lewm.py
README.md

# CDN으로 사용하기

jsDelivr

jsDelivr는 공개 GitHub 리포지토리를 별도 설정 없이 CDN으로 즉시 서빙합니다. 버전과 파일을 고르면 웹페이지에 바로 붙일 수 있는 링크와 예시 코드가 만들어집니다.

버전

파일

링크

예시

# 프로젝트 배지

DOI 바로가기

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

**Fin-JEPA: Joint-Embedding Predictive Representation Learning for Financial Time Series**

This repository contains the official implementation of Fin-JEPA, the first application of the Joint-Embedding Predictive Architecture (JEPA) to financial time series.

## Overview

Fin-JEPA learns 64-dimensional latent representations of daily equity features through a lightweight 367K-parameter architecture comprising:
- **PriceEncoder**: MLP with GELU activations and LayerNorm
- **TransformerPredictor**: 4-layer causal Transformer (n_heads=4, embed_dim=64)
- **SIGReg**: Sketched Isotropic Gaussian Regularization for collapse prevention

## Repository Structure

```
├── model.py              # Fin-JEPA model architecture
├── compare_arch.py       # Architecture ablation training script
├── experiment_e.py       # Downstream evaluation (VoE analysis)
├── compare_arch2.py      # Encoder variant + density ablation
├── analyze.py            # Analysis utilities
├── model_lewm.py         # LeWorldModel-aligned variant
├── ijepa.py              # I-JEPA baseline (MNIST)
├── leworldmodel.py       # LeWorldModel baseline (MNIST)
├── benchmark.py          # Benchmark utilities
├── hf/                   # Hugging Face Jobs training scripts
│   ├── train_jepa.py
│   ├── phase1_train.py
│   ├── phase2_train.py
│   ├── submit_hf.sh
│   └── ...
├── output/               # Ablation results (meta.json)
│   └── arch_*/
└── docs/
    ├── jepa_paper.tex    # Paper source (LaTeX)
    └── figures/          # Paper figures
```

## Reproducibility

See the paper's Reproducibility Statement for details. The architecture ablation experiments (Section 4.2) are fully reproducible using publicly available market data.

## Citation

```bibtex
@article{wang2026finjepa,
  title={Fin-JEPA: Joint-Embedding Predictive Representation Learning for Financial Time Series},
  author={Wang, Yihan and others},
  year={2026}
}
```

## License

MIT

fin-jepa

파일 탐색기

# CDN으로 사용하기

# 프로젝트 배지

Was this content helpful?

Recent Feedback

Download README