File size: 2,177 Bytes
a445ccd 6b11f75 a445ccd | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 | ---
license: mit
tags:
- finance
- retrieval
- candlestick
- k-line
- vicreg
- contrastive-learning
- RAG
---
# FinAlogy — VICReg Morphology-Aware Encoder
This repository contains the pretrained visual encoder and demo data for **FinAlogy**, a visual analogy retrieval system for financial K-line analysis.
- 💻 Code: [https://github.com/nice-zzy/FinAlogy](https://github.com/nice-zzy/FinAlogy)
- ▶️ Demo Video: https://youtu.be/SIreBcl__WU
---
## Files
| File | Description |
|---|---|
| `checkpoint_best.pth` | Pretrained VICReg encoder checkpoint (CLIP-ViT-B/32 backbone, 675 MB) |
| `finalogy_demo_instances.zip` | Demo K-line instances for quick inference testing |
---
## Model Description
The encoder is a **CLIP-ViT-B/32** backbone fine-tuned with **VICReg** self-supervised learning on DOW30 historical K-line data (2010–2021). It maps candlestick chart images into a 512-dimensional morphology-aware embedding space, capturing shape-level properties (body size, shadow structure, local trend direction) while remaining invariant to absolute price scales.
**Training details:**
- Backbone: CLIP-ViT-B/32
- Loss: VICReg (λ = µ = 5)
- Data: DOW30 components, 2010–2020 (train), 2021 (test)
- Window size: W = 5 trading days
- Silver-pair threshold: τ = 0.98
**Evaluation (test set, 2,075 unique anchors):**
| Method | 52D Alignment |
|---|---|
| Barlow Twins | 0.785 |
| SimSiam | 0.814 |
| **VICReg (ours)** | **0.910** |
> 52D Alignment measures cosine similarity between retrieved results and the query in the 52-dimensional morphological feature space — higher is better. VICReg achieves the best morphological alignment among all compared methods.
---
## Usage
```python
import torch
from huggingface_hub import hf_hub_download
# Download checkpoint
ckpt_path = hf_hub_download(
repo_id="ZiyaZhao/FinAlogy",
filename="checkpoint_best.pth"
)
# Load encoder
checkpoint = torch.load(ckpt_path, map_location="cpu")
# See https://github.com/nice-zzy/FinAlogy for full inference pipeline
```
For the full retrieval and report generation pipeline, see the [GitHub repository](https://github.com/nice-zzy/FinAlogy). |