Skip to content

Quickstart

Run the full pretrain–finetune training pipeline end-to-end on synthetic data in about a minute — no CSMAR data, no GPU — and come away with four asset-embedding CSVs (W2V-PT, W2V-FT, BERT-PT, BERT-FT). This is the paper's estimation in miniature.

Setup

git clone https://github.com/Initial-Singularity/AssetEmbeddingsInChina
cd AssetEmbeddingsInChina
uv sync --extra notebook

Run it

uv run jupyter lab examples/quickstart.ipynb
# or, headless:
uv run python examples/quickstart.py

The notebook uses the synthetic dataset shipped in examples/data/ — every number is generated, so no external data is needed.

What it does

The notebook runs the W2V-PT → W2V-FT → BERT-PT → BERT-FT chain — the same training stages the paper uses, on toy data:

  1. W2V-PT — train a Skip-gram Word2Vec on the synthetic portfolio "sentences".
  2. W2V-FT — continue training W2V on the fine-tune period, starting from the pretrained vectors (PT→FT).
  3. BERT-PT (pretrain) — train BERT on the pooled history, its embedding layer warm-started from the W2V-PT matrix (W2V→BERT).
  4. BERT-FT (finetune) — refine BERT per "quarter", the encoder initialized from the pretrained model (PT→FT) and the embedding layer warm-started from W2V-FT (W2V→BERT).

Each stage writes its embeddings as a (Token, Embed_1..Embed_d) CSV — the deliverable of the pipeline.

What you should see

The run ends with a summary table of the four embedding CSVs (model, training data, shape, path) and a quick peek at what they encode: the top-5 cosine neighbours of one probe stock in each space. The synthetic universe has latent cluster structure, so a stock's neighbours should stay within its cluster; the fine-tuned models pick up the FT-period drift together, while the initializations (W2V→BERT, PT→FT) keep all four embeddings comparable with one another. (Exact neighbours vary with the synthetic seed, but the cluster structure is stable.)

Next steps