Learning asset embeddings from partial institutional holdings in China's A-share market. Companion training code for the working paper "Do Asset Embeddings Need Complete Holdings? Lessons from China" (2026).
China's A-share market is institutionally distinct: retail investors dominate trading volume, institutional holdings account for only ~20% of floating shares, and the disclosure regime alternates between fully observed (semi-annual / annual) and partially observed (quarterly top-10) panels. Asset characteristics shift on quarterly horizons under industrial-policy and regulatory cycles. Whether asset embeddings need complete holdings to work — or whether the partial panels China actually discloses are enough — is the question the companion paper answers; this repository provides the training pipeline that learns the embeddings.
Following Gabaix et al. (2025), we extract latent asset characteristics from institutional holdings: each investor's portfolio is a sentence and each stock is a token. Rank order carries economic meaning — the largest position is the primary exposure, smaller positions are supplementary — just as word order carries meaning in language.
The estimation departs from Gabaix et al. in two respects, and both answer one question: how can latent demand factors be estimated when holdings are only partially observed? China's hybrid disclosure rules act as a censoring filter on the true holdings, and the two departures are designed for exactly the places where that censoring binds.
Pretrain–finetune instead of independent quarterly fits. Embeddings fitted quarter by quarter are not directly comparable over time: the demand system identifies each quarter's solution only up to a rotation, so quarter-to-quarter differences need not represent true changes in the asset. Pretraining once on the first 80% of holding observations (chronological split, cutoff 2023Q1: 73 quarters for pretraining, 6 for quarterly fine-tuning) and initializing every quarterly fine-tune from that base makes the rotations more likely to agree across quarters. Pretraining learns the common, stable holding structure; finetuning captures the quarter-specific variations on top of it.
Warm-starting BERT's embedding layer from Word2Vec. At both the pretraining and fine-tuning stages, BERT's embedding layer starts from the trained Word2Vec embeddings rather than random vectors (cold-start). Under top-ten disclosure, many assets appear in only a handful of reported portfolios in Q1/Q3 — the warm-start gives exactly these thinly observed assets a sensible prior.
Without either departure, the performance of the BERT embeddings generally deteriorates; Gabaix et al. find pretraining unnecessary in the uniformly disclosed US market, and the paper shows this is not the case in China.
- Three model families trained on institutional portfolios:
AssetBERT(transformer MLM),AssetW2V(Skip-gram/CBOW), andAssetRS(PCA/ICA recommender-system baselines). - The pretrain–finetune pipeline producing quarterly embedding sequences that are comparable over time.
- A text-embedding pipeline that builds LLM-based embeddings (OpenAI / Cohere / Voyage / Gemini / local HF models) from firm names, the paper's text-based comparison point.
- The data pipeline from raw CSMAR shareholding dumps to model-ready portfolio corpora, plus a config system and batch-generation tooling for reproducible experiment grids.
In the paper, the resulting embeddings outperform text embeddings, align with investor demand, and capture the state-ownership structure specific to the A-share market — and a controlled censoring experiment shows the top-ten disclosure rule China already imposes is enough for them to add value.
git clone https://github.com/Initial-Singularity/AssetEmbeddingsInChina
cd AssetEmbeddingsInChina
uv sync --extra notebook # notebook + ipykernel are demo-only extras
uv run jupyter lab examples/quickstart.ipynb # or: uv run python examples/quickstart.pyThe notebook runs the full W2V-PT → W2V-FT → BERT-PT → BERT-FT training chain on a synthetic dataset shipped in examples/data/ (CPU, ~1 minute) and produces four asset-embedding CSVs, the fine-tuned ones comparable to their pretrained starting points. Real production runs require CSMAR data; see docs/how-to/prepare-data.md.
The trained embeddings are archived on Zenodo under CC-BY-4.0: 10.5281/zenodo.20640781 (concept DOI — always resolves to the latest version).
asset_embeddings.tar.gz— AssetBERT / AssetW2V / AssetRS per-stock embeddings at dimensions 4–128 for 2023Q2–2024Q3.text_embeddings.tar.gz— full-dimensional masters for the nine text embedders plusregenerate.py, which rebuilds anylanguage × transform × dimension × quartervariant.
These are the learned representations, not CSMAR source data — download them directly without CSMAR access. See the deposit's own README.md for the schema and reproducibility notes. Reproducing the embeddings from raw data still requires CSMAR (Prepare data).
@unpublished{cheng2026asset,
title = {Do Asset Embeddings Need Complete Holdings? Lessons from China},
author = {Cheng, Zitong and Huang, Dashan and Liu, Xiaobin and Zeng, Tao},
year = {2026},
note = {Working paper},
}
@dataset{cheng2026asset_data,
title = {Asset \& Text Embeddings for the China A-share Market},
author = {Cheng, Zitong and Huang, Dashan and Liu, Xiaobin and Zeng, Tao},
year = {2026},
publisher = {Zenodo},
doi = {10.5281/zenodo.20640781},
}(Corresponding author: Xiaobin Liu, liuxb53@mail.sysu.edu.cn.)
The full docs live in docs/ and are published at https://initial-singularity.github.io/AssetEmbeddingsInChina/.
See CONTRIBUTING.md.
MIT — see LICENSE.