Code originally written in 2021/2022 and refactored to use uv and polars in 2026.

Code for learning distributed representations (embeddings) of financial assets from returns time series, relating to the paper Stock Embeddings: Learning Distributed Representations for Financial Assets [pdf].

Run notebooks/pointwise_embeddings.ipynb. It is the notebook for the original Stock Embeddings paper and walks the whole pipeline end-to-end: load returns, build target/context sets, train embeddings, and evaluate them on sector classification. After uv sync and dropping the data files into Data/ (see Data), it runs top to bottom in a few minutes.

Each notebook maps to a paper:

A related third paper, leveraging financial news for multimodal asset embeddings [pdf], does not have code in this repo.

The project is managed with uv. To create the environment and install the stock_embeddings package with all dependencies:

uv sync

Run the test suite:

uv run pytest

The notebooks import the installed package, e.g.:

from stock_embeddings import ReturnsData, ClassificationEmbeddings

Data files are too large to be stored in the repo but feel free to reach out to me via email: [email protected]

The code expects two files in Data/:

• Data/returns_df_611.csv — daily returns matrix (date index, one column per ticker)
• Data/historical_stocks.csv — ticker metadata (ticker, name, sector, industry)

The first 10 rows of each data file can be found in the sample files in the Data/ directory to show the required format.

ReturnsData locates these files by searching upward from the current working directory, so the notebooks run regardless of whether Jupyter is launched from the repo root or the notebooks/ directory.

@article{dolphin2022stock,
  title={Stock embeddings: Learning distributed representations for financial assets},
  author={Dolphin, Rian and Smyth, Barry and Dong, Ruihai},
  journal={arXiv preprint arXiv:2202.08968},
  year={2022}
}