A paper that integrates supply chain graphs with risk factor zoo to build asset pricing factors
The link to our paper draft is here.
Authors: Agostino Capponi, Jose Sidaoui, Jiacheng Zou. RA support by Dehan Cui.
Contact Jiacheng jiachengzou@gmail.com or Jose jas2545@columbia.edu for questions.
We provide the python code to automate building of 64 firm characteristics from Freyberger et al, RFS 2020, and these firms' stock returns for users with access to Wharton Research Data Services (WRDS).
First, users need to follow the instructions here to manually click a few download links.
Then, running our python backend code in the jupyter "FirmFeatures_Calculations.ipynb" (located in "Data" folder) produces a "features.csv" file of the panel data of all NYSE, AMEX, and NASDAQ traded stocks in the user-specified date range.
We document detailed formulas and variable names here.
GNN code.ipynb is a self-contained jupyter notebook that implements our model in the paper. The first part of the notebook contains all of our functions to implement the benchmark models in our paper (PCA, RP-PCA, NC Ridge, NC LASSO) as well as the full implementation of our Graph-Neural Network model and construction of our asset pricing factors. The notebook may be run in order. The input datasets are "features.csv" produced from the jupyter notebook in the data folder, "ff_mon.csv" which are the monthly FF5 factors, and "features_and_supply_chain.csv" which is our merged features dataset with the supply chain relationship data. The jupyter notebook GNN code additional.ipynb contains the code for the additional experiments added to the May 2026 version of our paper. These include the new NN multi-layer perceptron benchmark, the Hansen-Jagannathan distance experiments, the spanning regressions, our new SCG factor and deciles (sorted on the GNN predicted signal rather than the embeddings), the oversmoothing experiments for the GNN embeddings, and the multi-step prediction tests with transaction costs.