The notebook with the experiments is available here The study is a replication and enhancement of the clustering proposed by Han(2022), with our own KMeans Optimization
Data was harvested from Dacheng Xiu's web-site (https://dachxiu.chicagobooth.edu/download/datashare.zip), it is in GBs of size. To observe the data structure and do your own tests, see the provided samples from 2019 to 2021: "./data/sample_historic_characteristics.csv" Quality is questionable, and clean up was required - see notebook. The CRSP and various securities data was unavailable to me at the time of writing, there was no way to map PERMNO to SnP constituants nor make a reversal benchmark mentioned in the paper.
In abscence of adjusted return (accessible from WRDS), the momentum was used as a proxy.
1.DATE: The end day of each month (YYYYMMDD) 2.permno: CRSP Permanent Company Number 3-96. 94 Lagged Firm Characteristics (Details are in the appendix of Han(2022)) - lagged as these are released by CRSP with a delay, which we I assume is 1 month. 97.sic2: The first two digits of the Standard Industrial Classification code on DATE
The models were inspired by the paper:
@article{han2023pairs,
title={Pairs trading via unsupervised learning},
author={Han, Chulwoo and He, Zhaodong and Toh, Alenson Jun Wei},
journal={European Journal of Operational Research},
volume={307},
number={2},
pages={929--947},
year={2023},
publisher={Elsevier}
}Cite these papers if using their datasets:
@article{gu2020empirical,
title={Empirical asset pricing via machine learning},
author={Gu, Shihao and Kelly, Bryan and Xiu, Dacheng},
journal={The Review of Financial Studies},
volume={33},
number={5},
pages={2223--2273},
year={2020},
publisher={Oxford University Press}
}@article{gu2021autoencoder,
title={Autoencoder asset pricing models},
author={Gu, Shihao and Kelly, Bryan and Xiu, Dacheng},
journal={Journal of Econometrics},
volume={222},
number={1},
pages={429--450},
year={2021},
publisher={Elsevier}
}