anomalies-replication-python

This repository is a replication of the work of Chen and Zimmermann (2022), which highlights the successful replication of anomalies. Their project provides valuable data on firm characteristics and portfolio returns. And many reserachers are using their data in academic research. For the original study and more information, please refer to their project of Open Source Cross-Sectional Asset Pricing.

• Everything in Python (Chen and Zimmermann use Stata and R)
• Firm characteristics are in parquet format
  • Faster read and write speed
  • Smaller file size
  • For R users, to read parquet data, you can install arrow package install.packages("arrow")
  • For Stata users, to read parquet data, see this guidance: https://github.com/mcaceresb/stata-parquet
• 96 different ways to compute portfolio returns (see example in step 3 of How to use)
• Easy to update firm characteristics and portfolio returns

Replication process

I have followed the code structure and firm characteristic definitions outlined in Chen and Zimmermann's work to replicate various anomalies in Python. Specifically, I have replicated anomalies that rely on CRSP or Compustat data. Please note that anomalies requiring data from other sources, such as IBES or OptionMetrics, have not been included in this replication.

There are a total of 138 anomalies. To view the complete list of anomalies, please refer to the predictors.csv file.

Warning

• The code has been tested on macOS Sonoma version 14.1 with Python 3.10.13.
• And it is easy to run the bash files in any Linux system.
• Windows users should use with cautions (I do not have windows enviornment to test it)

How to use

0_a. Required packages

Before you proceed to run the code, please ensure that you have the following packages installed (the versions used for testing are shown in parentheses).

• numpy (1.26.0)
• scipy (1.11.3)
• pandas (2.1.2)
• pandas-datareader (0.10.0)
• wrds (3.1.6)
• python-duckdb (0.9.1)
• statsmodels (0.14.0)
• joblib (1.3.2)
• pandarallel (1.6.5)
• numba (0.58.1)

0_b. Clone or downlaod the repo

You can obtain the code in two ways:

Option 1: Clone the repository using the following command:

git clone https://github.com/mk0417/anomalies-replication-python.git

Option 2: Download the ZIP file by following these steps:

• Click on the "Code" button at the top right with a green background.
• Select "Download ZIP."
• Unzip the downloaded file.

0_c. WRDS access

Important

Data is extracted from WRDS cloud and you need to change the WRDS username

• Open the file of functions.py under code directiry
• Move to line 33
• Replace 'change to your wrds username' with your WRDS username (keep the quotation mark)

1. Download data

To download the necessary datasets, navigate to the data_download folder (under the folder of code) in your terminal and execute the following command:

sh run_download.sh

Important

• A prompt will appear and you need to type your WRDS credentials
• Answer y in the prompt to create .pgpass file

This will download all required datasets. A folder named data_download (under the folder of data) will be automatically created, and all raw data will be saved in this directory.

2. Generate firm characteristics

To generate firm characteristics, navigate to the predictors folder (under the folder of code) in your terminal and run the following command:

sh run_predictors.sh

This process will create a folder named predictors under the data folder, and it will store firm characteristics in this directory.

Convert parquet to csv (optional)

To convert parquet data to csv files, navigate to the folder of code in your terminal and run the following command:

python convert_to_csv.py

A folder named csv will be created under data/predictors, and all csv files are stored in the directory.

If you just need firm characteristics, then you can ignore step 3.

3. Portfolio returns

For portfolio returns, navigate to the portfolios folder (under the folder of code) in your terminal and run the following command:

sh run_port_ret.sh

This will compute portfolio returns, including long-minus-short returns. A folder named portfolios under the data directory will be automatically generated, and the returns will be stored there. Long-short returns are determined based on the characteristic-return relation. For instance, if asset growth and return are negatively associated, the long-short return for decile portfolios will be calculated as port1 minus port10.

There are various methods available for calculating portfolio returns, such as:

• ew_all_port5: equal-weighted portfolios & break points based on all stocks & quintile portfolios
• vw_nyse_port10_price5: value-weighted portfolios & break points based on NYSE stocks & decile portfolios & keep stocks if price>=5
• nodlst_ew_all_port10_price5_ex_fin: returns are not delisting adjusted & equal-weighted & break points based on all stocks & decile portfolios & keep stocks if price>=5 & exclude financial industry

Credit

Thanks Chen and Zimmermann for their contributions to open-source asset pricing research.