mxhofer/RiskyData-LDA

This project quantifies text-based risk disclosures as described in the Risky Data paper.

RiskyData-LDA Platform

Copyright (c) 2022 Maximilian W. Hofer & Kenneth A. Younge.

Repo maintainer: Maximilian W. Hofer (maximilian.hofer@epfl.ch)

  AUTHOR:  Maximilian W. Hofer  
  SOURCE:  https://github.com/mxhofer/OrgSim-RL  
  LICENSE: Access to this code is provided under an MIT License.  

The RiskyData-RL platform quantifies risk disclosures in IPO prospectuses using an LDA topic model.

Usage

Fork repo

Create your own copy of the repository to experiment with OrgSim-RL freely.

Install dependencies

git clone https://github.com/mxhofer/RiskyData-LDA.git

pip install -r requirements.txt

NB: You might need to add the geckodriver such that Bokeh can save images in PNG format:

conda install -c conda-forge firefox geckodriver

Prepare input data

We provide the Risky Data paper data in the data/ipo.csv file. You can replace this file with your own risk factor text as long as the column with text-based risk is split into paragraphs using the ---new_paragraph--- divider.

Run pipeline

To quantify textual risk factors, run the following scripts (in order):

1. 00-preprocess.py: cleans text data (tokenization, stemming, etc.)
   1. Input: Download IPO data and store it as data/ipo.csv (URL in 00-preprocess.py)
   2. Output: data/ipo_allText.pkl
2. 01-fit.py: fits an LDA topic model and writes paragraph-level topic loadings to disk
   1. Input: data/ipo_allText.pkl
   2. Output: data/ipo_allTopics.pkl
3. 02-normalize.py: normalizes topic loadings to firm-level risks using year and industry groups
   1. Input: data/ipo_allTopics.pkl
   2. Output: data/ipo_risk.xlsx

A note on performance: pre-processing text data and fitting the LDA topic model take time. On a MacBook Pro 2020, the pipeline takes ~2 hours to complete.

Inspect results

The results in data/ipo_risk.xlsx contain the aggregate risk disclosure and the individual risk factors for each IPO firm.