Note

• The code was tested on an Ubuntu 20.04 LTS machine running Python 3.8.10.

Setting Up the Python Env

• Install the requirements from requirements.txt using pip3 install -r requirements.txt. (Preferably in a virtual environment.)

Downloading Data

• Run the download_fillings.py for dowloading the data.
• Download and run mongodb.
• Make sure its running at 127.0.0.1:27017
• Then run the populate_db.py file to upload the cleaned data to the local mongodb service.

Documentation for each file

• download_filings.py : This for the downloading the forms for companies given in the PS. It uses a csv file consisting of the CIK for these companies.

• extract_cik.py : This was used to find the CIK for all the companies given in PS. We found the list of the companies similiar to the given name, then used an string matching algorithm (from difflib module in python 3) to find the closest name and then took CIK of that name. Except 10-15 names all of the companies and the CIK were easily and accurately found. We added the remaining CIK manually in the sheet.

• extract_ticker.py : This was used for mapping the CIK to the tickers.

• parse_10q.py: This file consist of function for parsing the 10Q forms.

• parse_form.py : This file consist of function for parsing the 10K forms.

• populate_db.py : This will read the downloaded files and upload the cleaned dataset to mongodb.

• For parsing the forms, we have used regular expressions to extract each section's text in form of paragraphs and corresponding tables in that section. They are stored in a JSON format and accessed as such.

Sentiment Analysis

• All ML models were run using hugging face's pipeline.
• For sentinment analysis, we have used finbert, a pre trained model using BERT on finanical text.
• It gives softmax score of three outputs - positive, negative, nuetral.

Summarization

Zero Shot Classification

Improvements

References

• https://github.com/jadchaar/sec-cik-mapper
• https://github.com/eabase/py-edgar
• https://github.com/jadchaar/sec-edgar-downloader
• https://airccj.org/CSCP/vol7/csit76615.pdf
• https://gitlab.com/juliensimon/huggingface-demos/-/blob/main/sec-filings/SEC_functions.ipynb
• https://huggingface.co/ProsusAI/finbert
• https://huggingface.co/facebook/bart-large-mnli

API ka documentation

• Link to the API : https://sec-dataset-downloader.herokuapp.com/create_dataset
• We were told to make an API to download dataset for any tickers for given forms and range of date.
• Parameters required for this API:

Example:

{
   "tickers" : ["AAPL", "MSFT"],
   "start" : "2021-01-01",
   "end": "2022-01-01",
   "forms" : ["10-K", "10-Q"]
}

• Method used: POST
• To use the API, send these parameters only, there is also an optional paramter called end which defaults to January 1st of 2022.
• Dataset consisting of only the forms filed between these date are returned.
• The dataset is returned in form of csv file.
• CuRL command for the same:

curl --location --request POST 'https://sec-dataset-downloader.herokuapp.com/create_dataset' \
--header 'Content-Type: application/json' \
--data-raw '{
   "tickers" : ["AAPL", "MSFT"],
   "start" : "2021-01-01",
   "forms" : ["10-K", "10-Q"]
}'

• NOTE : Downloading the dataset takes lot of space in backend, which might be too big for heroku. Kindly run on local system as it might be downloading GBs of data for processing depending on the time range, no of companies and no of filings.