nbettencourt/SC454k

SC454k is a dataset of roughly 454k news articles and press releases related to small cap stocks from nasdaq.com and attached with market data from the Trade and Quote Millisecond dataset (TAQ/MSEC) from Wharton Research Data Services. Inspiration for this dataset was taken from FNSPID: A Comprehensive Financial News Dataset in Time Series.

Schema

• Symbol: the ticker symbol of respective stock (e.g., AAPL)
• Security: the full name of stock associated with the ticker symbol (e.g., Apple Inc.)
• Sector: broader economic sector the stock belongs to, sourced from nasdaq.com (e.g., Technology)
• Industry: more specific business segment within 'Sector', sourced from nasdaq.com (e.g., Computer Manufacturing)
• URL: URL associated with the 'Article' column
• Date: the date the 'Article' was published, in the format of MMM DD, YYYY HH:MM:SS (e.g., Aug 06, 2024 12:11:00)
• RelatedStocksList: stocks/topics that are related or appear in the 'Article' (excluding the respective stock mentioned in the 'Symbol' column). Delineated by '|'s (e.g., Markets|INTC|UNH|DOW)
• Article: text of the news article or press release. Any specially formatted text (bold text, tables, etc.) are treated in a pseudo-LaTeX format. For example, bold text are surrounded by '**'s and tables are wrapped in \begin{table}{...}\end{table}
• Title: title of the news article or press release
• articleType: value denoting the type of 'Article', either 'News' or 'Press Release'
• Publication: source responsible for publishing the 'Article' (e.g., The Motley Fool)
• Author: person or entity responsible for publishing the 'Article'. May appear as an entity similar to the 'Publication' (e.g., Publication: Zacks, Author: Zacks Equity Research) or as individual person (e.g., Publication: Validea, Author: John Reese)
• Risk_Free_Rate: annualized 1-month U.S. Treasury bill yield at the time of the article's publication - rate is sourced from Federal Reserve Economic Data (FRED), where values are expressed as decimals
• weighted_avg_x_hours: stock price 'x' hours away from 'Date' (calculated from a weighted average of the 8 closest stock prices to 'Date')

Scraping/Pairing Market Data

Scraping the non-pricing aspects of this dataset was conducted across 10 EC2 t2.large instances across two parts: scraping links, then the content of these links. Puppeteer, a headless Chrome browser in JavaScript, was used for lightweight, quick scraping. Data was subsequently cleaned with a combination of PySpark and NumPy. Pairing this with market data was made available through Wharton Research Data Services and their extensive collection of financial datasets. One collection of these datasets is the Trade and Quote Millisecond dataset (TAQ/MSEC), consisting of millisecond level access to the tick-by-tick trade and quote data of all activity within the U.S. National Market System. The process of pairing this pricing data with the timestamped news article/press release data through their API was magnitudes more tricky than meets the eye. The issue comes down to one thing: latency. The granularity of the API call can only be day-level, meaning a simple call to get the pricing data for say Apr. 3, 2020 11:13:37AM will return around 10-15 thousand records of data, and because multiple prices are being retrieved for each news article/press release (I'm grabbing the price of the stock 4 days before, 2 days before, all the way to 30 days after with more granularity in between, totaling 18 different prices across 10 unique days), this is a large amount of data to pair, even with highly-optimized C++ code. Thus, the pairing algorithm went through a number of iterations to optimize efficiency such as promoting a more indexed-based approach to finding dates, a stringent approach to clearing out unused memory, and customized bash scripts for easy deployment of multiple EC2 instances. Ultimately, totaling around 1400 hours of compute across 41 (yes, 41) m7g.medium instances. Any values where no market data was available is denoted with a null value. The number of rows with complete market data totals just above 440k.

Future Updates

I expect to update this dataset on a recurring basis, and will be doing the same for large cap stocks, as well. Stay tuned!

LinkedIn: https://www.linkedin.com/in/nicholas-bettencourt/
Email: nbetts@g.ucla.edu