FiScrape

Financial news scraping and processing for machine learning.

FiScrape is a package for scraping financial news and applying natural langage processing for sentiment analysis. It leverages scrapy.

A search term can be queried across numerous online news sources. The resulting scraped headlines and standfirsts (and optionally the full text), along with metadata, will be timestamped and stored for each publication, along with the author and other relevent data.

It can be scheduled as a cronjob to run periodically. The sentiment scores are calculated and an email notificaion is sent for any new article containing a specified keyword.

Each news source on a particular topic can be stored as a seperate feature for timeseries analysis. e.g. For Bitcoin, all FT articles with "Bitcoin" in the title and/or text can comprise one set of sentiment feature. WSJ articles could be another set feature. This facilitates sentiment analysis on each publication individually, to see which, if any, yield the highest Information Coefficient or other measures of predictive merit.

Additionally, features may be further segmented by author.

Generated features can be stored locally as CSV or in a database using SQLAlchemy. The scraper can be set to run periodically.

The repository can be found at: Github-FiScrape

To install from git:

pip install git+git://github.com/PWE-Capital/FiScrape.git or git clone https://github.com/PWE-Capital/FiScrape.git

Dependencies

FiScrape requires Docker, Splash and this fork of Aquarium to scrape some websites that render in Javascript.

1. After pip installing FiScrape, download Docker at the above link.
2. As per the above Splash installation docs, pull the splash image with:

Linux:

$ sudo docker pull scrapinghub/splash

OS X / Windows:

$ docker pull scrapinghub/splash

3. Start the container:

Linux:

$ sudo docker run -it -p 8050:8050 --rm scrapinghub/splash

(Splash is now available at 0.0.0.0 at port 8050 (http))

OS X / Windows:

$ docker run -it -p 8050:8050 --rm scrapinghub/splash

(Splash is available at 0.0.0.0 address at port 8050 (http))

• Alternatively, use the Docker desktop app. Splash is found under the 'images' tab. Hover over it, click 'run'. In additional settings, name the container 'splash', and select a port such as 8050. Click 'run' and switch on the container before running scrapy. Switch it off after.

• In a browser, enter localhost:8050 (or whatever port you choose), and you should see Splash is working.

• The other dependencies will be automatically installed and you can run FiScrape as normal. $ sudo docker pull scrapinghub/splash for Linux or $ docker pull scrapinghub/splash for OS X.

3. Aquarium creates multiple Splash instances behind a HAProxy, in order to load balance parallel scrapy requests to a splash docker cluster. The instances collaborate to render a specific website. It may be necessary for preventing 504 errors (timeout) on some sites. It also speeds up the scraping of Javascript pages, and can also facilitate Tor proxies. To install Aquarium, navigate to your home directory and run the command:

cookiecutter gh:Saran33/aquarium

Choose default settings or whatever suits, splash_version: latest, set user and password, set Tor to 0.

4. a. To start the container (without Acquarium):

Linux:

$ sudo docker run -it --restart always -p 8050:8050 scrapinghub/splash (Linux) (Splash is now available at 0.0.0.0 at port 8050 (http).)

OS X / Windows:

or $ docker run -it --restart always -p 8050:8050 scrapinghub/splash (OS X) (Splash is available at 0.0.0.0 address at port 8050 (http).)

• Alternatively, use the Docker desktop app. Splash is found in the 'images' tab. Hover over it, click 'run'. In additional settings, name the container 'splash', and select a port such as 8050. Click 'run.'
• In a broweser, enter localhost:8050 (or whatever port you choose) and you should see Splash.
• The other dependencies will be automatically be installed and you can run FiScrape as normal.

4. b. Or to start the Splash cluster with Aquarium:

Go to the new acquarium folder and start the Splash cluster:

cd ./aquarium
docker-compose up

In a browser window, visit the below link to view Splash is working: http://localhost:8050/ To see the stats of the cluster: http://localhost:8036/

To run FiScrape:

To scrape all sites:

1. Navigate to the outer directory of FiScrape.
2. Open a terminal and run:

python3 fiscrape.py 

To run FiScrape GUI:

python3 fiscrape_gui.py

3. You will be prompted to enter a search term and the earliest publish date from which to scrape.

To scrape a specific site:

1. Navigate to the outer directory of FiScrape.
2. Open a terminal and run, for example, for the Financial Times:

scrapy crawl ft

3. You will be prompted to enter a search term. e.g. Bitcoin. You will then be prompted to enter the earliest date from which to scrape. e.g 2021-01-15. You can also enter t for today, yd for yesterday, w for the past week, m for month, or y for the start of the current year. For the matching articles returned from the query, each headline, standfirst (referring to the snippet on the site's search page), summary, image caption, content body, footnote, author name, author position, author bio, author email, author twitter, are returned to a database.
4. The default settings save the articles to a local SQLite database (which can be changed in settings.py). The DB can be read via SQL queries such as:

sqlite3 FiScrape.db
.tables
.schema article
.schema sentiment
select * from article limit 3;
.quit

Alternatively, the DB can be opened in the convenient DB Browser for SQLite.

To save the scraped data to a CSV as well as the DB, run:

scrapy crawl ft -o output.csv -t csv

To scrape the full HTML of a page:

Perform this test step in order to inspect a list of URLs, to see if they render in Javascript, or to simply scrape the raw HTML. Save a csv file with a list of URLs. The column should be called 'URL' or 'url.'

Open a shell in the top level FiScrape dir. Run the following command and it will prompt you for the url path. Enter d to read the default CSV file. The default CSV file can be changed in the "FiScrape/spiders/test_scr.py" file. Enter t to read the default test CSV file. The test CSV file can be changed in the "FiScrape/spiders/test_scr.py" file.

scrapy crawl test

or if you don't want to see the log:

scrapy crawl test --nolog

Sentiment Analysis

Additionally, FiScrape will return sentiment scores for each article's combined headline and standfirst, as well as seperate sentiment scores for the article's body. These sentiment scores will be stored in the database. The default sentiment analysis pipeline calculates the following sentiment scores:

TextBlob (TextBlob is a Python Library based on the Natural Language ToolKit (NLTK)):

• subjectivity (A float, ranging between 0 and 1. 0 is objective and 1 is extremely subjective)
• polarity (A float, ranging between -1 and 1. -1 is extremely negative sentiment and 1 is extremely positive sentiment.

VADER (Valence Aware Dictionary and sEntiment Reasoning lexicon-based sentiment scores):

• compound (A composite score, calculated by summing the valence scores of each word in the lexicon, adjusted accordingly, then normalized between -1 and 1. -1 is extremely negative sentiment and 1 is extremely positive sentiment.
• negative
• neutral
• positive The negative,neutral and positive scores are ratios describing the proportion of the text that falls into each category (so they should sum to 1, or close to it with floats).

Scheduling

• FiScrape can be scheduled to run periodically, as a cron job in the cron table, either on a remote Linux server or on a local machine (Linux or MacOS).
• In order to schedule the scraper:

Add conda to shell

(if using a conda env to run FiScrape)

• Activate anaconda to shell: echo $SHELL conda init <SHELL_NAME> e.g. conda init zsh or conda init bash

Add your environment to the fiscrape.sh script

• Find base conda path: conda info | grep -i 'base environment'

• Paste it into the script as the source, followed by /etc/profile.d/conda.sh

• e.g. source CONDA_PATH/etc/profile.d/conda.sh

• Add the name of your virtualenv or conda environment to the script: e.g. conda activate pweenv for conda or source pweenv/bin/activate for virutalenv.

Make the script executable

chmod +x fiscrape.sh

Turn on Splash Cluster

• In a seperate shell, wherever your aquarium folder is, in aquarium folder, run: cd ./aquarium docker-compose up

Test the bash script

./fiscrape.sh

Add the task to crontab

• Open crontab in nano text editor: EDITOR=nano crontab -e
• Copy the last line of the crontab_entry.sh file into the nano editor (or create a similar one based on the examples in crontab_entry.sh).
• Change the desirred frequency or fixed date/time for crawling.
• Crontab timing format examples: https://crontab.guru/examples.html
• Set the path to your local FiScrape outer directory:

Add the fiscrape.sh file path to the script:

• In the main FiScrape directory, open a shell and enter pwd. Copy the path.
• Add it to FISCRAPE_PATH in the fiscrape.sh file.
• Save the file (in nano, CTRL+X and then Y).

To check existing cronjobs:

crontab -l

Check if it ran:

If it ran, it will output zlog.txt to the FiScrape dir, and FiScrape.db will be saved to the sqllite_files dir (if none exists already). Alternatively, check with commands: whereis syslog grep fiscrape.sh” /var/log/cron tail -f /usr/bin/syslog | grep CRON

Email Notifications

Email notifications will be sent every time a new article is published, containing the selected keyword. To monitor multiple keywords, set up crontab entries for each keyword by copying the above steps. Each new article will be sent in a sperate email alert, with the headline as the Subject line, the body containing the keyword, article source, headline, snippet/standfirst, article link, and sentiment scores.

• If the body of the article was accessible (not paywalled for FiScrape), the sentiment scores are calculated for the headline and body. If not, the sentiment scores are based on the headline and snippet/standfirst, and the email will mention this caveat benath the scores. e.g.

To set up emails:

• Create a new Gmail or other email account. For Gmail:
• Allow less secure apps: https://myaccount.google.com/security
• Allow the gmail account: https://accounts.google.com/b/0/displayunlockcaptcha

Configure Email Settings

• In FiScrape/FiScrape/settings.py, change the email settings at the bottom of the file.

MAIL_FROM = '[email protected]'  # the 'from address'
MAIL_HOST = 'smtp.gmail.com'  # the SMTP server of your email provider
MAIL_PORT = 465  # post 465 is used for Gmail SSL, posrt 587 is used for Gmail TLS
MAIL_USER = '[email protected]'  # username for the 'from address'
MAIL_PASS = 'yourSecretpassword'
MAIL_TLS = True  # set to True if using TLS
MAIL_SSL = True # set to True if using SSL
mail_list = ["[email protected]", "[email protected]"]  # The alert recipients, sperated by commas
cc_list = ["[email protected]", "[email protected]"] # The alert CC recipients, sperated by commas ( Set to [''] if no one is to be cc'd)

• Emails will be sent whenever FiScrape is run, if any new articles are found.

ML Pipeline Example

In the NLP_Trading_Example.ipynb notebook, there is an example of how to use either the Python libraries Pandas and SQLAlchemy to retrrieve the data. Pandas can accept SQL queries. The example function creates a temporary DB table of sentiment scores, for a given publication. It then uses the temp table to form a Pandas DataFrame. That function can be called recursively for each source/publication in the DB, to generate individual sentiment features for each source.

The example aggregates the data into hourly candles, for the purposes of illustration. That could be useful for generating lagged or resampled sentiment features, although that may not be necessay if using a model such as a LSTM neural network.

The notebook includes some trend indicators such as a Kalman filter, and volatility features such as Bollinger bands and a Yang-Zhang estimator. Numerous Python libraries need to be pip installed to run this file. Installing Github-PWE_Analysis should satisfy the dependencies. You will also need to install the C++ TA-Lib

The data is then passed to a preprocessing pipeline. The categorical featues are converted to one-hot encodings and the rest of the features are scaled.

The example model is a basic Linear Discriminat Analysis classification model. This of course doesn't capture the sequential importance of features. The LDA is useful for dimensionality reduction though, so its outputs could be passed to a neural network.

Finally, the example contains a Backtrader backtesting environment, as well as Pyfolio analysis.