choose a forecast approach

[GitTorres]

• (nys) read 5 articles which used sentiment analysis in trading strategy
• (dl only to start) dl+f or dl only?
• if f, choose targets
• hypothesize relevant features
• identify data adjustments needed

[GitTorres]

source 1:

• what it did

  • features
  • financial news + social media feeds -> sentiment scores
  • historical price data
  • models
  • Random Forest / SVM (targeting short-term stock price movement)
  • data
  • 5 largest companies from S&P index (Apple, Microsoft, Amazon, Google, Facebook)
  • NASDAQ financial news headlines
  • StockTwits social media

• what we will do differently

  • use pre-built sentiment score from Quandl in addition to StockTwits data
  • combine price action logic with sentiment scores in trading strategy

[GitTorres]

source 2:

what it did: compare machine learning stock price forecast to human stock price forecast

conclusion: not very relevant. They spent effort investigating human's ability to comprehend text information and comparing to machine-learning algorithms. Not interesting.

[GitTorres]

source 3:

what it did: use Log Periodic Power Law Singularity (LPPLS) model to study and diagnose price bubbles in the Chinese stock market

[GitTorres]

source 4:

what it did: investigative study on how useful GDELT information is in improving the accuracy of BTC price / stock index price prediction

target: daily stock price

sources:

• GDELT
  • Global Database of Events, Location, and Tone is the "largest, most comprehensive, open source database of human society"

conclusion: read later; do not use for current task

[GitTorres]

source 5:

what it did: build a model of the directional change of company's stock price using sentiment of consumers' reviews conclusion: not useful

[GitTorres]

source 6:

what it did: sentiment scores based on twitter data was used to forecast stock price conclusion: not useful

[GitTorres]

will use decision logic only - no machine learning at first

ml if time permits