Stock Price Prediction with News Articles

1
Stock Price Prediction with News Articles

• Qicheng Ma
• CS224n Final Project

2
What is built

• Build classifiers to predict News Articles
  Stock Up/Down/Flat on one day
• Actually build paragraph mentioning Stock
  /-/0
• Tried Naïve Bayes and Max Ent classifier.
• Using WSJ/ReuterFF news 1994-1996, daily quotes
  from Y!Finance
• Trading based on Up/Down signals, fixed-value
  daily long/short weighted by Up/Down signal.
• More paragraphs with /- signal for one stock
  trade stock more
• distribute fixed investment amount for one day
  proportional to sum(signals)
• Compare to baseline buy-and-hold strategy, 5 vs
  2 monthly
• Based on last years project (Timmons and Lee)
  and other research papers
• some subtle differences, tried different things.

3
Results

• Training on 29 months data, test on 1 month.
  Repeat.

4
Results monthly returns

• NB not as good as Baseline (same return, higher
  volatility)
• MaxEnt 5.2 vs baseline 2.6, statistically
  significant at 95 confidence
• Precision more important than Recall
• want most signals to be profitable, less
  important to discover all profitable
  opportunities with risk of false positive.

5
Future Work

• Build a generic model first (e.g. STOCK earning
  exceeds expectation), then train stock-specific
  on features
• Predict real-value returns, instead of discrete
  classes
• Look at excess return (stock minus market),
  adjust for beta, inflation, etc.
• Use more frequent data. Real time stock quotes,
  news article streams (search engine quotes and
  feeds)
• Profit!