My project "Making Predictions in Extremely Volatile Markets" for CS229 Machine Learning at Stanford, Fall 2014.
- I predict the daily direction of movement in stock prices by utilizing classification machine learning techniques (SVM and logistic regression), training on 8 years of data and testing on the following 2 years.
- This technique has a prediction accuracy of 52-60%, depending on the stock.
- Even with this high error rate, it's easy to significantly boost risk-adjusted returns (measured by Sharpe ratio) by using the predictions in a simple trading strategy as compared to buy-and-hold. The trading simulation is quite lenient (no fees, slippage, etc) but the result still ain't bad!