Feature Selection
There are many features that are used in forecasting.
Bollinger Bands: These utilize the rolling means of the closing price of a stock. Using the closing price averages over N previous days, the Upper (Lower) Bollinger Band is given by the + (-) 2 sigma standard deviation around the mean. The price of a stock within this window is used to determine when to buy or sell.
Momentum: The momentum of a stock is given by taking the price at the current day, and subtracting off some day in the past.
Relative Strength Index: This is an indicator that is dependent on the average amount of positive returns in an N day period, and the average negative return.
We use each of these traditional forecasting features, using different windows to track both long and short term trends. Windows range from 5 to 30 days for most features. We add a few additional features, outside of these three. We include the difference in high/low and close/open prices as a measure of volatility. We also include two variables to track the time of year, for industries that have seasonal variations. Finally we include a log price, as variations in the price may be larger/smaller with more/less expensive stocks, and there are a wide range of prices.
Finally, we generate our target variables. Our targets are the predicted rolling means, as the relative fraction of the current rolling means. This is done to reduce the noise, as day by day prediction is impossible for stock prices, but predicting trends is within the bounds of reason.