MC3-Project-2

Updates / FAQs

2015-11-21

• Updated list of charts to generate.
• Q: In a previous project there was a constraint of holding a single position until exit. Does that apply to this project? Yes, hold one position til exit.

2015-11-22

• Q: Is that 5 calendar days, or 5 trading days (i.e., days when SPY was traded)? A: Always use trading days.

2015-11-23

• Q: Are there constraints for Python modules allowed for this project? Can we experiment with modules for optimization or technical analysis and cite or are we expected to write everything from scratch for this project as well? A: You can use whatever modules you like as long as you cite them. You've already written most of what you need though.

Overview

In this project you will transform your regression learner into a stock trading strategy. Overall, you should follow these steps:

• Train a regression learner (KNN or LinReg, or other of your choice with or without bagging) on data from 2008 to 2009. This is your in sample training data.
  • For your X values: Identify and implement at least 3 technical features that you believe may be predictive of future return. You should implement them so they output values typically ranging from -1.0 to 1.0. This will help avoid the situation where one feature overwhelms the results. See a few formulae below.
  • For your Y values: Don't use price, use future 5 day return. Remember you're trying to PREDICT the future.
• Create a plot that illustrates your training Y values in one color, price in another color and your model's PREDICTED Y in a third color. With this chart we should be able to see how well your learner performs and that your Y values are shifted back 5 days. You may find it convenient to zoom in on a particular time period so this is evident.
• Create a trading policy based on what your learner predicts for future return. As an example you might choose to buy when the forecaster predicts the price will go up more than 1%, then hold for 5 days.
• Create a plot that illustrates entry and exits as vertical lines on a price chart for the in sample period 2008-2009. Show long entries as green lines, short entries as red lines and exits as black lines. You may find it convenient to zoom in on a particular time period so this is evident.
• Now use your code to generate orders and run those orders through your market simulator. Create a chart of this backtest. It should do VERY well for the in sample period 2008-2009.
• Freeze your model based on the 2008-2009 data. Now test it for the year 2010 -- Plot that illustrates entry & exits, generate trades, run through simulator, chart the backtest.

Perform the above steps first using the data ML4T-399.csv, once you've validated success (it should work well), repeat using IBM data over the same dates. Remember 2008-2009 is training, 2010 is testing. You should have one set of charts for each symbol.

Summary of Plots To Create

1. Training Y/Price/Predicted Y: Create a plot that illustrates your training Y values in one color, price in another color and PREDICTED Y in a third color. With this chart we should be able to see how well your learner performs and that your Y values are shifted back 5 days. You may find it convenient to zoom in on a particular time period so this is evident.
2. Sine Data In Sample Entries/Exits: Create a plot that illustrates entry and exits as vertical lines on a price chart for the in sample period 2008-2009. Show long entries as green lines, short entries as red lines and exits as black lines. You may find it convenient to zoom in on a particular time period so this is evident.
3. Sine Data In Sample Backtest
4. Sine Data Out of Sample Entries/Exits: Freeze your model based on the 2008-2009 data. Now test it for the year 2010 -- Plot that illustrates entry & exits, generate trades,
5. Sine Data Out of Sample Backtest
6. IBM Data In Sample Entries/Exits: Create a plot that illustrates entry and exits as vertical lines on a price chart for the in sample period 2008-2009. Show long entries as green lines, short entries as red lines and exits as black lines. You may find it convenient to zoom in on a particular time period so this is evident.
7. IBM Data In Sample Backtest
8. IBM Data Out of Sample Entries/Exits: Freeze your model based on the 2008-2009 data. Now test it for the year 2010 -- Plot that illustrates entry & exits, generate trades,
9. IBM Data Out of Sample Backtest

Template and Data

You will use data in the ML4T/Data directory. In particular files named ML4T-399.csv, and IBM.csv.

Choosing Technical Features -- Your X Values

You should have already successfully coded the Bollinger Band feature. Here's a suggestion of how to normalize that feature so that it will typically provide values between -1.0 and 1.0:

bb_value[t] = (price[t] - SMA[t])/(2 * stdev[t])

Two other good features worth considering are momentum and volatility.

momentum[t] = (price[t]/price[t-N]) - 1

Volatility is just the stdev of daily returns.

Choosing Y

Your code should predict 5 day change in price. You need to build a new Y that reflects the 5 day change and aligns with the current date. Here's pseudo code for the calculation of Y

Y[t] = (price[t+5]/price[t]) - 1.0

If you select Y in this manner and use it for training, your learner will predict 5 day returns.

Contents of Report

• Your report should be no more than 8 pages long. Use 1" margins and no smaller than 10 point font. Your report should contain no more than 12 charts. Penalties will apply if you violate these constraints.
• Include the charts listed in the overview section above.
• Describe each of the indicators you have selected in enough detail that someone else could reproduce them in code.
• Describe your trading policy clearly.
• Discussion of results. Did it work well? Why? What would you do differently?

What to turn in

Turn your project in via t-square.

• Your report as report.pdf
• Your code as code.py

Extra credit up to 3%

Choose one or more of the following:

• Compare the performance of KNN and LinReg in this task. The instructor anticipates that LinReg might work well. If that turns out to be the case, how can that be? This is a non-linear task.
• Extend your code to create a "rolling" model that updates each day rolling forward.
• Extend your code to simultaneously forecast all the members of the S&P 500. Generate trades accordingly, and backtest the result.