76 lines
3.1 KiB
Markdown
76 lines
3.1 KiB
Markdown
This document is the final report for the machine learning for trading
|
|
course. I have implemented two manual strategies, a random tree
|
|
learner-based strategy and one based on Q-learning.
|
|
|
|
# Experiment 1
|
|
|
|
I have implemented two manual strategies. The first strategy buys on a
|
|
bullish MACD cross with a MACD smaller than zero and sells on a bearish
|
|
MACD cross with a MACD greater than one.
|
|
|
|
The second strategy uses MACD diff (the difference between the MACD and
|
|
the MACD signal), RSI, and price SMA with a period of eight. I have
|
|
plotted the metrics over their one, three, and five days return to find
|
|
reasonable thresholds for the strategy.
|
|
|
|
|
|
|
|
Based on the scatter plots, I have created a list of buy and sell
|
|
signals. Each signal uses the current number of shares owned and one of
|
|
the three indicators. The following figure shows the result for both
|
|
manual strategies compared to the benchmark. Both approaches do well in
|
|
the in-sample period but worse afterward, which I expected because I
|
|
cherry-picked the thresholds based on the in-sample period's scatter
|
|
plots.
|
|
|
|
|
|
|
|
Next, I have implemented a random tree-based strategy learner. The
|
|
learner uses a leaf size of five and no bagging. A smaller leaf size
|
|
would result in overfitting to the in-sample data. But as the following
|
|
screenshot shows, five works well, and the RT learner does well for the
|
|
out of sample data.
|
|
|
|
|
|
|
|
I have also implemented a strategy learner based on Q-learning. The
|
|
Q-learner uses fifteen training runs on the in-sample data. It mostly
|
|
does well for the out of sample data, but it looks like the RT-based
|
|
strategy learner is better.
|
|
|
|
I am using a bin-size of five for the three indicators mentioned before.
|
|
That results in 375 (3x5x5x5) states with only about 500 in-sample data
|
|
points. Probably the Q-learner is overfitting to the in-sample data.
|
|
Indeed, with bin sizes of four, the Q learner performs better for the
|
|
out-of-sample data.
|
|
|
|
|
|
|
|
# Experiment 2
|
|
|
|
Experiment 2 aims to show that the strategy learner trades differently
|
|
when there is a commission, and the impact is not zero. The RT-based
|
|
trader does not consider the commission value, but the Q-learning based
|
|
trader does.
|
|
|
|
However, it seems like a commission smaller than $10 does not affect
|
|
the number of trades significantly. Only when the commission is around
|
|
$50 or with a slippage of 1% we see considerably fewer transactions.
|
|
|
|
| commission | n_orders |
|
|
|------------|----------|
|
|
| 9.95 | 79 |
|
|
| 20 | 83 |
|
|
| 50 | 63 |
|
|
| 100 | 37 |
|
|
|
|
# Closing Remarks
|
|
|
|
Machine Learning for Trading is a great course. It gives an excellent
|
|
introduction to finance, trading, and machine learning without getting lost in
|
|
technical or mathematical details. I have enjoyed building decision tree
|
|
learners and a Q learner from first principles. At the same time, the course
|
|
accurately teaches powerful libraries such as NumPy and Pandas.
|
|
|