This project was created by @thomas211738 and @joshleeds
We wanted to see if technical analysis indicators were good predictors of a stock's future directionality. Since there are many technical analysis indicators, we decided to use as many as possible and apply them as features to a machine learning model to identify the best combination of technical analysis indicators that result in the best results. We did this for about 250 stocks to test how well the models worked for different stocks and avoid overfitting our model.
For this project, we used the following technical analysis indicators as features:
- Simple Moving Average (SMA)
- Exponential Moving Average (EMA)
- Relative Strength Index (RSI)
- Moving Average Convergence Divergence (MACD)
- Bollinger Bands
- Stochastic Oscillator
- Ichimoku Cloud Signals
- Aroon Indicator
- Williams %R
- On-Balance-Volume (OBV)
This is what one year's worth of data for our data set would look like.
As we can see, for each of the technical analysis indicators, they give a value of either a -1, meaning that particular indicator believes the stock will go down, 0 meaning it believes it will stay relatively the same, or a 1, meaning the stock will go up. We also have a Target column on the right, which indicates the stock directionality for the next day, with 0 meaning the stock price went down, and 1 meaning the stock price went up.
For the machine learning model, we used an LSTM neural network using keras in tensor flow. We trained 250 different models for the 250 companies that we chose. We first acquired 20 years' worth of stock data for each company and split it with 80% (16y) of the data to be trained on, and 20% (4y) of the data to be tested on. We also used the past 100 days of our data to predict the stock directionality one day in the future. We did this because we thought it would give us more accurate and better results since the model would have more data to analyze and work with. for our x data, we gave the model the past 100 days of our technical analysis indicator's values, and for the y data, we gave the model the stock directionality value for the next day. We also had an EPOCH of 10, meaning the model passed over the 16 years' worth of stock data 10 times.
To test our model's accuracy, we used the last 20% (4y) of our data. We needed to find out when our model predicted the stock going up, and then compare those instances with whether the stock actually went up or not. To do this, we used the precision_score function from sklearn, and found that we had an average precision score of 56.77% for the 250 models. This means that for every time the model predicted the stock to go up, it would go up 56.77% of the time.
To determine the profits or ROI someone would gain from applying this model to actual trading, we started by investing $1 in each company. Then, each time the model predicted the price going up, we would invest this $1 for the day. We would then keep updating this investment through the 4 years' worth of data to determine what our final investment is worth and ROI.
We found that the average ROI was 169% over the past 4 years. Our Highest ROI was for Apple, which had an ROI of 450% over the past 4 years. Our lowest ROI was for Coca-Cola, which had an ROI of 96%. Within that same 4 year period, the S&P 500 had an ROI of 45%. This means we had an average α of 124%.
While the ROI for Apple was 450%, this was because we were originally using Apple stock data when we were creating our model and so the model for it is quite overfit. In the future, we could try adding more features for the model to work with rather than removing ones that make the model less accurate. Another thing we can change is instead of the technical analysis indicators being restricted to values of -1, 0, and 1, we can make them have their actual values to let the model evaluate them better. Lastly, one thing we didn't calculate in this project was the Sharpe ratio to measure the performance of our investment in comparison to its risk. We can look into this in the future to measure the excess return of our investment per unit of risk.
In conclusion, this project helped us identify that technical analysis indicators play a big factor in determining future stock directionality. While many improvements could be made to this project, we were still able to learn a lot about machine learning, finance, and the intersection between them. We are very interested in the field of quantitative finance and want to pursue it more in the future.
This project is licensed under the MIT License.