It is quite common to observe that a lot of the initial application of machine learning techniques are on directly the market price (Open, High, Low, Close, Volume). As understanding of machine learning increase, the sophistication level starts to increase. Quants are starting to look for alpha by adding factors rather than on the market price directly.
In this tutorial, we will conduct an investigation into the fruitfulness of direct application of machine learning algorithms on market price alone.
A search on the web will uncover many examples of using classification based ML to predict the next day’s result. You can find some examples here. Typically, in the field of data science, one is happy with the accuracy of prediction. Indeed, the prediction accuracy of these ML (as Logistic Regression, Linear Discriminant Analysis (LDA), and Quadratic Discriminant Analysis(QDA)) when conducted on XMA (Australian 10-Year Treasury Bond) were 50%, 54%, and 63% respectively. Those are very impressive numbers. If you have a predictive accuracy level of 54% at a casino, you will be able to consistently make money from the house. The question we want to ask however is that, does having a high predictive accuracy level translate directly to returns and out-performance over simple buy-and-hold strategy?
Check out the backtesting tearsheets on Logistic and QDA for the answer. I did not include LDA backtest here. It is similar to QDA.
If you like what you read, do remember to subscribe.
What can we conclude from this? Direct application of ML to market data might have yielded fruitful results in the past. It is however no longer the case today.
I hope you have found this useful. If you would like to be kept informed. Do not forget to subscribe. Thank you.