Abstract: The report presents a new model that tries to predict the direction of returns of an instrument traded on electronic markets by recognizing patterns in Japanese Candlestick data by using a Random Forest Classifier. Rather than trying to predict the absolute magnitude, the model is trained to predict whether the market is going to make a significant upward move, a significant downward move, or is going to stay neutral in the next few bars. Hence the model is trained to classify input data to one of three possible classes, i.e. Buy, Sell or Hold. This report also devises a probabilistic model for normalizing Japanese Candlestick data and labelling the training dataset into the three mentioned classes. Experimental results conducted on the model trained on actual data shows a win rate of above 50% for Buy and Sell calls, which might translate to positive excess returns for a trading strategy basedon this model.
It might be easier to view PDFs from the drive.
Drive: https://drive.google.com/drive/folders/13U4AZqYCml2MDaW_0AFt20e4zZgI7Bxt?usp=sharing