What Language Do Stocks Speak

Stock prediction is a challenging and chaotic research area where many variables are included with their effects being complex to determine. Nevertheless, stock value prediction is still very appealing for researchers and investors since it might be profitable, yet the number of published research papers remains to be relatively small. The employment of advanced data analysis techniques has already been suggested by previous researches, such as the use of neural networks for stock price prediction, but practical implications of the majority of approaches are limited as they are concerned mainly with a prediction accuracy and less with the success in real trading with consideration of trading fees. We propose a novel approach for stock trend prediction that combines Japanese candlesticks (OHLC trading data) and neural network based group of models Word2Vec. Word2Vec is usually utilized to produce word embeddings in natural language processing tasks, while we adopt it for acquiring semantic context of words in candlesticks’ sequence, where clustered candlesticks represent stock’s words. The approach is employed for the extraction of useful information from large sets of OHLC trading data to improve prediction accuracy. In evaluation of our approach we define a trading strategy and compare our approach with other popular prediction models – Buy & Hold, MA and MACD. The evaluation results on Russell Top 50 index are encouraging – the proposed Word2Vec approach outperformed all compared models on a test set with a statistical significance.