STOCKGRAM : DEEP LEARNING MODEL FOR DIGITIZING FINANCIAL COMMUNICATIONS VIA NATURAL LANGUAGE GENERATION

Abstract

This paper proposes a deep learning model, StockGram, to automate financial communications via natural language generation. StockGram is a seq2seq model that generates short and coherent versions of financial news reports based on the client's point of interest from numerous pools of verified resources. The proposed model is developed to mitigate the pain points of advisors who invest numerous hours while scanning through these news reports manually. StockGram leverages bi-directional LSTM cells that allows a recurrent system to make its prediction based on both past and future word sequences and hence predicts the next word in the sequence more precisely. The proposed model utilizes custom word-embeddings, GloVe, which incorporates global statistics to generate vector representations of news articles in an unsupervised manner and allows the model to converge faster. StockGram is evaluated based on the semantic closeness of the generated report to the provided prime words.