Discovering Finance Keywords via Continuous-Space Language Models

Abstract

The growing amount of public financial data makes it increasingly important to learn how to discover valuable information for financial decision making. This article proposes an approach to discovering financial keywords from a large number of financial reports. In particular, we apply the continuous bag-of-words (CBOW) model, a well-known continuous-space language model, to the textual information in 10-K financial reports to discover new finance keywords. In order to capture word meanings to better locate financial terms, we also present a novel technique to incorporate syntactic information into the CBOW model. Experimental results on four prediction tasks using the discovered keywords demonstrate that our approach is effective for discovering predictability keywords for post-event volatility, stock volatility, abnormal trading volume, and excess return predictions. We also analyze the discovered keywords that attest to the ability of the proposed method to capture both syntactic and contextual information between words. This shows the success of this method when applied to the field of finance.

Convolutional Neural Networks for Financial Text Regression

“…However, concern can be replaced with pertain only if it does not have any sentiment polarity. It can be seen that expanding the lexicon using word embeddings, like previous works did (Tsai and Wang, 2014;Tsai et al, 2016;Rekabsaz et al, 2017), can be problematic and may end up with a lexicon expansion contain- ing semantically close but sentimentally far words. Another interesting word in the list is DMAA.…”

“…Stemming decreases the vocabulary size of the word embeddings and thus reduces the parameters of the model. Stemming is also required to use word vectors trained by Tsai et al (2016) since the corpora which is used to train the word embeddings consists of stemmed reports.…”

“…Instead of random initialization of the embedding layer of the model, initialization with pretrained word embeddings enables the model to capture contextual information faster and better. In our work, we used pretrained word embeddings supported by Tsai et al (2016). They used MD&A section of 10-K reports from 1996 to 2013 to train the word embeddings with a vector dimension of 200 by word2vec 6 continuous bagof-words (CBOW) (Mikolov et al, 2013).…”

“…• CNN-NTC-multichannel: Same as CNN-STC-multichannel but its transferred convolution layer is non-static Table 1 indicates that performance of our CNNsimple (baseline) model is comparable with EXP-SYN, the best model represented by Tsai et al (2016), which uses a manually created lexicon and POS tagger. Furthermore, the best predictions for the years 2008 and 2010 are achieved by the CNNsimple model.…”

Explaining Financial Uncertainty through Specialized Word Embeddings

“…They found that doing so both improves the performance of a Ranking Support Vector Machine (SVM rank ) as well as a Support Vector Regression (SVR) model with bag-of-word vectors as features and stock return volatility in the year after the filing date as label. Following up, Tsai et al [33] show that such an expanded dictionary can effectively be used to not only predict return volatility, but also post-event volatility (estimated with the Fama-French 3-factor model [7]) in the following year. Although the authors acknowledge that the regression on post-event volatility is sensitive with regard to the number of added candidates k [33, cf.…”