New NLP model improves stock market predictions

For fiscal investors, uncovering ways to efficaciously foretell the behaviour of stocks and shares is captious if they privation their investments to execute well. There are online sources of accusation connected the factors that thrust banal marketplace movements, ranging from quality items to fiscal reports. But processing models that tin gully connected these assorted forms of earthy connection information to make close predictions isn't easy. In fact, for the earthy connection processing community, it's a large challenge.

A radical of researchers astatine the Research Center for Social Computing and Information Retrieval astatine China's Harbin Institute of Technology person constructed a model that tin synthesize these aggregate information sources and the assorted forms of information they contain. Study results, published successful the KeAi diary AI Open, amusement that their exemplary achieves a higher AUC (area nether the precision-recall curve) people than existing models.

As writer Kai Xiong explains: "Financial texts incorporate word-level, event-level, and sentence-level information. Simply utilizing a azygous operation of words, besides known arsenic a azygous semantic unit, isn't capable to stitchery each the accusation you request for an effectual prediction model."

According to co-author Xiao Ding, the Heterogeneous Graph-based Sequential Multi-Grained Information Aggregation Framework (HGM-GIF) they person developed tin code this problem.

"To get the word-level information, the fine-grained data, our model uses a stopwords list—in different words, a database of words that should beryllium filtered retired erstwhile processing the earthy connection data. To get the lawsuit information, the medium-grained data, we usage an existing openIE instrumentality to extract a bid of lawsuit triples, comprised of subject, verb and object, from fiscal text. While to get accusation from the sentences, the coarse-grained data, we divided the sentences recovered successful fiscal text."

Author Li Du picks up the story: "To exemplary the affluent connections betwixt those assorted sets of data, we usage heuristic rules to physique connections betwixt words, lawsuit triples and sentences. This results successful a caller heterogeneous graph neural web that models their interactions."

In their model, words sequentially interact with substance (event triples and sentences) for accusation selection, lawsuit triples interact with lawsuit triples for lawsuit narration understanding, sentences interact with lawsuit triples for discourse accusation supplement, and lawsuit triples interact with sentences for accusation selection. Author Ting Liu adds: "We past brace the results with accusation astir the peculiar corp to nutrient the last banal marketplace prediction."

The squad besides conducted studies successful which they removed antithetic kinds of accusation and graph neural web layers from the exemplary to analyse the impact. According to writer Bing Qin, these 'ablation' studies showed that words, lawsuit triples, and sentences are each important for accusation selection, portion each accusation aggregation furniture is important for last banal marketplace prediction.

---

---

More information: Kai Xiong et al, Heterogeneous graph cognition enhanced banal marketplace prediction, AI Open (2021). DOI: 10.1016/j.aiopen.2021.09.001

Provided by KeAi Communications