Sentiment calculus for a method and system using social media for event-driven trading
It is often said that sentiment drives the stock market. How can a system that computes sentiment conveyed in social media chatter, evaluate publicly traded assets in real-time? What are the properties of such a system?
Sentiment calculus for a method and system using social media for event-driven trading was published on September 27, 2012. We would like to take the occasion and the date to describe the problem it addresses and the solution it provides.
Sentiment conveys human’s inclination, favorable, unfavorable or neutral, with respect to an object in a given event or situation, in the case at hand, publicly traded assets. This is the case in speech and conversations, in full fledged texts as well as in more often than not short social media chatters.
A core problem with the current methods in sentiment analysis is that no attention is provided to content of linguistic expressions. Statistical methods such as the “bag-of-words” and Machine Learning, including Deep Learning, typically disregard the syntactic relations between words. This is fatal for any attempt to extract accurate content from linguistic expressions, including sentiment, topics and events from texts.
The patent describes a solution to this problem. It uses a modular architecture and a deterministic sentiment computation. The computation is not based on sets of so called “content words” extracted form texts, which are then deprived from the syntax-semantic content, as it is the case in the “bag-of-words” approach. There is no training set required in the processing because the knowledge of language is built into the system, as it is as a part of humans’ genetic endowment.
Sentiment calculus for a method and system using social media for event-driven trading defines the properties of a systems that takes Social Media feeds and analyses each message and computes the sentiment of specific objects it includes, namely ticker symbols or any other market related objects of interest. The computation proceeds in real time, on the basis of the sentiment of its parts of each message in their specific syntactic arrangement.
The system encapsulates both the universal and domain dependent operations that generate and interpret human language. The syntax-semantic computation is universal as well as it is sensitive to domain dependent lexical features.
The universal computation relies on structure dependent operations, building hierarchical dependencies between the constituents of linguistic expressions. Implicit information is part of full-fledged texts as well as it is in short Social Media messages. The system recovers the content of unexpressed constituents, including verbal elements, e.g. $SPY very close to all time high. $VXX very close to all time low, as well as it recovers the antecedent of pronouns, e.g. $SPY just did the unthinkable. This Monday, it hit to 11-YEAR LOWS, thus enabling accurate sentiment computation.
The system is also domain dependent. For example, the lexical sentiment of the verb “sell” will be negative in the stock market domain, e.g. I sold my AAPL shares, but not other domains, e.g. I sold my Maserati. The knowledge base associated with the Inference Engine includes Named Entities recognized by the Inference Engine whose inference rules derive the relevant sentiment with respect to the specific Named Entities. For example, the knowledge base includes conceptual trees enabling conceptual dependencies between Named Entities, e.g. the stock symbol $AAPL, the Company name Apple and its products including the iPhone. Interestingly in the short text $AAPL has sold 1 Billion iPhones, the sentiment will be positive for $AAPL. How so? The inference system will interpret $APPL as standing for the Company Apple, the Agent of the selling of I Billion of its iPhone-product, and will yield a positive sentiment for $AAPL. Furthermore, the post-processing sentiment computation ensures the relevance of the sentiment per asset by identifying the pragmatic bias of the message. For example contrasting sentiment value will be computed to competing ticker symbols, company names, or products irrespective of whether or not the text include negative terms.
Sentiment calculus for a method and system using social media for event-driven trading describes an innovation that turns statistical and Machine Learning approaches to sentiment analysis around. It provides a method and system for computing sentiment conveyed by Social Media short messages on trading assets in real time. It does so deterministically, with an accuracy rate comparable to humans’ ability to interpret short texts. and with the capacity to process high volume and speed of supercomputers. Crucially, the natural language computation delivers content: something refreshing in today’s world of AI!