How the future of equity research is increasingly depending on AI
Equity research has been one of the bedrocks of the finance industry. It uses a fundamental approach which is bottom-up in nature and involves the scrutiny of financial statements, annual reports and company press releases. Due to the ease and simplicity of the process, it is one of the first industry techniques taught to business school and management students. Fundamental analyses involves little complicated math and as a result appeals to a multitude of individuals. Moreover, it brings along the “art” in investing which relies also on intuition and clairvoyance than simply arriving at decisions using numerical information.
by Mrinal Mishra.
Often it is not possible for asset managers or investors to track a multitude of stocks or asset classes and as a result they rely on “research” teams within banks or research boutiques. These firms keep one abreast of the latest developments by publishing frequent research reports. Traditionally, research was bundled into a package and offered en-suite along with other services. However, MiFID II requires brokerages to either ring fencing research expenses or raise fees to account for them. In a nutshell, since January 2018 asset managers started to pay for any research reports they want to access from the buy-side firms.
This calls for asset managers to develop in-house research capabilities. Most buy side companies do have in-house research teams but each individual is expected to follow a variety of assets thus limiting the depth of research undertaken. AI can help solve this conundrum by “putting the science back” in the art of investing. Natural language processing (NLP) can help analysts filter the first level of information from annual reports, press releases and news websites. This will allow them to get a condensed and relevant view of the document without having to spend time poring over the details. Moreover, this also frees up time for private or (not easily available) public information acquisition which could be beneficial for arriving upon an investment decision. Hence, information extraction using NLP and AI can be suitably applied to perform such tasks.
Unsupervised learning techniques maybe be used for topic analysis by obtaining a broad list of topics that a paragraph or a document section talk about. This allows one to understand the content of the documents on a first-pass level. Traditionally, algorithms can do a better job of classifying topics from documents than human beings. Finally, one can go a step further and try to predict (hopefully with success) earnings per share or target prices. This can be achieved by supervised learning algorithms which can be trained using past target price information.
However, there remain some challenges to gainfully using such models for estimating financial outcomes. First, financial vocabulary is extremely specific, tailored and high on jargon. For example simple words like stock, share and bond have different meanings in English language and finance. As a result trying to run models which are trained using text from Wikipedia or IMDB might not move the needle much in finance. Second, while there are pre-trained word to vector representations (Word2vec or GloVe) available for general use, they are not very accurate when we employ them on annual reports or financial documents. Hence, a pre-trained word embedding using finance text (and/or documents) is very much the need of the hour.