Using technology to navigate a new era of data-driven investment
Advances in artificial intelligence and machine learning technologies have unearthed a treasure trove of opportunities for finance practitioners keen on leveraging digital data to maintain their edge
The emergence of the digital age since the early 2000s has given rise to an exponential surge in the volume of digital data, a phenomenon that has rewritten the playbook for finding investment opportunities. In the modern data-centric world, finance practitioners can access a multitude of data sources – from traditional ones such as market prices, financial statements and analysts’ estimates – to alternative datasets such as crowdsourced information, textual, location and transactional data, among others.
Unsurprisingly, the domain of unstructured data – which includes text documents, such as news, financial reports, earnings call transcripts and social media content – makes up an estimated 80 per cent of total data and has posed a significant challenge due to its complexity and sheer volume.
Opportunities emerge
Yet, where there are challenges, there are opportunities. With the advent of artificial intelligence (AI) and machine learning (ML) technologies, we have seen these hurdles converted into a treasure trove of opportunities for finance practitioners keen on leveraging this data to maintain their edge.
As data grows in breadth and depth, the need for advanced data processing capabilities has become increasingly paramount. This necessity has triggered a wave of research and innovation in AI, with ML techniques emerging as a powerful tool for sifting through these vast and diverse datasets. The development of high-performance computing, particularly the adaptation of graphics processing units (GPUs) for general-purpose calculations, has been instrumental in accelerating the training and applications of ML algorithms across various sectors.
A significant catalyst in the ML revolution has been the rise of open-source software, which has democratised access to sophisticated techniques. Indeed, the bright outlook for AI applications has encouraged large tech corporations like Google, Facebook, Amazon and Microsoft to invest heavily in the field of ML. These names have committed resources to create development frameworks for their own use but have also been enthusiastic in releasing them as open-source to the research community.
Popular ML frameworks include PyTorch (released by Facebook in early 2017), TensorFlow (open-sourced by Google Brain in 2015) and MXNet (main backers being Amazon and Microsoft) and consist of a high-level interpreted language built upon a fast and low-level compiled backend to access computation devices including GPUs. These names have certainly contributed to develop a collaborative and active community, made up of researchers, practitioners and ML enthusiasts, and to the popularity of data science competition platforms such as Kaggle.
For finance practitioners and investment managers, these frameworks have opened up the possibility of fine-tuning deep learning models to serve specific tasks. They can be used to predict stock returns, risk-adjusted returns, and even to determine directly optimal portfolio weights which maximise portfolio risk-adjusted returns on a given timeframe.
Large language models emerge
The impact of technology on finance is most vivid with the emergence of large language models (LLMs), which have made tremendous progress in recent years and are now providing us with unparalleled tools to make sense of text information. The advent of attention-based networks – artificial neural networks able to highlight the relevant interactions within sequences – started in 2017 with the introduction of the BERT model by Google, and culminates today with very powerful models like OpenAI's GPT4, revolutionising our capabilities of treating text data. Recent models can attain and sometimes even exceed human competency at reading and annotating text, providing a technology that can massively accelerate and improve the process of integrating textual data from multiple sources into the investment process.
A recent widely advertised LLM in the financial context is BloombergGPT, trained with large bodies of text (hundreds of billions of private and public tokens), which probably already powers Bloomberg chat assistants and other tools on the platform.
Proprietary models, however, are not always available for fine-tuning, and today, we are witnessing a surge in open-source efforts aimed at facilitating the fine-tuning process for financial tasks. Among these are academic initiatives such as FinGPT, which simplifies the process by limiting the number of trainable parameters to a few millions, and tech company offerings like Meta's recently released LLaMA 2, a fully open-source model that provides remarkable performance with 10 times fewer parameters than GPT3. ClimateBERT is another interesting, fine-tuned model, which seeks to improve its understanding of companies’ references to climate change and their approach to greenhouse gas emissions.
The vast availability of such open-sourced models allows investment managers and finance players to build their own custom LLM model by fine-tuning pretrained ones to optimally extract information from a specific set of text data sources and carry out tasks needed by the financial practitioners. Whether to predict stock returns, screen out stocks with ESG controversies or identify thematic trends in the market, the applications are endless and offer the financial industry a fantastic potential for innovation.
The finance practitioners who skilfully harness these tools to navigate the vast expanse of data are likely to stand at the forefront of the investment industry.
Valentin Betrix is systematic equity fund manager at RAM Active Investments
