AI Transformers Revolutionize Financial Text Analysis

The early approaches in the field of NLP captured broad structure but struggled to understand nuances. To move beyond simple word matching researchers had to win the fight against complexity (the “curse of dimensionality”¹), by representing words as points in a (vector) space of reasonably low dimensions – so called word embeddings. After Bengio et al. (2003)2 introduced neural networks to compute work embeddings models like Word2Vec significantly popularized and operationalized these embeddings. The striking observation was that this generation of models captures both semantic and syntactic relationships, making it one of the foundational techniques in modern NLP.

In our previous episode, we explored word embeddings like GloVe3 (which at the time was the biggest competitor of Word2Vec) and how the Systematic Equity team applies it to turn earnings call transcripts and 10-K filings into investment signals aimed at enhancing risk-adjusted returns. GloVe and other word embedding methods are powerful tools – they capture relationships between words and can often comprehend context surprisingly well, especially when applied to large, domainspecific datasets.

However, word embeddings still map each word on a single, fixed representation, regardless of the specific sentence or context in which it appears. After several failed attempts to overcome this limitation the 2017 paper “Attention is All You Need”⁴ eventually introduced transformer models, which dynamically interpret words based on their surrounding context. For example, in traditional word embedding models, the word “bank” would have the same meaning whether it appears in “river bank” or “bank account”. In contrast, transformer models can distinguish between these meanings by considering the entire sentence, understanding whether “bank” refers to the side of a river or a financial institution. Unlike static word embeddings, transformers read like humans – adaptively interpreting language based on context. Though more complex and resource-intensive, they offer a deeper and more flexible understanding of language.

At the heart of this innovation is a mechanism called self-attention, which lets the model weigh the importance of each word in a sentence, or even a document, relative to every other word. This means the model doesn’t just look at words one-by-one but considers how all the words in an article relate to each other to grasp the full meaning.

Imagine you’re describing a holiday: “I took my dog to the beach, and she loved playing with the waves.” A transformer model understands that “she” refers to “my dog” and that “playing with the waves” is something enjoyable happening on the beach. It pays attention to these connections across the sentences, rather than interpreting each word separately. This ability to connect words and ideas across large chunks of text or even full documents helps transformers understand language much more like a human would, making them powerful tools for tasks like translation, summarization, and sentiment analysis. Transformer technology is the basic building block of state-of-the-art LLMs, where the term “large” refers to the billions or even trillions of parameters they contain.

However, FinBERT also has some known limitations. For examples, it sometimes misclassifies courteous phrases like “Have a great day” or “Enjoy your meal” as positive sentiment, while texts like “no problem” as negative sentiment, even though these are neutral in financial contexts. Such errors can be mitigated through fine-tuning or by using more advanced prompting techniques with newer LLMs like GPT-4, which can better interpret subtle nuances⁷.

While the previous examples were based on short sentences, we now turn to a more scalable approach to sentiment analysis suitable for longer financial documents, namely net sentiments, that works for documents of any length. To compute net sentiment, a document is first divided into chunks. Sentiment is then evaluated for each chunk individually. The results are aggregated from a chunk level to the full document level by comparing the number of positively scored chunks to negatively scored ones, yielding an overall net sentiment score⁸

Based on this net sentiment score, we had another observation which may challenge FinBERT’s ability to accurately detect sentiment at the level of individual companies. While the model could identify negative sentiment during the COVID-19 pandemic (as depicted in the grey area in the chart below), it is important to note that some sectors consistently exhibit more positive sentiment than others. For example, communication services (yellow line in the chart below) and information technology (orange line) always exhibit more positive sentiment than Materials (purple line). This sector bias may require additional adjustments in the later stage of the investment process, particularly for investors who prefer to avoid sector-specific exposures because they do not believe taking such risks is compensated over the long term.

Relying solely on off-the-shelf models comes with additional limitations, which can be grouped into two main issues:

• Issue 1:In the early days, most NLP research was conducted within academic institutions, but over time, large corporations have taken the lead. This shift raises concerns around compliance, data ownership, data privacy, and copyright – critical considerations in the highly regulated financial industry.
• Issue 2:Additionally, these models continue to grow in size and complexity. However, as highlighted in the Chinchilla paper⁹

Consequently, off-the-shelf models may not be the best choice to meet our unique requirements. To maintain control over data governance and optimize model performance in the financial context, we must conduct our own research and develop customized transformer models tailored to our strategy.

To address this, we have customized the FinBERT model specifically for our use cases, ensuring it was better aligned with the unique requirements of our investment process. We then conducted a comprehensive simulation using sentiment signals generated by our tailored FinBERT model. The result of our backtest was encouraging: the simulated strategy achieved an information ratio of approximately 0.6 (before cost) against the S&P 500 index (as shown below). This demonstrates that machine evaluated sentiment, when properly calibrated and integrated into a disciplined investment framework, can contribute meaningfully to portfolio performance.

While these results are promising, we expect more to come. Thus far, this article has focused solely on sentiment. However, the potential of LLMs lies well beyond sentiment analysis. These models are capable of more sophisticated tasks, such as understanding casual relationships, or even reasoning. As LLMs continue to evolve, we see exciting opportunities to harness their capabilities for deeper insights and more sophisticated analysis.

Based on the promising results from our sentiment-based strategy, we went a step further and asked ourselves: Could this technology eventually replace a financial analyst?

A core function of fundamental analysis is building valuation models using data from balance sheets and cash flow statements. Traditionally this requires skilled analysts to interpret and contextualize company fundamentals. To test whether LLMs could assist in this domain, we leveraged OpenAI’s GPT-4 model via the OpenAI API, processing reports of thousands of companies in parallel.

Our experiments showed that LLMs can indeed extract key figures, compute ratios, and summarize broad trends with reliability. This aligns with research like the Chicago Booth paper¹⁰, which demonstrated that LLMs can ingest cash flow statements and balance sheets to evaluate future earnings improvement.

However, recent studies underscore the limitations of current reasoning models. Shojaee et al. (2025)¹¹ systematically examined the abilities of large reasoning models and found they face a collapse in accuracy and consistency beyond certain levels of complexity. While LLMs excel at broad pattern recognition and can generate convincing reasoning and analytic chains for simple tasks, their computational reasoning deteriorates for nuanced and high-complexity problems, such as those frequently encountered in comprehensive equity analysis. The study further shows these models often fail to employ explicit algorithms and may produce inconsistent explanations when faced with complex scenarios.

This suggests that, although LLMs can reliably extract figures, identify trends, compute ratios, and provide analytical starting points, they are not yet able to fully replicate the depth, logic, and rigor of a skilled human analyst, especially when the financial situation demands nuanced judgement or the construction of bespoke investment rationales. In practice, LLMs should be viewed as a supportive tool but their output always requires careful review and interpretation by experienced investment professionals. Therefore, in our view, the human analyst’s expertise, critical reasoning, and intuition remain indispensable in translating computational analysis into highest quality investment decisions.

Par conséquent, selon nous, l’expertise, le raisonnement critique et l’intuition de l’analyste humain restent indispensables pour traduire l’analyse computationnelle en décisions d’investissement de la plus haute qualité.

In summary, this paper has shown how the evolution of natural language processing – from early word embeddings like GloVe to advanced transformer models such as FinBERT – is reshaping the landscape of quantitative finance. By moving beyond static, context-free approaches to models that understand language in its context, we have been able to extract actionable insights from unstructured financial data.

Our experience demonstrates that off-the-shelf models, while powerful, often fall short in meeting the specific demands of financial analysis, particularly when it comes to sector bias, compliance, and data privacy. By developing and tailoring our own models, we have meaningfully improved our sentiment analysis, as evidenced by the simulation that outperformed the S&P 500 on a sector-neutral basis.

As these technologies continue to advance, our focus will remain on refining our models, integrating new capabilities, and responsibly applying AI-driven insights to our Systematic Equity strategies like Best Styles and Powered by AI strategies. Ultimately, the combination of domain-specific AI and human expertise positions us well to decode financial narratives more effectively and to uncover new sources of value in ever-evolving markets.