The future of sentiment: Leveraging BERT for financial market data

Text-based data has historically been a rich source of alpha for both discretionary and systematic traders. However, a trade-off has always existed between these two distinct users. Humans are inherently better at understanding text data, whilst machines are orders of magnitude quicker. In this insight, we outline the recent advances in natural language processing (NLP) and high-performance computing which have:

• Further enhanced the processing speed advantage enjoyed by systematic traders.
• Closed the gap between human and machine capabilities in natural language understanding

Deep learning models have revolutionised the field of Quantitative Finance and wider society over the recent years. Artificial, convolutional and recurrent neural networks have become mainstays in many high-performing funds for their ability to outperform traditional statistical or machine learning models. Taking a broader perspective, Large Language Models (LLMs) have become front and centre across a broad range of technical and non-technical fields. Generative models, with GPT leading the market, are being deployed at a record pace across a wide range of disciplines. Whilst generative LLMs have dominated news coverage for their novelty, the rate of progress of discriminative LLMs has also been formidable. Google’s BERT model has emerged as being the best-in-class transformer architecture for sentiment analysis. As outlined in figure 1, this model has been able to achieve up to double digit improvements over the previous state-of-the-art sentiment models in the General Language Understanding Evaluation (GLUE) test.

Figure 1: Scores for state-of-the-art sentiment analysers on the General Language Understanding Evaluation (GLUE) benchmark.

The progress that is presented by BERT models brings the opportunities for systematic traders to better leverage text data. Not only does the base model outperform the other models in the GLUE test, it has the added advantage of being fine-tuneable with relatively low amounts of data. This can further enhance their performance when applied to domains which have technical ‘jargon’ or non-standard use of language. In the following case study, we display the flexibility of BERT models when deployed to a variety of market domains.

The ability to process significant amounts of text is one of the key advantages that fully systematic traders rely on to extract alpha from text data. To maintain this edge, traders must ensure that they have easy access to high-performing models – or the ability to train them themselves. At the same time, it must be feasible to deploy these models in a live environment in a cost-efficient manner that still enables high throughput.

One of the key advantages of modern LLMs is the ease with which they can be accessed. LLMs are high-variance models that require enormous amounts of training data. It would be a huge challenge for any trader or company to gather, clean and test the amount of data that would be required to train an LLM. For example, the base BERT model discussed above has 110 million trainable parameters. However, many models are open-sourced. Subsequent model fine-tuning to make them more tailored to a given domain typically requires a significantly reduced amount of training data. As such, the barrier to adopting BERT models is extremely low.

Given that they can be accessed relatively easily, the remaining question is how feasible it is to incorporate them into ETL or data pipelines. Below is a summary of the runtime requirements for a single predict from one of Hugging Face’s FinBERT models. A single thread on CPU with 2.3GHz clock speed can, in the base case, process 20.16 pieces of text data per second. However, many optimisations can be made. The first is to switch from a standard tokeniser to a fast tokeniser. When working with models through Hugging Face, as is highly recommended when using Python, this is as simple as changing one line of code. Whilst the performance effects of this can vary, in the tests used this led to a roughly 74% increase in processing speed for the benchmark used here. Deploying the model on a GPU/TPU will yield the most significant speed boosts. In this instance, moving the model onto a 9.1 TFLOP GPU leads to per-second prediction capacity of 261, over a tenfold improvement from our baseline. Indeed, compute power far in excess of this is typically available through all modern cloud service providers.

Figure 2: The number of per second predictions that can be computed using the Hugging Face Python package and the base BERT model. Values are based on batched predictions: batch size = 256; 1,000 batches.

Finally, we display a case study on how to effectively use modern sentiment models for alpha extraction, focusing specifically on financial news. This includes an exploration of the logic required to go from raw news data to daily news signals. These signals rely on BERT models for sentiment analysis. As can be seen, using the BERT base model fails to capture the significant alpha that exists in the dataset. However, by switching to a FinBERT model, trained on analysts’ reports and other stock news, we can derive a high-performing strategy. This single difference in strategy logic and the resulting difference in performance displays the high potential for BERT models to be used across different domains. Assuming access to a relatively small amount of well-labelled training data, individuals or institutions can develop highly performant BERT models that are tailored to the specific task at hand. In this instance, the FinBERT model only required 10K training samples.

Figure 3:  Comparison of the cumulative log returns for two trading strategies between March 2019 and March 2022, where both strategies rely on daily sentiment signals derived from LSEG’s Financial News Service.

The gap between the two lines highlights the difference in performance that can be captured when using a model fine-tuned on domain-specific data (FinBERT) versus a more general model (base BERT).  Given the existing support for fine-tuning and deployment of BERT models, it seems that the potential to effectively leverage text data is higher now than it ever has been.