Encode GPT

Bloomberg has made an unprecedented move by introducing BloombergGPT, a proprietary AI model trained for financial tasks, to the world of finance. This model is powered by the immense amount of financial data that Bloomberg has amassed over the last forty years, and is designed to be a conversational and accessible chatbot. While this could bring great promise to the sector, it could also bring unforeseen consequences that could severely impact investments and drive irrationality in the financial markets.

<h2>Bloomberg has made an unprecedented announcement that could change how we understand finance and invest our money forever.</h2>But why? Well, it’s pretty simple.Think about all the financial data that one of the biggest financial institutions in the world has amassed over the last forty years, and then picture a conversational and accessible chatbot that has all that knowledge inside.<blockquote>That’s BloombergGPT.</blockquote>BloombergGPT represents the merge between the most crucial sector in the world, finance, with one of the fundamental shifts in the history of human progress, Generative AI.But BloombergGPT could also bring unforeseen consequences that can severely impact your investments and drive the overall irrationality of the financial markets to new records.<h1>Differential data, differential outcomes</h1>Bloomberg has decided to enter the seemingly unexplored world of financial LLMs, AI models trained for financial tasks, due to the great promise these models have.However, BloombergGPT, not to be confused by its name with OpenAI’s family of products, is an end-to-end proprietary Bloomberg solution.In other words, they haven’t simply connected their data to OpenAI’s GPT-4 API and fine-tuned a model to their needs — i.e. further train an already pre-trained model like GPT-4.Quite the contrary, they’ve created a model from scratch.Particularly, they’ve based its development on BLOOM, an open-source project led by the co-founder of Hugging Face that created a large language model (LLM) that, in contrast to OpenAI’s GPT family, is accessible to anyone, while being very similar in size.But this isn’t what makes BloombergGPT really special.<h2>363 billion reasons to become your best financial friend</h2>Albeit being in one of the most competitive industries today, BloombergGPT has one characteristic that makes it unique.It’s the model trained in the biggest financial dataset in history, 363 billion tokens (a token being a set of characters or words).<blockquote>This means that this model can be accurately described as the LLM that has been more largely exposed to financial data that we’ve ever seen, vastly embarrassing any other possible comparison.</blockquote>But that’s not all.To maximize results, they’ve taken a new approach to training, doing so using a diverse dataset where around half of it is strictly in-domain, with examples like financial reports and news, and with the other half being public, general knowledge information available mainly on the Internet.This has allowed this model, according to the performed evaluations, to become the first-ever LLM that excels at finance while being reasonably good at other fields, all the while being much smaller than state-of-the-art models like GPT-4.But how is this even possible?<h2>50 billion reasons not to go larger</h2>Bloomberg is very open on the particularities of the model, starting with it being ‘just’ 50 billion parameters big.Naturally, one is inclined to think: if these models get better as they grow in size, why just 50 billion parameters when we have models like GPT-3 at 175 billion parameters, Google Brains’ PaLM at 580 billion, or the humungous Megatron-LM at an insane 1 trillion?The truth is that the number of parameters isn’t arbitrary, and you need to take into account many factors like, for instance, the size of your dataset.<blockquote>If your train a model that’s too big for your dataset, your model will overfit.That is, it will learn to adjust completely to your data by memorizing it, making it unsuitable for data outside the training corpus.On the contrary, if your model is too small for your dataset, your model will underfit, meaning it won’t be great at predicting by lacking the necessary size to learn your data correctly.</blockquote>For those reasons, using <a target="_blank" rel="noopener noreferrer nofollow" class="ae nq" href="https://openreview.net/forum?id=iBBcRUlOAPR">Hoffmann et al</a>’s guidelines for ideal model size for training, they settled for 50 billion.But the surprising thing is that, despite being considerably smaller than other models, it vastly outperforms them in financial topics thanks to its unique training dataset.All in all, everything here sounds great, but what did Bloomberg actually achieve?<h2>Great calculator, better psychologist</h2>The evaluation of BloombergGPT involved several financial tasks. These included:<ol><li>FPB: A sentiment classification task on sentences from financial news</li><li>FiQA SA: Predicting aspect-specific sentiment in financial news and microblog headlines.</li><li>Headline: A binary classification task determining if a news headline in the gold commodity domain includes certain information.</li><li>NER: A named entity recognition task on financial data for credit risk assessment.</li></ol><blockquote>Named Entity Recognition involves correctly identifying entities in data.For instance, for the sentence “Twitter has open sourced its algorithm”, the machine should be capable of tagging Twitter as an Organization.On the flip side, for the sentence “Twitter releases its newest feature”, Twitter could be tagged as an Organization or a Product, making it a non-trivial exercise.</blockquote>5. ConvFinQA: A task requiring answering conversational questions based on S&amp;P 500 earnings reports, reporting exact match accuracy on the public development set.Unsurprisingly, BloombergGPT performed the best among all models (GPT-Neo X, OPT 66B, and BLOOM 176B) in four of the five tasks (ConvFinQA, FiQA SA, FPB, and Headline) and ranked close second in NER, as we can see below:<img src="https://miro.medium.com/v2/resize:fit:700/1*1qbFOT_lfBmLESbMSDX1Rg.png" alt="">Source: BloombergGPT: A Large Language Model for FinanceTherefore, we could soon see BloombergGPT being used to analyze market sentiment, interpret news headlines, elaborate financial documents, and even make trading decisions.Awesome, but at what cost?<h1>Giving our money to robots</h1>Of all the possible contributions BloombergGPT is going to have, I bet that the one that will give Bloomberg a bang for its buck will be sentiment analysis.<h2>A sentiment machine</h2>BloombergGPT was evaluated on five internal aspect-specific sentiment analysis tasks, all in English.The tasks were:<ol><li>Equity News Sentiment: Predicting sentiment expressed in news stories toward a company.</li><li>Equity Social Media Sentiment: Predicting sentiment expressed in financially-relevant social media content toward a company.</li><li>Equity Transcript Sentiment: Predicting sentiment expressed in transcripts from company press conferences toward a company.</li><li>ES News Sentiment: Predicting sentiment expressed in news stories toward a company regarding its environmental and social policies.</li><li>Country News Sentiment: Predicting sentiment expressed in news stories toward a country’s economy.</li></ol>BloombergGPT outperformed all other models on these sentiment analysis tasks, as you can see below:<img src="https://miro.medium.com/v2/resize:fit:700/1*X6-HH0Zh7uz1dFkTOkpsxQ.png" alt="">Source: BloombergGPT: A Large Language Model for FinanceNow, I want you to think for a second about the potential implications of this.<h2>Making the emotional markets even more irrational</h2>If we were capable of calculating what percentage of investment decisions are strictly emotional decisions, especially those made on a short-term maturity, we would probably be way above a comfortable value.In fact, situations like the ones many public companies suffer every day, like seeing 10% of your value erased for a wrong statement by your chatbot, as Google saw with Bard, signal how irrational and nonsensical the financial markets are.Nevertheless, hundreds of millions of dollars are moved based on headlines, biased-analyst predictions, and ill-informed decision-making.For instance, people started to dump their positions in Republic First Bancorp by confusing this bank’s name with the troubled First Republic Bank (hysterically, this has led to a monthly decrease in value of 30% for this bank’s shares, simply by having a similar name to the other bank).Therefore, how dangerous would it be if this tool becomes commonly used by market participants?<h2>Uncontrolled decisions and volatility</h2>Imagine that all investment banks' trading departments and other powerful market participants all started using the same tool to evaluate market sentiment for more efficient day trading.Suddenly, money worth potentially billions of dollars would be dependent on how well this tool gauges sentiment from news headlines or financial reports, allowing companies and retail investors to make sudden investments in search of that extra penny.Naturally, if large bag holders consolidated into the same tools, volatility would heavily increase, as a great part of market participants would receive the same alpha and arrive at the same conclusions based on the same biased model, while making those unable to leverage these technologies get caught swimming naked.A market with higher inequality and volatility?And what would happen then if the models develop specific, unexplained biases toward certain companies, forcing them to have a negative sentiment towards them and, consequently, turning the whole market against them?<blockquote>In summary, my point is that the extreme convenience of these models could make them relevant factors in deciding overall market movements in a market already fueled with rampant speculation and clumsy decision-making.</blockquote>For example, I would be very surprised if trading bots aren’t automatically enhanced with these models to make automated trading decisions.Of course, one could argue for the contrary, stating that these models will also be used to get a better understanding of companies to make better decisions based on the huge knowledge they possess.But the extreme velocity of the market’s reaction to good or bad news, added to the natural tendency humans will have to accommodate themselves and automate certain decisions in jobs where stress and long hours are the common trends, leads me to believe that the potential outcome of these models could be potentially far more negative than positive.<h1>A final word</h1>BloombergGPT is only the tip of the iceberg when it comes to AI innovation, which gives a sense to how difficult is to stay up-to-date with this technology.But if you subscribe to my newsletter, you can do that with just 5 minutes a week, as you’ll deep dive into innovation topics in an easy-to-read, digested manner, providing you with the latest news and insightful innovations on AI and Crypto.<blockquote>Become aware of the technologies that are shaping our future. Be ready to profit from that knowledge.</blockquote>

 Bloomberg has made an unprecedented announcement that could change how we understand finance and invest our money forever. But why? Well, it’s pretty simple. Think about all the financial data that one of the biggest financial institutions in the world has amassed over the last forty years, and then picture a conversational and accessible chatbot that has all that knowledge inside. That’s BloombergGPT. BloombergGPT represents the merge between the most crucial sector in the world, finance, with one of the fundamental shifts in the history of human progress, Generative AI. But BloombergGPT could also bring unforeseen consequences that can severely impact your investments and drive the overall irrationality of the financial markets to new records. Differential data, differential outcomes Bloomberg has decided to enter the seemingly unexplored world of financial LLMs, AI models trained for financial tasks, due to the great promise these models have. However, BloombergGPT, not to be confused by its name with OpenAI’s family of products, is an end-to-end proprietary Bloomberg solution. In other words, they haven’t simply connected their data to OpenAI’s GPT-4 API and fine-tuned a model to their needs — i.e. further train an already pre-trained model like GPT-4. Quite the contrary, they’ve created a model from scratch. Particularly, they’ve based its development on BLOOM, an open-source project led by the co-founder of Hugging Face that created a large language model (LLM) that, in contrast to OpenAI’s GPT family, is accessible to anyone, while being very similar in size. But this isn’t what makes BloombergGPT really special. 363 billion reasons to become your best financial friend Albeit being in one of the most competitive industries today, BloombergGPT has one characteristic that makes it unique. It’s the model trained in the biggest financial dataset in history, 363 billion tokens (a token being a set of characters or words). This means that this model can be accurately described as the LLM that has been more largely exposed to financial data that we’ve ever seen, vastly embarrassing any other possible comparison. But that’s not all. To maximize results, they’ve taken a new approach to training, doing so using a diverse dataset where around half of it is strictly in-domain, with examples like financial reports and news, and with the other half being public, general knowledge information available mainly on the Internet. This has allowed this model, according to the performed evaluations, to become the first-ever LLM that excels at finance while being reasonably good at other fields , all the while being much smaller than state-of-the-art models like GPT-4. But how is this even possible? 50 billion reasons not to go larger Bloomberg is very open on the particularities of the model, starting with it being ‘just’ 50 billion parameters big. Naturally, one is inclined to think: if these models get better as they grow in size, why just 50 billion parameters when we have models like GPT-3 at 175 billion parameters, Google Brains’ PaLM at 580 billion, or the humungous Megatron-LM at an insane 1 trillion? The truth is that the number of parameters isn’t arbitrary, and you need to take into account many factors like, for instance, the size of your dataset. If your train a model that’s too big for your dataset, your model will overfit. That is, it will learn to adjust completely to your data by memorizing it, making it unsuitable for data outside the training corpus. On the contrary, if your model is too small for your dataset, your model will underfit, meaning it won’t be great at predicting by lacking the necessary size to learn your data correctly. For those reasons, using Hoffmann et al ’s guidelines for ideal model size for training, they settled for 50 billion. But the surprising thing is that, despite being considerably smaller than other models, it vastly outperforms them in financial topics thanks to its unique training dataset. All in all, everything here sounds great, but what did Bloomberg actually achieve? Great calculator, better psychologist The evaluation of BloombergGPT involved several financial tasks. These included: FPB: A sentiment classification task on sentences from financial news FiQA SA: Predicting aspect-specific sentiment in financial news and microblog headlines. Headline: A binary classification task determining if a news headline in the gold commodity domain includes certain information. NER: A named entity recognition task on financial data for credit risk assessment. Named Entity Recognition involves correctly identifying entities in data. For instance, for the sentence “Twitter has open sourced its algorithm”, the machine should be capable of tagging Twitter as an Organization. On the flip side, for the sentence “Twitter releases its newest feature”, Twitter could be tagged as an Organization or a Product, making it a non-trivial exercise. 5. ConvFinQA: A task requiring answering conversational questions based on S&P 500 earnings reports, reporting exact match accuracy on the public development set. Unsurprisingly, BloombergGPT performed the best among all models (GPT-Neo X, OPT 66B, and BLOOM 176B) in four of the five tasks (ConvFinQA, FiQA SA, FPB, and Headline) and ranked close second in NER, as we can see below: Source: BloombergGPT: A Large Language Model for Finance Therefore, we could soon see BloombergGPT being used to analyze market sentiment, interpret news headlines, elaborate financial documents, and even make trading decisions. Awesome, but at what cost? Giving our money to robots Of all the possible contributions BloombergGPT is going to have, I bet that the one that will give Bloomberg a bang for its buck will be sentiment analysis. A sentiment machine BloombergGPT was evaluated on five internal aspect-specific sentiment analysis tasks, all in English. The tasks were: Equity News Sentiment: Predicting sentiment expressed in news stories toward a company. Equity Social Media Sentiment: Predicting sentiment expressed in financially-relevant social media content toward a company. Equity Transcript Sentiment: Predicting sentiment expressed in transcripts from company press conferences toward a company. ES News Sentiment: Predicting sentiment expressed in news stories toward a company regarding its environmental and social policies. Country News Sentiment: Predicting sentiment expressed in news stories toward a country’s economy. BloombergGPT outperformed all other models on these sentiment analysis tasks, as you can see below: Source: BloombergGPT: A Large Language Model for Finance Now, I want you to think for a second about the potential implications of this. Making the emotional markets even more irrational If we were capable of calculating what percentage of investment decisions are strictly emotional decisions, especially those made on a short-term maturity, we would probably be way above a comfortable value. In fact, situations like the ones many public companies suffer every day, like seeing 10% of your value erased for a wrong statement by your chatbot, as Google saw with Bard, signal how irrational and nonsensical the financial markets are. Nevertheless, hundreds of millions of dollars are moved based on headlines, biased-analyst predictions, and ill-informed decision-making. For instance, people started to dump their positions in Republic First Bancorp by confusing this bank’s name with the troubled First Republic Bank (hysterically, this has led to a monthly decrease in value of 30% for this bank’s shares, simply by having a similar name to the other bank). Therefore, how dangerous would it be if this tool becomes commonly used by market participants? Uncontrolled decisions and volatility Imagine that all investment banks' trading departments and other powerful market participants all started using the same tool to evaluate market sentiment for more efficient day trading. Suddenly, money worth potentially billions of dollars would be dependent on how well this tool gauges sentiment from news headlines or financial reports, allowing companies and retail investors to make sudden investments in search of that extra penny. Naturally, if large bag holders consolidated into the same tools, volatility would heavily increase, as a great part of market participants would receive the same alpha and arrive at the same conclusions based on the same biased model, while making those unable to leverage these technologies get caught swimming naked. A market with higher inequality and volatility? And what would happen then if the models develop specific, unexplained biases toward certain companies, forcing them to have a negative sentiment towards them and, consequently, turning the whole market against them? In summary, my point is that the extreme convenience of these models could make them relevant factors in deciding overall market movements in a market already fueled with rampant speculation and clumsy decision-making. For example, I would be very surprised if trading bots aren’t automatically enhanced with these models to make automated trading decisions. Of course, one could argue for the contrary, stating that these models will also be used to get a better understanding of companies to make better decisions based on the huge knowledge they possess. But the extreme velocity of the market’s reaction to good or bad news, added to the natural tendency humans will have to accommodate themselves and automate certain decisions in jobs where stress and long hours are the common trends, leads me to believe that the potential outcome of these models could be potentially far more negative than positive. A final word BloombergGPT is only the tip of the iceberg when it comes to AI innovation, which gives a sense to how difficult is to stay up-to-date with this technology. But if you subscribe to my newsletter, you can do that with just 5 minutes a week, as you’ll deep dive into innovation topics in an easy-to-read, digested manner, providing you with the latest news and insightful innovations on AI and Crypto. Become aware of the technologies that are shaping our future. Be ready to profit from that knowledge. Bloomberg has made an unprecedented move by introducing BloombergGPT, a proprietary AI model trained for financial tasks, to the world of finance. This model is powered by the immense amount of financial data that Bloomberg has amassed over the last forty years, and is designed to be a conversational and accessible chatbot. While this could bring great promise to the sector, it could also bring unforeseen consequences that could severely impact investments and drive irrationality in the financial markets. 

Bloomberg has Just Launched ChatGPT for Finance

https://us-prod-content.primeobjects.net/v1/

DouHub Team

types

All partners

All competitors

All customers/clients

All investors

All service/product providers

All knowledge/idea providers

Organization Administrator

Content Administrator

User Administrator

License Administrator

User

Help you with product management, marketing, R&D, and operations to ensure a successful SaaS business from Zero to One.

Bloomberg has Just Launched ChatGPT for Finance

More Readings