The release of ChatGPT and Bing’s incorporation of it has led to a series of high-profile AI releases. While the singularity-precipitating moment of AGI (artificial general intelligence) is not upon us quite yet, generative AI and these Turing-test-acing chatbots will certainly shift the landscape in many industries, including our own industry of finance and investing.
Here we look at areas within finance that are likely to be affected, how they’ll be affected, risks thereof, and where our own products come in, whether those products are already on the market or on the roadmap.
Generative AI can help develop more advanced trading algorithms by analysing large datasets, recognising patterns, and generating predictions or decisions to capitalise on market inefficiencies. Just how early trading algos affected markets, this could lead to increased high-frequency trading and algorithmic trading strategies, potentially impacting market dynamics and liquidity. One question is whether Generative AI would exacerbate flash crashes or mitigate their severity.
Another question is the composition of the “large datasets”, and we believe this will go well beyond numeric inputs to include language content, especially as market psychology relies significantly on current events. Our own data, i.e., news and other textual sources of information, are ready to form the basis for NLU-driven recommendations. We already might for some of our clients that are algo-based.
AI-driven tools can optimise investment portfolios by analysing historical data, generating predictions, and recommending asset allocation strategies. Generative AI can also help identify uncorrelated assets to improve risk management through diversification. With large data sets, this identification can be particularly tricky, especially for retail and at-home investors who don’t have a team to help them find correlations. The “generative” part would be transforming the occluded patterns into human-actionable language to notify the humans that two assets would be a great diversification pair.
Financial analysis and research
Generative AI can process vast amounts of information, including financial statements, news articles, and economic reports, to generate insights and recommendations for investors. This could lead to a more data-driven approach to investment research and decision-making, even for traditional investors that mainly relied on “gut instinct” and intuition when making the final decision to pull the trigger on that big investment.
Because our products are geared heavily towards financial analysis and research, we see two major caveats here that our products could improve upon: AI “hallucinations” and source mapping. As Microsoft embarrassingly showed during a demonstration, Generative AI may well live up to its name and generate completely false outputs based on conjured-out-of-the-ether data. These are sometimes termed hallucinations.
Moreover, because accuracy is paramount in financial decision-making and research, having access to original sources is critical. With CityFALCON, users can always find the original articles or company filings or other source material to make a final verification of data integrity. In our industry, accuracy is almost as important as having any data at all.
Another area where we already have a usable product is sentiment analysis. NLP-based AI can analyse large volumes of text data, such as news articles, social media posts, and earnings calls, to gauge the sentiment of those sources. Our sentiment analysis is applied to individual articles (and actually at the clause-level), and the resulting data is mapped to topics and groups of topics (even aggregated at the country level).
Our approach uses supervised methods to ensure the algorithms stay on-track, and we can control the perspective: an issue with unsupervised and/or crowdsourced approaches here is that different people will have different sentiments for the same event – this is inevitable when you have buyers and sellers on opposite sides.
Furthermore, because of that buy/sell dichotomy, we do not give a buy/sell recommendation but a positive/negative sentiment for the article or topic. We believe this approach is the most sensical for our users, who could be on the buy or the sell side of any transaction.
Robo-advisors and personal finance
Generative AI can power robo-advisory platforms, providing personalised financial advice and investment management services to retail investors at a lower cost than traditional human advisors. Robo-advisory services have been on the market for a while now, and it is very likely they will augment their current approaches with each new iteration of AI technology. They also need the data for advising, some of which could be news, filings, and company reports, which we could supply via our API – a beautiful automated web of data flows.
With its core strength of analysing vast amounts of data (has the motif established itself yet?), AI models can help assess and manage risk more effectively by identifying patterns and predicting potential losses or defaults. Of course, this area is ripe for algorithmic discrimination and regulations are very likely to tightly control how algorithms can be used – someday, as so far governments have been far behind the curve in regulating new technologies. The aforementioned diversification could be risk management for individuals (and even internal corporate teams), based on numbers and text information processed through NLU.
Generative AI can help financial institutions automate compliance tasks, such as reporting, based on the forms and structures such institutions already have in place. Regtech is a popular space for startups and we fully expect this area only to grow. This is tangential to us, but we could help provide some of the reports and news to complement the existing data regtechs utilise.
One specific type of (currently way hyped up) generative AI, chatbots, can play a valuable role in the financial markets through a multitude of applications.
Chatbots can handle routine customer inquiries and support tasks, such as account balance checks, transaction history, and password resets. Putting all of this in one place and accessible through a single sentence is the true value add, since even old-school banks have automated this in one way or another by 2023.
For our own chatbot, users could check their watchlist sentiment, find prices for their watched assets, pull up tables of financial metrics, and request the original source filings all in a single window. With a voice component, this could become as easy as speaking. Not only does this improve the user experience, it could help users process information into knowledge in a new way and illuminate previously-hidden investment ideas.
One part of the CityFALCON mission is to educate users on financial markets. Our blog can do this for more complex topics, and the CItyFALCON chatbot could further provide information about products, investment strategies, and market trends. This can help increase financial literacy and empower users to make informed decisions about their investments. Of course a major risk is bad recommendations, and we will always press users to do verify their strategies – we are not pivoting from NLP and NLU into robo-advising just because we have a chatbot.
While we are not currently making recommendations, other companies may use chatbots to offer personalised financial advice based on a user’s financial goals, risk tolerance, and investment preferences. By leveraging AI-driven algorithms, chatbots can recommend tailored investment strategies, helping users build and manage their portfolios more effectively. This would be particularly useful for robo-advisors to explain asset allocation and receive real-time feedback from their users, an intriguing potential for NLU-based feedback loops from chatbot-to-user-and-back.
Market news and analysis
Conversely, we are in the business of providing users with real-time market news, updates, and analysis. Users can receive relevant information about stocks, bonds, or other financial instruments right in the chatbot. We envision our chatbot allowing a user to interact with our platform and any data therein as it already exists, which means not only news but corporate documents, sentiment, and even elements like trending assets can be understood within the CityFALCON chatbot.
Trading and execution
At some point we may integrate with brokers to allow trading from CityFALCON, and our chatbot could enable users to execute trades directly through the chat interface. Of course, brokers are very likely to have their own chatbots that will assist users in this regard too. NLU processing of user inputs would be absolutely critical here, because chatbot-based fat finger trades would be catastrophic to broker reputations.
RISKS AND COUNTERMEASURES
While generative AI and chatbots offer several potential benefits, major risks remain and new ones are likely to arise. These include concerns about data privacy (whose data is used as the source and will it leak?), algorithmic bias and discrimination, copyright infringement and stolen credit, increased market volatility (especially if those flash crashes worsen), and the potential for AI-generated market manipulation. Furthermore, as AI systems race ahead, the regulatory landscape will need to adapt to address these emerging challenges and do so in an environment of accelerating change.
Another area of concern is unquestioning users, who may start to accept any chatbot responses as infallible. (A separate issue would be complete lack of trust due to high-profile misses and hallucinations). This greatly raises the stakes for anyone implementing a chatbot in the financial space. Imagine the legal nightmare for the company whose chatbot says “buy Stock S at Price P” and a whole slew of retail investors get burned. A small disclaimer “this is for educational purposes only” may not satisfy the courts if the clear corporate messaging is “u. This is a major reason we are taking a supervised learning approach, even though the efforts are higher than unsupervised approaches and we are not making recommendations but the information basis for making such decisions.
The supervised approach will also enable us to rectify bias earlier, as we can be vigilant for it and retrain the models whenever bias may start to creep in. This would ensure we remain compliant with any regulations that may affect our part of the industry (which is, to be clear, far less regulated than those dealing with actual payments, transactions, and capital flows).
Overall, generative AI, NLU applications, and chatbots have the potential to enhance user experiences, streamline operations, and improve the delivery of financial content for users, either enterprise or retail. They’re also likely to pose plenty of challenges to regulators and those implementing the technologies, and we are poised to not only deliver the demanded services but also to mitigate risks with supervised learning and being conscious of those risks from the start.