Broker to Trading

AI at the crossroads

Up until very recently AI was prevalent, but invisible. AI, via machine learning, powers large-scale digital technologies. AI makes it possible to make sense of huge amounts of data. This data is a consequence of the Internet and mass connectivity. Huge amounts of data are generated, which fortunately AI is tailor-made to look at and find patterns of interest. This works behind the scenes. However, generative AI has taken AI out of the shadows.

Generative AI is very recent, but it is based on the same type of programs as behind-the-scenes AI. Generative AI is an advance in Human-Computer Interaction (HCI), in particular Natural Language Processing (NLP). What this means is that generative AI is a way to talk with the AI, in ways that seems like a natural conversation. The reality is that a huge amount of processing goes on behind the scenes. Gen AI leverages the economies of scale provided by data centers (as do many technologies that are scaled to operate on the Internet).

But what is it that the human is interacting with? In essence, they are talking with a database that is built from Internet content. So to an extent, the human is talking with Internet content. Internet content itself covers a vast array of content that developed apart from the Internet.

But at its core, the AI is trained on the kind of things that one encounters when surfing the web. Trained means that the AI has absorbed this content, and created an image of it, which itself provides clues to what it means. Thus the AI can simulate an intelligent response when asked about its database (which comes from the Internet), as the intelligence is hard-wired into the image. So can this technology help traders?

What can gen AI do for the trader?

The database of the AI chatbot contains a vast amount of information, from specialised information about trading to general knowledge which may apply to trading. The advance of AI is that it can work out what parts of this are relevant to a question. There are big caveats that this does not always work out, that chatbots have to be trained in very detailed ways to simulate human responses, and that the information given can be completely wrong. There are ways around this, which is to verify all information that is crucial via an Internet search engine.

Let's think about a typical trader. They are facing a complex market which oftentimes does the completely unexpected. Or even if it does the expected, it does it only for a time that does not help the trader, who is looking for a sustained value difference between the start and end of the trade. Traders rely on all sorts of information to give them something to leverage off.

However, the AI chatbot may possess this kind of information. They do not usually have real-time information, as they have to rely on Internet searches to find current information. But markets update much faster than this. A technical indicator can be seen as a way to try and make this chaotic information fluctuation more stable. But the problem is that it fundamentally is not stable, so indicators often fail. Chatbots will not help here, as they cannot process information in real-time, at market speed. But they may be able to provide a contextual understanding of whatever information flux is currently happening.

This is why we have argued in other articles for chatbots as an adjunct for information ordering. Orderly information can provide a way to leverage decision-making in chaotic market conditions. Outside of this, chatbots can be useful for trade preparation, for finding out more about technical indicators, and for analysing data, such as statements. However, traders may want chatbots to do more. But for chatbots to do more may need advances in their underlying technology. This is where quantum computing comes into view.

What can quantum computing do?

Quantum computing cannot currently do a huge amount. This is because the technology has not been 'solved' in the way that digital technology has, allowing for improvement, iterations, and adaptation to use cases. Quantum computing is an idea, with parts of it implemented in technology. The core idea is that at the quantum level, particles can be in a superposition of multiple states at the same time. At some point, these states can collapse into a single state.

In classical computing, parallelism is used to speed up sequential processes. But as anyone who has designed a parallel algorithm knows, this is a complex process that requires that the sequential algorithm be translated into a parallel format. Additionally, the parallelism is limited by the nodes available to compute each parallel process.

Quantum computing is a step beyond this, with a type of massively parallel computation, without requiring separate nodes to make each distinct parallel calculation. That is, the quantum effect of superposition can maintain multiple computations in a single state, which, ideally, gets collapsed into the solution, without an inherent time taken to do this. This can be contrasted with classical parallel computation, where there is an inherent time taken for each process to complete its calculation steps.

The problem of sequential to parallel in classical computation is amplified however, as many problems cannot be specified in this way. But there is another problem, which is that the technology to do this is inherently unstable. The outside environment naturally intrudes upon the computational process. That is, it is very hard to separate the computation from its surroundings, as it is making use of a natural, fundamental process that does not tend to separate processing elements (in fact, it evolves as a unified whole, until it decoheres).

This is why the core of quantum computers is super-cooled and highly protected, in effect to isolate the computation from the environment. But this does not work out so well, currently. However, it may be that the algorithms that power gen AI might be suited to quantum parallelism, assuming it can become more effective.

Quantum parallelism and generative AI

The core technology driving gen AI is neural nets. These help create the image used by the chatbot, providing the chatbot with ways to work out what is relevant to the current input. But neural nets are pretty static, once trained. This is why chatbots are not so good at real-time information.

But imagine if the neural net could dynamically adjust itself. Even relatively static current neural nets absorb massive amounts of power. Making this more dynamic may be possible with huge investments in computing technology and data centers. But quantum computing might leapfrog over all this, by allowing a dynamic, responsive chatbot. Thus the trader could have a companion that could support them in the chaos of market information.

This is speculative, and there may be ways to do this without requiring advances in a novel technology. For example, using prompts to help clarify what it is that the trader is seeing or thinking in the chaos of the market, and thus try and overcome emotional issues with trading. What is real-time with gen AI is the human-like response to a question (which is why they are revolutionary).

A future trading LLM

Let's imagine that quantum parallelism has managed to make LLMs reactive to real-time data, such as the turmoil of markets. They also have the capacity current LLMs have, to communicate with a human and reason, or at least appear to convincingly, such that it does not make too much of a difference.

However, we still have to take account of the 'difference'. Currently, it can be taken into account by fact-checking. But this may be impossible in a complex, fast-moving trading market. But it is also taken account of by prompting and by interlacing the AI reasoning with human reasoning (which comes down to a cognitive effort and good, layered prompting). This is itself dependent on the trader's own experience.

In this future LLM, we might assume that there have been advances in AI reasoning, such that the human trader does not have to depend on any sense of augmentation. So the chatbot is providing genuine insights, that can also adapt with the trader's experience, of course, in a virtuous circle of knowledge.

Advanced AI reasoning

What might advanced AI reasoning be like, and how might it be useful for a trader? Currently, humans reason in different ways. Traders often make inferences from market-based data, that can vary depending on the type of trading. For example, earnings releases for Stocks and news data releases for Forex (with some crossover, of course, in this complex, interlaced system). But things are so chaotic and move so fast in real-time trading, that it is hard to apply advanced, sustained reasoning, such as can be done with static objects.

Traders may rely on their brains to do this, with intuition, i.e. what may be advanced reasoning applied to complexity, based on experience, and the way the mind works. But maybe a future LLM could do this, finding complex, causal chains in market events, acting as a reasoning system, to augment the trader's intuition, or to ground it, at least.

For now, the trader has to deal with complex markets without such a technology. However, a wide range of approaches have been developed, from high-speed order infrastructure to facilitate rapid orders in and out of the market, to exploiting micro-patterns, to low-cost trading.