Cause and Effect

Large Language Models are powerful machine learning models centered around natural language understanding. 

While having certain strengths, they also have several weaknesses, for example: 

Hallucinations - can create incorrect information. 

Causal reasoning - lack genuine causal understanding.

Outdated data - retraining is expensive and so keeping models updated with new data can be challenging.

Explainability - LLMs are complex and hard to interpret - leading to problems with transparency and trust.

In our AI hedge fund, and recently in our new tech entity Cognisca, we have been very focused on the above problems, particularly around causal understanding.

Why is this important? Well, understanding cause and effect is a key aspect of financial markets intelligence, but more importantly also general purpose (human-level) intelligence. 

It’s something that we as human beings do very well every day, and take for granted. 

So how do we tackle some of the above problems? Our approach combines the best of both worlds:

• A live, semantic representation of financial markets - call it a causal reasoning engine that is continually updated as markets themselves update. There's a lot more to it than this single sentence communicates (more details given below).

• Combined with an LLM for natural language input, it becomes a live reasoning engine for markets, that simultaneously mitigates some of the key issues with LLMs on their own.  

Coming back to those weaknesses again:

Hallucinations - our data representation is based on true market data and facts.

Causal reasoning - the representation encodes causal information by default. 

Outdated data - no longer an issue as the entire architecture is continually updated. It also deals with regime shifts and uncertainty.

Explainability - we map the relationships learnt from markets into a formal logical reasoning layer. This sounds complicated, but really what it does, is to provide formal, traceable, explanations. These are also updated in synchronisation with data updates. 

The result of the above is a financial markets reasoning copilot.

The system can be more data efficient than a pure LLM - we can generalise from far fewer examples / much less training data.

It can also handle abstract relationships such as categories, and hierarchies. This is useful for generalising to new situations.

And finally it can also incorporate prior knowledge more easily. There is no reason AI systems have to learn every relationship from data. Human beings enter the world with a certain amount of 'cognitive scaffolding' in order to better recognise faces for example. And AI systems should be no different - using some prior knowledge or relationships can be a major advantage.

We'll be releasing a version of the system soon, so check here for updates.

For the Curious Minds

As always, here with some more technical details for those that are interested.

The architecture described above is a combination of LLM and our own causal reasoning engine. Overall it's a neuro-symbolic architecture that uses a broader set of data processing techniques and data representations to mitigate some of the issues with a pure transformer architecture approach.

The causal reasoning engine takes in faster data updates (a more statistical process), transforms those into more abstract sets of rules and relationships (which are still dynamic and uncertain), and then at the highest level, transforms those into formal logical rules, based on a version of Prolog.

Formal logical rules based systems are a call-back to previous iterations of AI and are symbolic in nature. It's thought that much of the processing in the human brain, particularly around decision making, is also symbolic.