As AI adoption accelerates across financial services, explainability remains one of the industry's most significant challenges. Brandon Davies explores the development of a Human in the Loop AI model designed to generate forward-looking market insights while maintaining transparency, accountability and regulatory acceptability. Drawing on behavioural science, conviction narratives and machine learning, the model seeks to provide a more explainable alternative to traditional foundation models. The article examines how organisations can balance innovation with governance while improving confidence in AI-driven decision-making across increasingly complex and dynamic financial markets.
Just over 12 months ago I and my colleague Professor Ali Kabiri wrote an
article for this magazine explaining our initial research on the development of
a “Human in the Loop” AI model.
It was a project which we felt would potentially allow us to gauge risk
in financial markets in a forward-looking way and had the potential to both out
perform conventional Large Language Models.
Moreover, we felt it had the scope to do that in a way that produced explainable
results and ran on a small fraction of the tokens that foundation LLM models
used.
It has taken us a year to complete and test the model, and the results
have been very encouraging.
Econometric testing with macrofinancial data is the first real test of
the validity of any proposed new model, and this is where we set our initial
goal for our model.
These results have been encouraging and have subsequently led to a paper
that goes into a detailed statistical appraisal of the model’s structure and
performance.
The result is something that, while being very technical, shows that the
outturns of the model not only give original insight into the workings of the
model but also of the markets themselves.
This is perhaps the most important aspect of the model, since it has uncovered
explanative variables that are not seen in the advanced foundation LLMs against
which we have calibrated our model.
What is of possibly more importance to many risk managers is that the model’s
results are not a “black box”, though through advanced statistical analysis of
the results we are able to explain why it produces the results it does.
For any heavily regulated industry this is a very important aspect of any
model as their use is likely to need approval from regulators within whatever
section of the finance industry would be proposing its use.
The reasons why we concentrated on building a “Human in the Loop”
model are explained in detail in our original article for Connect
Magazine.
To briefly recap, it is because financial markets are different from many
aspects of the natural world, where results are replicable due to the existence
of laws of nature.
Typical examples would be physics, chemistry or biology where outturns are
overwhelmingly stable with regard to changes in inputs.
The financial world is very different. Indeed, financial markets exhibit
great variation in outturns given very similar circumstances at different
points in time, since financial market outturns are dynamic and conditional,
and the markets themselves are subject to heterogeneity of factors and idiosyncratic
effects.
Basically, this means the factors affecting market outturns are many and
are unstable in both number and importance.
From Adam Smith to John Maynard Keynes, economists have recognized that
the way people interpret financial circumstances can vary greatly across time.
Most famously, Keynes called this “Animal Spirits,” though you might also
recognize other descriptors such as “Irrational Exuberance,” “The Madness of
Crowds,” or even “Risk Appetite.”
All are essentially descriptors for some effect that we know affects market
outturns but which we cannot measure.
But with AI, and especially LLMs, we can turn words into data. Yet words are
not simply constructs from a dictionary definition. They combine with other
words to give meaning and, importantly for our model, they create emotion.
To incorporate emotion into our model, we have used Conviction Narrative
Theory, which explains how humans make decisions under radical uncertainty.