AUTONOMOUS-FOX
Three streams of interconnected research. One objective: a general-purpose model of market behaviour.
Our Expertise
Stream 1: Learning the Building Blocks of Market Behaviour
Raw market data is noisy, constantly changing, and full of information that is useless for modelling. Before anything else can work, the data needs to be transformed into something meaningful.
Our first research stream investigates whether continuous market behaviour can be compressed into a discrete set of learned representations, each capturing a distinct pattern of how an asset is behaving at a given moment. Think of it as learning a vocabulary of market behaviours directly from data, without defining those behaviours in advance. The goal is not just compression. It is to create a representational layer that can distinguish between market states at a much finer resolution than existing methods, separating genuine structure from noise. This representational layer is the foundation on which everything else in our research programme is built.
Stream 2: Discovering How Markets Change
Markets are not static. The statistical properties of asset behaviour change continuously: volatility expands and contracts, dependencies between assets strengthen and dissolve, and the patterns that drive price formation evolve over time. Existing statistical methods can infer behavioural states from market data, but they are limited in resolution. In practice, only a handful of distinct states can be reliably resolved, imposing hard limits on the granularity of what the model can distinguish.
Our second research stream builds on the representations learned in Stream 1 to discover fine-grained behavioural states at a far higher resolution, potentially distinguishing hundreds of distinct modes of market behaviour. These states are not predefined categories but structures that emerge through learning, along with the temporal dynamics that govern how markets transition between them. The result is a much richer and more granular picture of how market behaviour evolves through time than existing methods can achieve.
Stream 3: Mapping the Connections Between Assets
No asset moves in isolation. Assets influence one another through relationships that strengthen, weaken, and rewire themselves as conditions change. Most models either ignore these connections or assume they are fixed. They are not.
Our third research stream investigates how these dynamic relationships can be learned directly from data using graph-based methods. Rather than relying on pre-specified relationships, our models discover the relational structure between assets and track how it evolves. Crucially, this layer is conditioned on the market states discovered in Stream 2, so the relationships themselves adapt to the prevailing environment.
How They Fit Together
These three streams form a deliberate architectural sequence. Stream 1 transforms raw data into meaningful representations. Stream 2 discovers how market behaviour evolves through those representations. Stream 3 maps the connections between assets, conditioned on the current state of the market. Together, they are intended to produce a general-purpose model of market behaviour that can support a wide range of downstream applications.
The objective is depth of understanding, not breadth of application. A model that truly learns how markets behave can support many applications as a consequence. But the understanding comes first.
How They Fit Together
The modelling infrastructure we are building is general-purpose by design. While our primary focus is foundational research, the components we develop have direct applicability to a range of problems in quantitative finance.
Statistical arbitrage is one natural application. Traditional statistical arbitrage relies on fixed patterns and historical relationships that erode as they become widely known. A model that understands how assets behave in relation to one another within an evolving market network can detect transient opportunities that static methods miss, and adapt as those opportunities shift. Execution optimisation, risk management, portfolio construction, and market making are equally natural areas for downstream applications, each drawing on different aspects of the same underlying model.
We are not building trading systems. We are building the intelligence layer that trading systems have been missing.