Building a financial super-intelligence requires solving two problems, at a depth no prior system has reached.
The first is understanding the market. Every investor, whether running a quantitative fund or reading 10-Ks, starts from the same question: why do some assets earn more than others over the long run? The no-arbitrage condition implies the existence of a pricing kernel , which weights every possible future state of the world and must satisfy , for every traded return . A single equation unifies the CAPM, consumption-based models, the Fama-French factors, options pricing, and every other model of expected returns as alternative specifications of the same object. The same answers two questions at once: what returns should we expect, and how should risk be priced? What varies across models is the proposed functional form of — what it depends on, how it varies across states of the world, and why. The problem is that we do not know what looks like. Understanding what markets know means learning the pricing kernel from data: identifying which firm characteristics co-vary with , with what signs and magnitudes, and with what stability across regimes.
The second is understanding the individual investor. A financial system that knows everything about markets but nothing about the investor cannot give useful personalized advice. Goals, constraints, tax situations, behavioral tendencies, time horizons, the particular anxieties and aspirations of a specific financial life: these are not inputs that a model can infer from a brief questionnaire. They require deep, continuous, longitudinal modeling of a person across every dimension of their financial existence.
Both problems reduce to the same research question: what is the low-dimensional structure that generates the high-dimensional behavior we observe? For markets, it is the latent economic forces that drive asset prices, hidden behind the noise of daily returns. For the individual, it is the latent factors that shape financial behavior, risk tolerance, and the formation of goals, hidden behind the surface of transactions and stated preferences. The two problems sit in different literatures and use different data, but the underlying mathematical question is the same.
Neither problem can be solved without the other. A system that deeply understands markets but does not understand the investor optimizes against the wrong objective: the optimal efficient frontier in the abstract is not optimal for this specific person at this specific moment in their financial life. A system that understands the investor perfectly but does not know what markets actually price will produce recommendations that feel right and end up being wrong. Autonomous requires both the market knowledge and the concerns of the individual investors, built to interact and inform each other.
These problems define the research discipline at ATG. The work treats finance as a system and as an empirical science: we build, measure, and iterate. We'll publish as we go: methods, results, and dead ends alike. Our goal is to build trust in the systems we are developing and we believe sharing these results and building in the open is central to that ambition. It is also how research at this frontier compounds. Problems of this scale and complexity are not solved alone.