Markets generate the signal they price
The dynamics of capital markets are fundamentally shaped by the continual adversarial learning between market makers and market takers, driving price discovery, market efficiency, and the evolution of asset prices. The market is not merely an information processor, but is also an information generator that generates new information based on the interactions and trades between liquidity providers and takers. We propose a new framework for understanding market dynamics and modeling the evolution of asset returns through deep generative modeling, capturing the adversarial relationship between market makers and market takers in a setting of continuous information generation, processing, and price discovery. This framework connects the dynamics of market microstructure with markets in long-term horizon and explains the existence of market anomalies, such as dislocations, shocks, bubbles, regime shifts, and other abnormal asset pricingrelated phenomena.
We introduce the Generative Market Hypothesis (GMH): the claim that markets do not merely process exogenous information but actively generate new, priceable signal through the adversarial interaction of makers and takers. We formalize price discovery as a generative-adversarial game and instantiate it as Market GAN, a composable nexus that learns the conditional distribution of order flow under regime shifts. In equilibrium the framework recovers martingale pricing; in transition it exposes exploitable structure.
Price is not a sufficient statistic
Classical efficient-market theory treats price as a sufficient statistic for public information. Yet the act of trading is itself informative: every fill updates the latent state other participants condition on. We argue this feedback is not noise to be averaged away, but a generative process to be modeled directly.
- —Information is endogenous to the order book, not exogenous to it.
- —Maker / taker imbalance leads regime shifts.
- —Equilibrium pricing should emerge, not be assumed.
A two-player game for the tape
Let the market be a game between a maker population G that proposes liquidity and a taker population D that discriminates mispricing. At equilibrium neither profits at the other's expense and prices form a martingale. Away from equilibrium — during regime shifts — the residual is tradeable edge.
Compose, adversarially train, deploy
- 01Generator learns conditional order-flow distributions seeded from historical microstructure.
- 02Discriminator scores realized mispricing against the generated tape.
- 03Both networks compose as nexuses and assemble into the Nexus-α strategy.
𝔼x[ log D(x) ] + 𝔼z[ log(1 − D(G(z))) ]
At the adversarial equilibrium, expected excess return is zero — price discovery is complete. The edge lives in the gap.
adversarial loop diagram ]
Figure 1. Market GAN architecture — generator G proposes liquidity; discriminator D scores mispricing.
Out-of-sample, simulated venues
Simulated, net of fees. Past performance is not indicative of future results.
equity curve ]
Figure 2. Cumulative net return; edge concentrates around detected regime shifts.
Equilibrium holds; edge is transitional
Recovered equilibrium prices satisfy the martingale property to within estimation error, supporting GMH's central claim. Exploitable structure concentrates around detected regime shifts — consistent with the hypothesis that generated signal is strongest when the market's own narrative is changing.
A deployable nexus, not a metaphor
GMH reframes price discovery as a generative-adversarial process, and Market GAN operationalizes it as a deployable nexus. Future work composes additional nexuses — Nexus-β, Nexus-γ — onto the same adversarial substrate.