AI-driven Detection

Raw variance signals extracted from markets are noisy if used directly. This feature defines how Variance Markets stabilizes those signals by detecting belief regimes using an AI-driven classification layer. This layer does not predict outcomes and does not estimate “true” probabilities. Its only role is to identify coherent behavioral regimes inside a live prediction market.

The input to this feature is a set of observable time-series derived from active markets on Polymarket.

For a given market m, the system observes the following variables over time t:

• p(t): market price (implied probability)

• dp/dt: price velocity

• v(t): traded volume

• l(t): liquidity or depth proxy

These variables are sampled continuously while the market remains unresolved.

Regime Hypothesis

Before resolution, a prediction market does not operate under a single belief model. Instead, it exhibits multiple belief regimes, corresponding to different reaction speeds, confidence levels, and time horizons.

At any time t, the market is assumed to be in one regime R(t), where:

R(t) ∈ { r1, r2, …, rk }

Each regime represents a coherent belief dynamic with internally consistent behavior.

Regime Classification

The AI layer estimates, for each time step, the probability that the market is in a given regime:

P(R = ri | phi(t))

The specific model (clustering, hidden-state model, regime-switching network, etc.) is implementation-dependent. The protocol only relies on the deterministic output.

The active regime at time t is defined as the regime with the highest probability:

R_hat(t) = argmax over ri of P(R = ri | phi(t))