Core Models and Methods

AQFM does not organize around standalone predictive models. It organizes around methods for estimating, steering, and stabilizing the loss distribution under intervention.

The central object is the conditional loss distribution:

$$L \sim f(l \mid \nu, \mathcal{C})$$

Models are only useful insofar as they improve the ability to change the shape of this distribution through controls.

Core methods in AQFM fall into three functional layers: state estimation, response modeling, and distributional control.

---

1. State Estimation (Measure → Latent System Reconstruction)

State estimation is the process of reconstructing the hidden structure of the system that generates losses.

Because fraud is only partially observable, AQFM treats all signals as indirect projections of a latent environment $\nu$.

State estimation methods focus on inferring:

• exposure concentration across system segments
• adversarial pressure intensity
• behavioral instability and regime shifts
• vulnerability structure in product and payment flows
• sensitivity of losses to different control levers

This layer is not about predicting fraud events. It is about building a live representation of the system state that determines the shape of the loss distribution.

The output of state estimation is a structured view of where risk mass is accumulating and how it is evolving.

---

2. Response Modeling (Model → Control-to-Distribution Mapping)

Response modeling describes how the loss distribution changes when controls are applied.

Instead of asking whether a transaction is fraud, AQFM asks:

“How does this control change $f(l \mid \nu, \mathcal{C})$?”

This introduces a shift from predictive modeling to counterfactual distribution modeling.

Response models estimate:

• how interventions shift probability mass in the loss distribution
• how tail risk reacts to specific controls
• how variance changes under different policy configurations
• how adversaries adapt to control pressure
• how exposure migrates across cohorts and channels

The key modeling object is not a label, but a distributional response function:

$$\Delta f(l \mid \mathcal{C}_1 \rightarrow \mathcal{C}_2)$$

A model is valuable if it improves understanding of how controls reshape the loss distribution under real system constraints.

---

3. Distributional Control (Decide → Stochastic Ordering of Interventions)

Distributional control is the mechanism by which AQFM actively shapes risk.

Controls are not selected based on individual outcomes, but on their effect on the entire induced loss distribution.

AQFM evaluates interventions using stochastic dominance:

• First-order effects: controls that shift the entire loss distribution left
• Second-order effects: controls that reduce dispersion and compress tail risk without necessarily changing the mean

Both are valid forms of improvement because they alter the structure of the loss-generating process.

Methods in this layer include:

• selective friction allocation based on state
• cohort-level control differentiation
• dynamic thresholding across risk surfaces
• intervention portfolios across multiple system layers
• tail-targeted suppression strategies

The objective is to systematically reduce the probability and severity of extreme outcomes, not just optimize average fraud rates.

---

4. Adaptive Distributional Calibration (Adapt → Stability Under Drift)

Adaptation ensures that distributional control remains valid as the system evolves.

Because both adversaries and internal controls reshape the environment, the mapping:

$$\mathcal{C} \rightarrow f(l \mid \nu, \mathcal{C})$$

is non-stationary.

Adapt methods monitor:

• degradation of first-order dominance relationships
• weakening of tail compression effects
• shifts in how controls map to loss outcomes
• migration of risk across previously stable segments
• emergence of new high-impact regions in the distribution

When these changes occur, the system recalibrates state estimation and response models to restore effective distributional control.

Adapt is therefore not retraining. It is maintenance of control validity over a moving loss distribution.

---

Unifying Principle

Across all layers, AQFM methods are evaluated by a single criterion:

How effectively do they improve control over the shape of the loss distribution under uncertainty and adversarial adaptation?

Prediction accuracy is secondary. Local classification is incidental.

The primary object is the stability, tail behavior, and directional movement of the loss distribution under intervention.