Probability Modeling

What Probability Modeling Produces

Repute’s probability modeling layer converts normalized structural signals into probabilistic assessments of structural readiness and risk. The output is not a price prediction. It is a structural signal expressed as a probability — the likelihood that the currently observed structural conditions are consistent with historically identified breakout or collapse patterns.

Logistic Modeling Framework

The BPM (Breakout Probability Model) uses a logistic regression framework. The logistic function maps a weighted linear input to a probability in [0, 1]:

P = 1 / (1 + e^(−z))

Where:

z = w₁·AVI_norm + w₂·SIS_norm + w₃·NSM_norm + b

The weights w₁, w₂, w₃ are calibrated on historical data associating structural signal combinations with observable breakout events (defined as: 30%+ member growth + AVI sustaining above baseline for 14 days post-event).

Time-Series Acceleration Analysis

AVI is computed via time-series acceleration analysis — a two-pass process:

Pass 1: Signal computation For each time window t, compute the normalized mention/attention volume V_t.

Pass 2: Acceleration detection Compute the second derivative of the signal time series:

Acceleration_t = (V_t − V_{t-1}) − (V_{t-1} − V_{t-2})

Positive acceleration means the rate of change is itself increasing — not just that the signal is growing, but that it is growing faster over time. This is the structural definition of velocity in the AVI model.

The SustainabilityFactor is applied to prevent single-window spike events from generating falsely high acceleration readings.

Anomaly Weighting

Detected anomalies from the anomaly detection pipeline are incorporated into the probability model as negative weights:

P_adjusted = P × (1 − AnomalyWeight × AnomalyScore)

Where AnomalyWeight is calibrated based on the historical correlation between anomaly types and subsequent structural collapse. High-symmetry spikes carry the highest anomaly weight; single-metric deviations carry the lowest.

This adjustment ensures that a high BPM driven by bot-inflated inputs is corrected before output.

Cross-Cluster Convergence

Cross-cluster convergence detects when multiple independent creator clusters — previously unrelated narrative graphs — begin propagating the same narrative simultaneously.

This is a structurally significant signal because it means the narrative is crossing cluster boundaries organically, reaching net-new audiences without centralized coordination.

Detection Method

For narrative N, convergence is detected when:

• Two or more previously disconnected creator graph components begin sharing edges related to N
• The edge formation occurs within a short time window (< 72 hours)
• The connecting creators are verified as independent (not coordinated accounts)

Cross-cluster convergence is a rare and strong signal. When detected, it increases the NSM score and triggers an acceleration alert in the signal feed.

Model Outputs

All outputs are probabilistic structural signals — not price predictions, not trading signals, and not financial recommendations.