Signal Weighting Methodology

Each dimension is composed of multiple signals, and each signal carries a different predictive weight. Within the engagement dimension, executive meeting attendance carries more predictive weight than email response time — a customer whose executives stop attending meetings is at higher risk than one whose email responses slow down. Within the competitive position dimension, a named competitor appearing in the customer's job postings carries more weight than a general RFP announcement. The signal weights determine how much each behavioral observation moves the health score, and incorrect weights produce misleading scores.

1. Start with Expert-Informed Weights Interview your most experienced account managers. Ask: "Which behaviors most reliably predict churn? Which behaviors most reliably predict expansion?" Their answers are your initial weights. An experienced account manager who says "when the executive sponsor stops attending quarterly reviews, we lose the account within 12 months" is providing the weight for the executive attendance signal based on pattern recognition across hundreds of interactions.
2. Calibrate Against Historical Outcomes Compare the signal weights against actual outcomes. For every account that churned in the past two years, calculate what the health score would have been under your current weighting three months before churn. If the score did not detect the risk, the weights need adjustment. The signals that were present before churn but did not move the score enough are under-weighted. The signals that moved the score but were not present before churn are over-weighted.
3. Iterate Quarterly Signal weights are not permanent. The behaviors that predicted churn two years ago may not predict churn today — market conditions change, competitive dynamics shift, and customer expectations evolve. Review the weight calibration quarterly by comparing predicted health trajectories against actual outcomes. Adjust weights that are systematically over-predicting or under-predicting.