Blend Your Small-Group Experience with Industry Benchmarks—Know When Your Data is Too Thin to Trust
python# Limited Fluctuation Credibility (Bühlmann-Straub) import math def calculate_credibility_factor(exposure, expected_claims=1082, k=1000): """ Calculate credibility Z based on exposure size exposure: Number of member-months expected_claims: Expected number of claims k: Full credibility standard (typically 1,000 - 1,500 claims) """ # Square root rule for partial credibility if exposure < 12: # Less than 1 year return 0.0 z = math.sqrt(expected_claims / k) # Cap at 1.0 (full credibility) return min(z, 1.0) def credibility_weighted_estimate(own_experience, benchmark, credibility_z): """ Blend own experience with industry benchmark """ weighted_estimate = (credibility_z * own_experience) + ((1 - credibility_z) * benchmark) return { 'own_experience': own_experience, 'benchmark': benchmark, 'credibility_factor': credibility_z, 'weighted_estimate': weighted_estimate, 'own_weight_pct': credibility_z * 100, 'benchmark_weight_pct': (1 - credibility_z) * 100 } # Credibility by Group Size # (Assuming 1.2 claims PMPM avg) # 50 lives (600 member-months/year): # Expected claims: 720 # Credibility: sqrt(720/1082) = 81.5% → 0.815 # Estimate: 81.5% own + 18.5% benchmark # 125 lives (1,500 member-months): # Expected claims: 1,800 # Credibility: sqrt(1800/1082) = 129% → capped at 1.00 (full credibility) # Estimate: 100% own experience # 25 lives (300 member-months): # Expected claims: 360 # Credibility: sqrt(360/1082) = 57.7% → 0.577 # Estimate: 57.7% own + 42.3% benchmark # Example Application: # 75-life group (900 member-months) # - Own PMPM: $1,450 (one large claim spike) # - Industry benchmark: $950 # - Credibility: sqrt(1080/1082) = 99.9% ≈ 1.00 # - Weighted: (1.00 × $1,450) + (0.00 × $950) = $1,450 # # BUT if we use Large Claim Pooling: # - Remove claims >$100K, recalculate # - Own PMPM (pooled): $875 # - Credibility on pooled: 1.00 # - Weighted: $875 (the volatile spike is removed)
Apply actuarial credibility theory to small groups. Blend your experience with industry benchmarks. Smooth random noise, preserve real signals, forecast with confidence.
Apply Credibility Theory→