P(X) Model Backtest
Three projection models compared against actual January & February 2026 data. V1 pipeline P(X) only → V2 + flat inflow forecast → V3 recency CDFs + seasonal inflow + dynamic correction.
Projection Error — All Models
Distance from final actual. Closer to zero = better.
Active Pipeline vs. Closed
Pipeline shrinkage as loans close through the month.
V1 Pipeline-Only: Proj = Funded + Σ P(close) using all-time CDFs (2022–2026, 16.5K loans), calendar days
V2 + Flat Inflow: Proj = V1 + 141 × (1 − entered%) — 6-month trailing avg of same-month closers, one number for every month
V3 + Recency + Seasonal + Dynamic:
• Recency CDFs — trailing 12-month data only (~6K loans), reflects current processing speed (+3–4% faster)
• Seasonal inflow — trailing 3-month avg × seasonal ratio (Jan: 103, Feb: 99 vs V2's flat 141)
• Dynamic correction — after day 7, compares actual vs expected inflow pace, blends adjustment (capped ±50%, max 40% blend)