This is the 2D top-down companion to the 3D funnel sim: the same lead-scoring mechanism, viewed from directly above so the three funnel rings read as concentric circles instead of stacked disks. Each simulated lead has two hidden features — engagement (0–1) and budget fit (0–1). The true chance a lead converts follows a fixed logistic function nobody gets to see directly:
P(convert | lead) = σ(1.3·engagement + 0.9·budget − 1.1 + noise)
σ(z) = 1 / (1 + e^(−z))
Your predictive-analytics model only gets to estimate that probability from the two weights you control:
score(lead) = σ(w₁·engagement + w₂·budget − 1.1)
- w₁, w₂ sliders — the feature weights your model assigns; the closer they track the true (unknown) 1.3 / 0.9, the more the model's ranking matches reality.
- Pursue threshold — leads scoring above it get sales follow-up (moved to the inner ring); leads below it stay on the outer ring. A low threshold chases everyone (wastes effort on weak leads); a high one only chases the safest bets (misses real opportunities).
- Model quality — balanced accuracy: the average of how many true converters were correctly pursued (sensitivity) and how many true non-converters were correctly skipped (specificity). A model with badly mismatched weights hovers near 50% — no better than a coin flip.
- Each quarter, revenue = converted leads × average deal size. Compounded quarter over quarter, better weight tuning translates directly into faster business growth — the orange bar chart on the right tracks cumulative revenue since Q1.
This mirrors real predictive-analytics deployments: the model never sees ground truth, only correlated features — the whole discipline is choosing weights (via regression, gradient boosting, etc.) that track the hidden outcome as closely as possible.