METHOD
Method and limits
The exact formulas of the demo model. Nothing is hidden, and what is missing is stated.
01 —
The formulas
cadence: g = (k·g_obs + w·g_prior) / (k + w), with k = n − 1, w = 3, g_prior = 120 days activity: p = 1 / (1 + exp((d − 2.5·g) / (0.6·g))), d = days since last order expected orders: λ = p · H / g, H = 365 days margin per order: m = basket · margin rate · (1 − return rate) − €8 net future value: NFV = max(0, λ · m · f), f = (1 − e^(−r·H)) / (r·H), r = 10 % a year 80 % band: 10th and 90th percentiles of a gamma law of shape n + 1, scaled to the NFV (Wilson–Hilferty approximation) cost ceiling = expected effect × lower bound of the NFV
This is the demo’s own code: it runs in your browser and can be read in this site’s js/model.js file.
02 —
The guardrails
Four guardrails, evaluated in order. The first one crossed is flagged as the deviation; the rest are listed without emphasis.
- history: at least 2 orders
- margin after action ≥ margin floor (12 % by default), where margin after action = (basket · rate · (1 − returns) − €8 − cost) / basket
- proposed cost ≤ cost ceiling
- proposed cost ≤ maximum budget per customer (€400 by default)
03 —
Default parameters
| PARAMETER | VALUE |
|---|---|
| Horizon | 365 j / 365 d |
| Discount rate | 10 % |
| Cadence prior | 120 j / 120 d |
| Prior weight | 3 |
| Variable cost per order | 8 € |
| Expected effect of the action | 10 % |
| Margin floor | 12 % |
| Maximum budget | 400 € |
| Minimum history | 2 |
04 —
What the demo cannot do
- The model is simplified: one cadence per customer, a logistic activity probability, a gamma band. It knows neither seasonality, nor the range bought, nor the real logistics and support costs.
- The expected effect of the action (10 %) is a demo assumption, not a measurement. It is calibrated on real experiments.
- Sample customers are fictitious. No result on this site describes a real customer.
- The production engine adds what this model ignores: gateway-product archetypes, forward and reverse logistics costs, support costs, acquisition cost, batch calibration of the band, and a key-account tier.