What does Pareto Whales and VIP Tiers conclude, in numbers, about “Who Will Be Next Year's Whale: BG/NBD and the Prediction of High Value”?

Answer: BG/NBD models the purchase rate and silent attrition separately and estimates its parameters from three summaries: the number of repeat purchases, the time of the last purchase and the length of observation. On the synthetic book, two of the four parameters are estimated closely and two drift because the likelihood surface is flat. At the level of the whole book the conditional forecast differs from the simulation by about four per cent, and the hit rate for next year's top 2 % is 0.625, above the 0.450 of the spend ranking.

Scope of the numbers: Parameters are illustrative and all calculations are generated by the book's programs. This is general business analysis, not personalised legal, tax or financial advice, and it promises no income to anyone.

中文版

Data table

The numbers below are printed verbatim by the chapter's Python program and can be reproduced by running it. standard-library Python onlyprinted values that match the answer: 0.450, 0.625

top 40 of 2,000 customers; year-2 revenue 104,253
oracle top-40 share of year-2 revenue 0.525
ruleprecisionshare of year-2 revenue
year-1 spend0.4500.382
repeat count0.2500.257
BG/NBD value0.6250.445
BG/NBD gain over year-1 spend 0.175 (bar for upkeep 0.1)
Show the full program
# Chapter 3. p33: which ranking rule finds next year's top 2 percent?
# Model: 2,000 synthetic BG/NBD customers (parameters of p31) with a
# personal mean basket m ~ lognormal(3.0, 1.0) and basket noise
# exp(N(-0.08, 0.4)); the first purchase is at week 0. Rules rank on
# year-1 data only: (1) year-1 spend, (2) repeat count x, (3) BG/NBD
# expected purchases times a shrunk average basket (weight 2 on the
# book mean). Truth is year-2 spend. Precision = overlap of the top 40.
import math
import random

R, AL, A, B = 0.6, 4.0, 0.8, 2.5


def simulate(n, rng):
    book = []
    for _ in range(n):
        lam = rng.gammavariate(R, 1.0 / AL)
        p = rng.betavariate(A, B)
        m = math.exp(rng.gauss(3.0, 1.0))
        s1 = [m * math.exp(rng.gauss(-0.08, 0.4))]
        s2 = []
        t, x, tx = 0.0, 0, 0.0
        while True:
            t += rng.expovariate(lam)
            if t > 104.0:
                break
            v = m * math.exp(rng.gauss(-0.08, 0.4))
            if t <= 52.0:
                x, tx = x + 1, t
                s1.append(v)
            else:
                s2.append(v)
            if rng.random() < p:
                break
        book.append((x, tx, s1, math.fsum(s2)))
    return book


def hyp2f1(a, b, c, z):
    terms, term, k = [1.0], 1.0, 0
    while abs(term) > 1e-15 and k < 5000:
        term *= (a + k) * (b + k) / ((c + k) * (k + 1)) * z
        terms.append(term)
        k += 1
    return math.fsum(terms)


def expect(x, tx, T, t):
    f = hyp2f1(R + x, B + x, A + B + x - 1, t / (AL + T + t))
    num = (A + B + x - 1) / (A - 1) * (
        1.0 - ((AL + T) / (AL + T + t)) ** (R + x) * f)
    den = 1.0
    if x > 0:
        den += A / (B + x - 1) * ((AL + T) / (AL + tx)) ** (R + x)
    return num / den


def top(scores, k):
    idx = sorted(range(len(scores)), key=lambda i: (-scores[i], i))
    return idx[:k]


if __name__ == "__main__":
    rng = random.Random(39933)
    book = simulate(2000, rng)
    k = 40
    y2 = [b[3] for b in book]
    total2 = math.fsum(y2)
    avgs = [math.fsum(b[2]) / len(b[2]) for b in book]
    mbar = math.fsum(avgs) / len(avgs)
    rules = {
        "year-1 spend": [math.fsum(b[2]) for b in book],
        "repeat count": [b[0] + 1e-6 * math.fsum(b[2]) for b in book],
        "BG/NBD value": [expect(b[0], b[1], 52.0, 52.0) *
                         (2.0 * mbar + math.fsum(b[2])) / (2.0 + len(b[2]))
                         for b in book],
    }
    truth = set(top(y2, k))
    print("top {} of {:,} customers; year-2 revenue {:,.0f}".format(
        k, len(book), total2))
    print("oracle top-40 share of year-2 revenue {:.3f}".format(
        math.fsum(y2[i] for i in truth) / total2))
    print("rule           precision  share of year-2 revenue")
    prec = {}
    for name in ("year-1 spend", "repeat count", "BG/NBD value"):
        pick = top(rules[name], k)
        prec[name] = len(truth.intersection(pick)) / k
        print("{:<14} {:>9.3f}  {:>10.3f}".format(
            name, prec[name], math.fsum(y2[i] for i in pick) / total2))
    print("BG/NBD gain over year-1 spend {:.3f} (bar for upkeep 0.1)".format(
        prec["BG/NBD value"] - prec["year-1 spend"]))

Single source and citation

Fan, Y. (2026). Pareto Whale Auto-Detection, High-Net-Worth Self-Selection Signals and Ten-X VIP Tier Escalation (Version 1.0.0), Chapter 3: Who Will Be Next Year's Whale: BG/NBD and the Prediction of High Value. 5D Library. DOI pending

DOI pending · monograph page goes live after publication · 5DL-B-399 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026399,
  author    = {Fan, Yafu},
  title     = {Pareto Whale Auto-Detection, High-Net-Worth Self-Selection Signals and Ten-X VIP Tier Escalation},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

Other chapter answers in this book