5D Library › 弃单挽回动力学、沉没数据续费棘轮与高客单顾问筛选机制 · English

第一章 弃单跟进的时序与次数

一个会自己衰减的回访强度

买家把商品放进购物车却没有付款,之后有一部分人会自己回来。把「此刻回来下单」的瞬时概率写成回访强度 h0(t)=ae−t/τh_0(t) = a\,e^{-t/\tau}:刚离开时最高,随兴趣淡去而下降。每一封跟进在发出时刻 ss 给强度加上一个衰减的脉冲 b(s)e−(t−s)/κb(s)\,e^{-(t-s)/\kappa},其大小 b(s)b(s) 随买家离开越久而越小;第 kk 封因习惯化而打折,两封靠得太近时后一封的效果也打折。回访在 7 天(168 小时)之内计算。

这个模型常被称为「霍克斯过程」,但并不准确。霍克斯过程描述的是事件会激发更多同类事件的自我激发过程 (Hawkes 1971);而回来下单是吸收事件,一个人下单之后就不再「回访」。这里的强度是被跟进从外部激发的,借用的只是指数核的形式。

推导

# Chapter 1. p01: When to send three follow-ups after an abandoned cart.
# Natural return of an abandoner has hazard h0(t) = a exp(-t/tau) (hours).
# A follow-up sent at time s adds an excitation b(s) exp(-(t-s)/kappa),
# where b(s) = b0 exp(-s/rho) falls as interest fades, and the k-th
# message is weaker by a factor f^k (habituation) and by 1-exp(-gap/G)
# when it follows the previous one too closely. Return is absorbing,
# so this is an externally excited intensity, not a self-exciting Hawkes
# process. We compare the schedule 0.25 h / 24 h / 72 h with the best
# three times on a grid, counting returns within 7 days (168 h).
from itertools import combinations
from math import exp

A, TAU = 0.02, 12.0          # natural return: about 21% eventually
B0, RHO, KAPPA, F, G = 0.08, 30.0, 2.0, 0.8, 12.0
HORIZON, DT = 168.0, 0.05
GRID = (0.25, 1, 2, 4, 8, 12, 24, 36, 48, 72, 96, 120)

def weights(times):
    w, prev = [], None
    for k, s in enumerate(times):
        g = 1.0 if prev is None else 1 - exp(-(s - prev) / G)
        w.append(F ** k * g * B0 * exp(-s / RHO))
        prev = s
    return w

def returned(times, horizon=HORIZON, dt=DT):
    times = sorted(times)
    ws = weights(times)
    cum, t = 0.0, 0.0
    while t < horizon:
        lam = A * exp(-t / TAU)
        for w, s in zip(ws, times):
            if t >= s:
                lam += w * exp(-(t - s) / KAPPA)
        cum += lam * dt
        t += dt
    return 1 - exp(-cum)

def curve(times, step=1.0):
    out, cum, t = [], 0.0, 0.0
    times = sorted(times)
    ws = weights(times)
    while t <= HORIZON:
        out.append((t, 1 - exp(-cum)))
        sub = 0.0
        for i in range(20):
            u = t + (i + 0.5) * step / 20
            lam = A * exp(-u / TAU)
            for w, s in zip(ws, times):
                if u >= s:
                    lam += w * exp(-(u - s) / KAPPA)
            sub += lam * step / 20
        cum += sub
        t += step
    return out

if __name__ == "__main__":
    base = returned(())
    rule = returned((0.25, 24, 72))
    print("   natural return within 7 days: {:.3f}".format(base))
    print("   schedule 0.25 h / 24 h / 72 h: {:.3f}  (+{:.1f} points)".format(
        rule, 100 * (rule - base)))
    scores = {c: returned(c) for c in combinations(GRID, 3)}
    best = max(scores, key=scores.get)
    rb = scores[best]
    print("   best grid schedule {} h: {:.3f}  (+{:.1f} points)".format(
        "/".join("{:g}".format(x) for x in best), rb, 100 * (rb - base)))
    print("   the rule reaches {:.0%} of the best gain".format(
        (rule - base) / (rb - base)))
    near = sum(1 for v in scores.values() if v - base >= 0.9 * (rb - base))
    print("   {} of {} grid schedules reach 90% of the best gain".format(
        near, len(scores)))
#    natural return within 7 days: 0.214
#    schedule 0.25 h / 24 h / 72 h: 0.369  (+15.5 points)
#    best grid schedule 0.25/12/24 h: 0.384  (+17.0 points)
#    the rule reaches 91% of the best gain
#    33 of 220 grid schedules reach 90% of the best gain

在示意参数下,完全不跟进时 7 天内有 0.214 的买家自己回来。按 0.25 小时、24 小时、72 小时发三封,回访率升到 0.369,多出 15.5 个百分点;在 12 个候选时刻中挑三个的网格最优是 0.25 小时、12 小时、24 小时,回访率 0.384,多出 17.0 个百分点。经验法则达到最优增益的 91%,而且 220 个候选时序中有 33 个都达到最优增益的 90%。

回访比例随时间的变化:不跟进、经验时序与网格最优时序(p01)。

注

实务含义。 最优面很平:精确到小时的时序并不重要,重要的是第一封尽早发出,因为买家的兴趣与跟进的效果都随时间衰减。第三封放在 72 小时,在这个模型中贡献很小。

该发几封

推导

# Chapter 1. p02: How many follow-ups? Marginal recovery against the
# cost of unsubscribes. Messages go out at 0.25 h, 12 h, 24 h, 48 h,
# 72 h, 120 h (in that order, the first k used). Each message makes a
# share u of recipients unsubscribe; an unsubscribed address is worth
# V in future revenue. Per 1,000 abandoners, order value RM 98.
import importlib.util

def load(name):
    spec = importlib.util.spec_from_file_location(name, name + ".py")
    m = importlib.util.module_from_spec(spec)
    spec.loader.exec_module(m)
    return m

p01 = load("p01_return_hazard")
TIMES = (0.25, 12, 24, 48, 72, 120)
ORDER, U, V = 98.0, 0.004, 60.0

def ledger(k, n=1000):
    base = p01.returned(())
    rec = p01.returned(TIMES[:k]) - base
    gain = n * rec * ORDER
    lost = n * (1 - (1 - U) ** k) * V
    return rec, gain, lost

if __name__ == "__main__":
    prev = 0.0
    best, best_net = 0, 0.0
    print("   k  recovered  marginal RM  unsub cost RM  net RM")
    for k in range(1, len(TIMES) + 1):
        rec, gain, lost = ledger(k)
        net = gain - lost
        print("   {}  {:9.3f}  {:11,.0f}  {:13,.0f}  {:6,.0f}".format(
            k, rec, gain - prev, lost, net))
        prev = gain
        if net > best_net:
            best, best_net = k, net
    print("   best number of messages: {} (net RM {:,.0f})".format(
        best, best_net))
#    k  recovered  marginal RM  unsub cost RM  net RM
#    1      0.117       11,436            240  11,196
#    2      0.152        3,464            479  14,421
#    3      0.170        1,804            717  15,987
#    4      0.179          846            954  16,596
#    5      0.182          301          1,190  16,661
#    6      0.183           55          1,426  16,481
#    best number of messages: 5 (net RM 16,661)

每多发一封,就有一小部分收件人退订,而一个退订的地址失去了它未来的价值。按 0.25、12、24、48、72、120 小时依次增加:第一封每 1,000 名弃单者带来 RM 11,436,第二封 RM 3,464,第三封 RM 1,804,第四封 RM 846,第五封 RM 301,第六封只有 RM 55;退订成本则每封约增加 RM 240。扣除退订后净收入在五封时最高(RM 16,661),但第四、第五封合计只多出很少。

挽回收入与扣除退订后的净收入随跟进封数的变化(p02)。

假说 · 尚未验证

作者的假说(尚未验证): 在作者的报告产品中,第一封在 1 小时内发出、共发三封的跟进,与第一封在 24 小时发出的跟进相比,7 天回访率至少高出 5 个百分点。什么会证明它错: 随机实验中两种时序的回访率差距小于 5 个百分点。

小结

经验时序「15 分钟、24 小时、72 小时」在示意模型中达到最优增益的 91%,最优面很平。跟进的边际收益迅速递减;第四封以后几乎只剩退订的代价。

参考文献

Hawkes, Alan G. 1971. “Spectra of Some Self-Exciting and Mutually Exciting Point Processes.” Biometrika 58 (1): 83–90.

Fan, Y. (2026). Cart-Abandonment Recovery Kinetics, Sunk-Data Retention Ratchets and High-Ticket B2B Qualification Mechanisms: Multi-Pulse Follow-Up Timing, Switching-Cost Accumulation and Screening Equilibria [弃单挽回动力学、沉没数据续费棘轮与高客单顾问筛选机制] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090209 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).