What does Cold-Start Price Testing conclude, in numbers, about “Contextual Bandits: Market, Channel, New versus Returning, and the Limits of Segment Pricing”?

Answer: Writing market, channel and new versus returning as context brings prices closer to each group's price sensitivity, but at launch scale the gain is limited: in Beian Studio's fictional sample, the segmentation ceiling is 1.095 times the single price, 5,928 yuan more over six thousand visitors, and the best contextual rule earns only 879 yuan more than knowing the best single price from day one. LinUCB's theoretical α is absurdly wide and must be scaled to reward noise; logistic Thompson sampling fits a shop's batch reality better. The most important finding is that the algorithms teach themselves to charge returning customers more, and that in this sample doing so earns almost nothing: charging new and returning customers alike loses only 0.005 of revenue. So the limits sit outside the algorithm: same customer, same time, same price, guaranteed by hash assignment; consistent display prices, with channel prices carrying public conditions and a capped spread; only public discounts between new and returning customers, never a loyalty penalty; and a former price in a time-limited offer must be a price that was really applied. The next chapter turns to repricing over time.

Scope of the numbers: Laws on price display and segment pricing are described only in general terms and are not legal or tax advice. Marked boxes separate published evidence, runnable calculations, sources and the author's hypotheses, each stating what would refute it; the source studied is the "Guoxu" chapter of the Guanzi on gathering and releasing goods and the "level" price.

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: 0.005, 5928

menuyuan/visitorshare of V0lost vs V0prices
V0 eight cell prices11.3861.0000.000[99, 129, 169, 219]
V1 market x channel11.3270.9950.005[219, 129, 129, 99]
V2 V1 + spread cap 1.3510.9100.9580.042[169, 129, 129, 129]
V3 one price per market10.8510.9530.047[169, 99]
V4 one price for all10.3980.9130.087[129]
V5 list + new-customer coupon11.0180.9680.032[219, 99, 40, 10]
without the 0.20 coupon limit V5 would pick [219, 129, 60, 40] for 11.100 per visitor
V5 prices: [list A, list B, coupon A, coupon B]; V1/V2: A/site, A/community, B/site, B/community
over 6,000 visitors: V0 - V4 = 5,928 yuan; V1 - V4 = 5,572; V5 - V4 = 3,715
Show the full program
# Chapter 5. p53: What each fairness rule costs, in yuan per visitor, on the fictional truth of p50.
# Menus compared (every price must be one of the arms or, for the coupon menu, an arm minus a coupon):
#   V0 eight cell prices, new and returning priced apart (no limits)
#   V1 one price per market x channel, new and returning see the same price
#   V2 as V1, and the dearest shown price at most 1.35 x the cheapest
#   V3 one price per market (both channels show the same price inside a market)
#   V4 one price for everyone
#   V5 one list price per market, plus a disclosed first-order coupon for new customers (steps of 10 yuan,
#      at most 0.20 of the list price, so the list price stays the price most buyers could really pay)
# The coupon menu is honest only if returning customers really pay the list price. EU readers: personalised
# prices must be disclosed under Directive 2019/2161; this program is arithmetic, not legal advice.
import itertools
import math

ARMS = [99, 129, 169, 219]
COUPONS = [0, 10, 20, 30, 40, 50, 60]
COUPON_MAX = 0.20
CAP = 1.35
MS, CS, RS = 0.6, 0.7, 0.75


def conv(m, c, r, p):
    k = 0.22 + 0.18 * m + 0.20 * c - 0.10 * r
    z = -2.4 - 0.2 * m + 0.3 * c + 0.6 * r - k * (p - 99) / 40
    return 1 / (1 + math.exp(-z))


def w(m, c, r):
    return (MS if m == 0 else 1 - MS) * (CS if c == 0 else 1 - CS) * (RS if r == 0 else 1 - RS)


CELLS = list(itertools.product((0, 1), repeat=3))


def value(price_of):
    return math.fsum(w(*x) * price_of(x) * conv(*x, price_of(x)) for x in CELLS)


rows = []
v0 = {x: max(ARMS, key=lambda p: p * conv(*x, p)) for x in CELLS}
rows.append(("V0 eight cell prices", value(lambda x: v0[x]), sorted(set(v0.values()))))
pairs = list(itertools.product((0, 1), repeat=2))
best1, best2 = None, None
for combo in itertools.product(ARMS, repeat=4):
    menu = dict(zip(pairs, combo))
    v = value(lambda x: menu[(x[0], x[1])])
    if best1 is None or v > best1[0]:
        best1 = (v, combo)
    if max(combo) <= CAP * min(combo) and (best2 is None or v > best2[0]):
        best2 = (v, combo)
rows.append(("V1 market x channel", best1[0], list(best1[1])))
rows.append(("V2 V1 + spread cap {:.2f}".format(CAP), best2[0], list(best2[1])))
best3 = max(((value(lambda x: pa if x[0] == 0 else pb), (pa, pb)) for pa in ARMS for pb in ARMS))
rows.append(("V3 one price per market", best3[0], list(best3[1])))
best4 = max((value(lambda x: p), p) for p in ARMS)
rows.append(("V4 one price for all", best4[0], [best4[1]]))


def coupon_menu(la, lb, ca, cb):
    def price(x):
        lst, cp = (la, ca) if x[0] == 0 else (lb, cb)
        return lst - cp if x[2] == 0 else lst
    return value(price)


best5, unc = None, None
for la, lb, ca, cb in itertools.product(ARMS, ARMS, COUPONS, COUPONS):
    v = coupon_menu(la, lb, ca, cb)
    if unc is None or v > unc[0]:
        unc = (v, [la, lb, ca, cb])
    if ca > COUPON_MAX * la or cb > COUPON_MAX * lb:
        continue
    if best5 is None or v > best5[0]:
        best5 = (v, [la, lb, ca, cb])
rows.append(("V5 list + new-customer coupon", best5[0], best5[1]))
top = rows[0][1]
print("menu                              yuan/visitor   share of V0   lost vs V0   prices")
for name, v, prices in rows:
    print("{:32s}  {:12.3f}   {:11.3f}   {:10.3f}   {}".format(name, v, v / top, 1 - v / top, prices))
print("without the 0.20 coupon limit V5 would pick {} for {:.3f} per visitor".format(unc[1], unc[0]))
print("V5 prices: [list A, list B, coupon A, coupon B]; V1/V2: A/site, A/community, B/site, B/community")
print("over 6,000 visitors: V0 - V4 = {:,.0f} yuan; V1 - V4 = {:,.0f}; V5 - V4 = {:,.0f}".format(
    6000 * (top - best4[0]), 6000 * (best1[0] - best4[0]), 6000 * (best5[0] - best4[0])))

Single source and citation

Fan, Y. (2026). Zero-History Cold-Start Price Testing, Van Westendorp Sensitivity Bounds and Bayesian Adaptive Repricing (Version 0.1.0), Chapter 5: Contextual Bandits: Market, Channel, New versus Returning, and the Limits of Segment Pricing. 5D Library. DOI pending

DOI pending · monograph page goes live after publication · 5DL-B-133 · v0.1.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026133,
  author    = {Fan, Yafu},
  title     = {Zero-History Cold-Start Price Testing, Van Westendorp Sensitivity Bounds and Bayesian Adaptive Repricing},
  year      = {2026},
  publisher = {5D Library},
  edition   = {0.1.0},
  note      = {DOI pending}
}

Other chapter answers in this book