We study how a seller can learn a reserve price for a second-price auction from censored historical bid data. Because values below past reserves are unobserved, even unlimited data may not reveal which reserve earns the most revenue. Under a monotone hazard rate (MHR) assumption, we characterize exactly when an optimal reserve is identified, even though the value distribution may not be: either revenue has not yet begun to fall at the lowest historical reserve, so the best reserve lies among the observed bids, or the censored fraction is as large as MHR allows, which pins down the entire distribution. When identification fails, we derive the reserve that minimizes worst-case revenue regret over distributions consistent with the data, and prove that this unique minimax reserve and its regret depend on just two observable quantities: the censored fraction and the hazard just above the lowest historical reserve. With N historical bidders, the optimal reserve is learned with expected worst-case regret O(N−1/2) when it lies among the observed bids, and O(N−1) from the censored count alone when the censored fraction is as large as MHR allows and the optimum lies at or below the lowest reserve. In the nonidentifiable regime, plugging estimates of the two quantities into the minimax reserve gives expected worst-case regret within O(N−2/5) of the population minimum under local density smoothness, with a matching minimax lower bound on a smooth class. Beyond MHR, we solve the minimax problem for regular and unrestricted distributions and prove an O(N−2/5) learning bound in the regular nonidentifiable regime, given known bounds on the boundary parameters. Experiments on synthetic and eBay data show gains from exploiting MHR and the benefits of the robust rule when parametric models are misspecified.
Figures & tables
Figure 1.1 : A nonidentifiable instance. Two value distributions agree at and above the historical reserve λ=1 , so they induce the same distribution of censored records. Their unique optimal reserves are 2/3 and 1/2 . The seller observes the total probability below the boundary, but not its distribution within the shaded region. Both distributions have a monotone hazard rate (see Section 3 ).
Figure 3.1 : The observed tail fixes the endpoint (λ,A) and caps the body’s slope at h0 . Every MHR body is a convex nondecreasing curve between the chord and the backward tangent. The bounds coincide in the fully identified case, A=λh0 .
Figure 4.1 : Straight-line replacements of a compatible cumulative hazard. When the alternative reserve is higher, two chords raise the curve between the reserves. When it is lower, two extended lines lower it. Each replacement can only increase the alternative reserve’s revenue advantage and gives a body hazard with at most one upward jump. Dashed: the original curve; gray: the band of Figure 3.1 ; the shaded strip lies between the two reserves.
Figure 4.2 : The two worst-case losses cross at r†≈0.621 . Raising the reserve beyond this point increases the loss from posting too high; lowering it increases the loss from posting too low. Their common value is Δ≈0.00524 for λ=1,A=1.5,h0=2,n=3 .
Regime
Case
Boundary condition
Rule
Regret rate
Identifiable
Tail-optimal
λh0≤1
Empirical tail
O(N−1/2)
Fully identified
A=λh0≥1
Linear completion
O(N−1)
Nonidentifiable
—
A<λh0 , λh0>1
Plug-in
O(N−2/5)
Table 5.1: Two regimes, three cases: expected excess worst-case regret of the corresponding rules under MHR, under the conditions stated below. At A=λh0=1 , the linear-completion guarantee also applies.
MHR
Regular
Unrestricted
Instance
λ
Worst-case
Actual
Worst-case
Actual
Worst-case
Actual
Rayleigh (1)
1.5
1.65%
0.20%
4.66%
0.49%
22.07%
0.24%
Weibull (1.5,1)
1
0.80%
0.18%
3.98%
0.05%
18.41%
0.56%
Weibull (2.5,1.5)
2
3.37%
0.19%
9.74%
<0.01%
40.09%
0.58%
Table 7.1: Regret of the minimax reserves under three distribution classes ( n=3 ), with exact censored fraction and observed tail. Worst-case regret ranges over compatible distributions in the indicated class; actual regret uses the listed distribution. All percentages use the listed distribution’s RG∗ . Appendix D.6 gives the reserves and computation details.
Figure 7.1 : Learning the robust reserve ( n=3 ). (a) Mean excess worst-case regret WG−Δ in revenue units, over 200 replications. (b) Mean worst-case regret in percent of RG∗ for Rayleigh (1) , with N=2000 and 80 replications per λ , on a linear axis capped at 6% ; the linear-completion rule exceeds it at λ=1.44 (its peak, 9.1% ), 1.56 , 2.28 and 2.40 . Bars show one Monte-Carlo standard error; some are smaller than the markers, and the linear-completion rule in (a) is drawn without them.
Worst-case regret
Actual regret under each distribution
Rule
WG
Weibull G
Equalizer
Exponential
Tangent
Plug-in, Auto-selector
1.46
0.32
0.00
1.36
1.35
Plug-in (Weibull hazard)
1.43
0.31
0.00
1.36
1.34
Weibull MLE
2.49
0.00
0.12
2.49
0.43
Gamma MLE
2.97
0.04
0.22
2.97
0.22
Lognormal MLE
3.17
0.07
0.26
3.17
0.16
Table 7.2: Four distributions with indistinguishable censored records. Mean worst-case and actual regret in percent of the original Weibull distribution’s RG∗ ( N=3000 , 200 paired replications). Actual regret compares each distribution with its own optimal revenue. Numerical minimax regret is Δ=1.35% ; Monte-Carlo standard errors are at most 0.02 percentage points.
Censored MLE
Item
λ
N
Full-info
Plug-in
Auto-selector
Weibull
Gamma
Lognormal
Linear compl.
Boundary
Drop
Palm Pilot
155
1156
0.08
1.39
1.44
2.26
2.61
2.77
1.64
1.64
2.34
Palm Pilot
199
1156
0.08
1.68
1.68
1.52
1.30
1.25
3.64
10.84
11.14
Palm Pilot
219
1157
0.08
0.63
0.63
1.05
1.41
1.90
1.53
27.73
27.98
Table 7.3 : Held-out regret on the recorded Palm Pilot bids , in percent of each split’s test optimum ( n=3 , 1000 bidder splits, 20,000 test auctions per split; Monte-Carlo standard errors at most 0.13 points). N is the mean number of training records. Full-info fits a reserve to the uncensored training records.
Worst-case
Held-out
Item
λ
K
m
N
α
G(λ)
Regime
Detected
Δ
Auto-selector
Weibull
Auto-selector
Weibull
Boundary
Palm Pilot
10
32
347
400
0.13
0.02
ident.
ident.
0
0.03
1.07
0.15
2.52
7.56
Palm Pilot
50
9
73
113
0.35
0.14
ident.
nonident.
0
4.82
1.57
6.88
3.39
6.84
Palm Pilot
100
9
62
113
0.45
0.31
ident.
ident.
0
0.06
2.19
0.04
4.18
3.57
Palm Pilot
150
14
90
175
0.49
0.48
nonident.
nonident.
0.14
0.94
2.38
1.17
2.32
0.04
Palm Pilot
175
39
145
488
0.70
0.58
nonident.
nonident.
1.65
3.00
3.89
2.52
2.95
2.41
Table 7.4 : Sellers’ own opening bids on Palm Pilot ( n=3 ). K auctions with opening bid within one cent of λ , m records, N=12.5K inferred bidders, and α=1−m/N ; G(λ) and the regime are those of the refit, and Detected is the auto-selector’s branch. Weibull is the censored Weibull MLE. Worst-case regret WG in percent of RG∗ and held-out regret on 20,000 auctions among reference bidders in percent of the test optimum, means over 50 draws of the winner imputation (standard errors at most 0.26 points, conditional on the recorded auctions); Δ in percent of RG∗ .
Appendix figures & tables27 assets
Supplementary material from the paper’s appendix.
Appendix
Symbol
Meaning
n , N
Number of bidders in the target auction; number of historical records subject to censoring at the lowest reserve λ , one per bidder who faced it.
G , F , λ
True distribution, a compatible completion, and the censoring boundary (the lowest historical reserve).
F(λ;G)
Ambiguity set: all completions of the observed tail; subscripts restrict the shape.
hF , HF
Hazard and cumulative hazard; MHR means HF convex.
ϕF , ΦF
Virtual value and single-bidder revenue curve in quantile coordinates; regularity means ΦF concave.
RF(r) , RF∗
Second-price auction revenue at reserve r , and its optimal value.
Appendix
Table A.1 : Recurring notation. Boundary quantities belong to the observed tail and are common to all compatible completions.
Figure A.1 : Pointwise CDF ranges for unrestricted, regular, and MHR completions of the same observed tail. A curve inside a shaded band must still satisfy its class’s shape constraint; unattained unrestricted boundaries are limits of absolutely continuous completions.
Figure A.2 : Replacing a convex cumulative hazard by a two-level body. A higher alternative reserve uses two chords to raise the curve between the reserves; a lower one uses the maximum of two extended chords to lower it. The preserved heights make each replacement improve nature’s revenue comparison.
Figure C.1 : The regular families share boundary point (qλ,c0) , with c0=λqλ , and slope cap ϕ0=λ−1/h0 . Gray shading gives the pointwise feasible range; curves must also be concave. Ss gives the worst loss against lower alternative reserves, and Et against higher ones. The price ray Φ=rq intersects each displayed curve at the seller’s quantile; a reserve below a curve’s lower support endpoint has quantile 1 . Moving right, toward larger survival quantiles, means moving to lower bidder values.
Rule
How the reserve is chosen
Full-info
The full-information reserve of the held-out eBay study: the exact revenue maximizer on simulated auctions among the uncensored training bidders. It sees the values that censoring hides from every other rule and serves as the oracle row.
Plug-in
Algorithm 5.1 , at the projected boundary estimates.
Auto-selector
Algorithm 5.2 : empirical tail rule on [λ,10λ] (on [λ,rmax] with rmax the larger of 4λ and twice the largest observed record in the opening-bid study of Section 7.4 ) when the raw h0≤1/λ ; plug-in otherwise.
Plug-in (Weibull hazard)
Algorithm 5.1 with the boundary hazard replaced by the censored Weibull fit’s hazard at λ ; the count estimate AN is kept.
Censored MLE
Maximum likelihood fit in a Weibull, gamma (shape at least one) or lognormal family. Observed bids contribute their fitted density; no-bid records contribute the fitted probability below λ . Post the fitted distribution’s optimal reserve. It is an oracle MLE when the full distribution belongs to the fitted family.
Linear completion
The linear-completion reserve rc(AN) of ( 5.3 ): it uses only the censored fraction and is optimal when the body hazard is constant below λ .
Appendix
Table D.1: Reserve rules compared in Section 7 and Appendix D .
Figure D.1 : Mean actual regret under 441 two-level completions and the equalizer completion of Weibull (2.5,1) , censored at λ=(log2)2/5 ( n=3 , N=3000 , 200 paired replications). Each completion is compared with its own optimal reserve; percentages divide by RG∗ . Actual regret can fall below Δ .
Figure D.2 : Mean actual regret under three compatible completions of Weibull (2.5,1) , censored at its median ( n=3 , 200 paired replications per N ). Percentages use RG∗ . Each rule posts the same reserves across panels. Plug-in (Weibull hazard) is the plug-in with a fitted Weibull boundary hazard. The right panel has a linear axis. Bars show one Monte-Carlo standard error. Actual regret can fall below Δ .
Item
λ
G(λ)
A
h0
λh0
Regime
RG∗
rG∗
Δ
r†
Palm Pilot
155
0.50
0.692
0.0080
1.23
nonident.
156.74
145.19
0.45%
149.02
Palm Pilot
199
0.70
1.202
0.0169
3.35
nonident.
156.74
145.19
2.14%
143.38
Palm Pilot
219
0.80
1.619
0.0255
5.59
nonident.
156.74
145.19
3.36%
142.37
Xbox
74
0.50
0.699
0.0134
0.99
ident.
85.73
74.19
0
74.19
Xbox
111
0.70
1.204
0.0137
1.52
nonident.
85.73
74.19
0.39%
84.18
Xbox
141
0.80
1.614
0.0137
1.93
nonident.
85.73
74.19
0.39%
84.17
Appendix
Table D.2: The nine eBay cells ( n=3 ). G(λ) is the censored fraction, rG∗ the refit’s optimal reserve, Δ the numerical minimax regret in percent of RG∗ and r† the minimax reserve; λ , RG∗ , rG∗ and r† are in dollars.
Significant decrease
Contrast 103Cλ (90% interval)
Item
bw ×
pointwise
sup- t
Peak (90%)
λ1
λ2
λ3
Palm Pilot
0.5
no
no
252 [247, 255]
24.2 [20.8, 28.0]
21.0 [15.6, 26.5]
13.6 [7.5, 19.6]
Palm Pilot
1
no
no
255 [252, 259]
20.4 [18.0, 23.0]
16.4 [13.6, 19.4]
11.5 [9.0, 14.2]
Palm Pilot
2
no
no
266 [264, 269]
12.9 [12.0, 13.8]
9.3 [8.6, 10.0]
6.6 [6.1, 7.1]
Xbox
0.5
yes
yes
120 [103, 205]
−5.1 [ − 8.6, − 1.4]
0.5 [ − 3.5, 4.9]
4.0 [0.5, 8.2]
Xbox
1
yes
yes
113 [106, 119]
−5.4 [ − 8.7, − 2.0]
−1.3 [ − 4.8, 2.6]
2.1 [ − 1.0, 5.6]
Appendix
Table D.3 : Kernel-hazard diagnostic on the reference records ( 2000 bidder-bootstrap draws). “bw × ” is the multiple of Silverman’s bandwidth; “Peak” is the location of the full-sample hazard’s maximum with its 90% bootstrap interval; the contrasts are at the item’s three boundaries of Table D.2 .
Kernel fit
Isotonic MHR refit
Item
λ
A
h0
Regime
Δ
A
h0
Regime
Δ
Palm Pilot
155
0.72
0.0080
nonident.
0.46%
0.69
0.0080
nonident.
0.45%
Palm Pilot
199
1.23
0.0169
nonident.
2.16%
1.20
0.0169
nonident.
2.14%
Palm Pilot
219
1.64
0.0255
nonident.
3.42%
1.62
0.0255
nonident.
3.36%
Xbox
74
0.75
0.0134
ident.
0
0.70
0.0134
ident.
0
Xbox
111
1.40
0.0209
nonident.
1.51%
1.20
0.0137
nonident.
0.39%
Appendix
Table D.4: Boundary scalars, regime and minimax regret (percent of the fit’s own RG∗ ) of the unconstrained kernel fit and of the isotonic refit at the nine boundaries. “Infeasible” means A>λh0 , so no MHR completion exists.
MHR
Regular
Unrestricted
Item
λ
Regime
Worst-case
Actual
Worst-case
Actual
Worst-case
Actual
Palm Pilot
155
nonident.
0.45%
0.04%
1.59%
0.03%
9.24%
1.22%
Palm Pilot
199
nonident.
2.14%
0.01%
5.08%
<0.01%
24.54%
1.46%
Palm Pilot
219
nonident.
3.36%
0.02%
7.46%
0.21%
33.35%
2.18%
Xbox
74
ident.
0
0
0
0
8.21%
1.49%
Xbox
111
nonident.
0.39%
0.38%
5.19%
0.18%
25.07%
0.66%
Appendix
Table D.5: Regret of the minimax reserves under three distribution classes on the nine eBay cells ( n=3 ), in percent of RG∗ , with exact censored fraction and observed tail; the layout of Table 7.1 . Worst-case regret ranges over the compatible distributions in the indicated class; actual regret uses the refit G . On the two identifiable cells the observed tail contains a shared optimum, so regular minimax regret is zero (Proposition 6.1 ).
Figure D.3 : Learning the robust reserve on the Palm Pilot refit ( n=3 ); the layout of Figure 7.1 . (a) Mean excess worst-case regret WG−Δ in dollars over 200 replications on the three Palm Pilot cells: the plug-in, the censored Weibull MLE and the linear-completion rule. (b) Mean worst-case regret in percent of RG∗ with N=2000 and 80 replications per λ , on a linear axis; the frontier is at λ=rG∗≈145 dollars. Bars show one Monte-Carlo standard error; some are smaller than the markers.
Worst-case regret
Actual regret under each distribution
Rule
WG
Refit G
Equalizer
Exponential
Tangent
Plug-in, Auto-selector
2.32
0.01
0.00
0.39
0.72
Plug-in (Weibull hazard)
2.46
0.01
0.01
0.34
0.92
Weibull MLE
2.96
0.27
0.09
0.84
0.04
Gamma MLE
2.85
0.22
0.07
0.77
0.08
Lognormal MLE
2.70
0.16
0.04
0.69
0.14
Appendix
Table D.6: Four distributions with the censored records of the Palm Pilot refit at λ=199 ; the layout of Table 7.2 . Mean worst-case and actual regret in percent of RG∗ ( N=3000 , 200 paired replications). Actual regret compares each distribution with its own optimal revenue. Numerical minimax regret is Δ=2.14% ; Monte-Carlo standard errors are at most 0.02 percentage points except in the Linear completion row ( 0.12 ). Rules that post the same reserves share a row.
Cell
Δ
Rule
N=3,000
N=32,000
WG
under G
WG
under G
Palm Pilot, λ=155
0.45
Plug-in
0.61
0.06
0.52
0.04
Weibull MLE
2.48
0.70
2.47
0.70
Gamma MLE
2.95
1.02
2.94
1.01
Lognormal MLE
3.08
1.11
3.07
1.10
Palm Pilot, λ=199
2.14
Plug-in
2.32
0.01
2.20
0.01
Appendix
Table D.7: Worst-case regret and actual regret under the refit G on the five nonidentifiable Palm Pilot and Xbox cells, in percent of RG∗ ( n=3 , mean over 200 paired replications). Monte-Carlo standard errors are at most 0.04 points.
Censored MLE
Item
λ
N
Full-info
Plug-in
Auto-selector
Weibull
Gamma
Lognormal
Linear compl.
Boundary
Drop
Cartier
555
274
8.16
14.46
5.60
10.90
11.85
13.52
14.60
14.46
7.53
Cartier
956
274
8.15
9.19
9.86
11.11
11.47
14.04
9.98
4.56
7.04
Cartier
1274
274
8.32
9.47
9.75
10.49
11.05
12.27
9.06
6.38
8.11
Xbox
74
322
1.09
1.82
1.88
2.37
1.93
3.12
1.33
1.33
1.83
Xbox
111
321
0.99
2.22
2.63
2.31
1.96
3.25
1.79
17.84
24.14
Appendix
Table D.8 : Held-out regret on the recorded Xbox and Cartier bids, in percent of each split’s test optimum; the layout and experimental settings of Table 7.3 . Standard errors at most 0.37 points.
Worst-case
Held-out
Item
λ
K
m
N
α
G(λ)
Regime
Detected
Δ
Auto-selector
Weibull
Auto-selector
Weibull
Boundary
Cartier
10
12
75
118
0.36
0.01
ident.
ident.
0
5.76
0.48
10.18
6.81
23.66
Cartier
100
8
39
78
0.50
0.11
ident.
ident.
0
1.06
3.89
6.94
14.18
22.62
Cartier
500
9
49
88
0.44
0.46
ident.
nonident.
0
2.80
0.03
12.01
5.35
12.01
Xbox
10
13
124
142
0.13
0.05
ident.
nonident.
0
5.18
0.05
8.66
1.24
8.66
Xbox
50
15
112
164
0.32
0.34
ident.
ident.
0
0.00
0.11
0.45
1.39
3.59
Appendix
Table D.9 : Sellers’ own opening bids on Xbox and Cartier; the layout of Table 7.4 , with N from the item’s own reference rate. “fallback” marks fewer than mmin records.
λ
Rate
N
α
Detected
Reserve
Auto-selector
Weibull
10
8
348†
0.00
ident.
149
0.15
2.58
10
10
348†
0.00
ident.
149
0.15
2.58
10
12.5
400
0.13
ident.
149
0.15
2.52
10
15
480
0.28
ident.
149
0.15
1.25
10
20
640
0.46
ident.
149
0.15
21.05
50
8
74†
0.01
nonident.
50
6.84
1.58
Appendix
Table D.10: Sensitivity of the Palm Pilot opening-bid study to the imputed bidder count: the reference rate (records per auction) replaced by 8 , 10 , 15 and 20 , with 12.5 the rate of Table 7.4 ( n=3 , 50 draws of the winner imputation). Detected is the auto-selector’s majority branch, Reserve its mean posted reserve in dollars, and Auto-selector and Weibull the held-out regret of the auto-selector and the censored Weibull MLE in percent of the test optimum. † : K times the rate is below the m observed records, so N=m+1 .
Worst-case regret
Actual regret
Item
λ
n
Δ
Plug-in
Weibull
Gamma
Lognormal
Linear compl.
Boundary
Plug-in
Weibull
Palm Pilot
155
2
0.82
1.15
4.49
5.43
5.70
1.35
1.35
0.09
1.44
Palm Pilot
155
3
0.45
0.62
2.50
2.98
3.11
0.75
0.75
0.06
0.72
Palm Pilot
155
5
0.13
0.17
0.73
0.85
0.88
0.22
0.22
0.02
0.17
Palm Pilot
155
10
0.00
0.01
0.03
0.03
0.03
0.01
0.01
0.00
0.00
Palm Pilot
199
2
4.50
4.95
5.28
5.05
4.79
16.40
42.01
0.20
0.51
Appendix
Table D.11: Minimax regret and mean worst-case and actual regret against the number of bidders, in percent of RG∗ ( N=3000 , 40 paired replications; standard errors at most 0.11 points for the plug-in and the MLEs and 0.43 for the linear-completion reserve).
Item
Winner’s record
Records
RG∗
rG∗
λ
G(λ)
Regime
Δ
r†
Palm Pilot
Kaplan–Meier
1652
156.7
145.2
155
0.50
nonident.
0.45
149
199
0.70
nonident.
2.14
143
219
0.80
nonident.
3.36
142
Palm Pilot
closing price
1652
155.8
145.1
155
0.50
nonident.
0.47
149
199
0.71
nonident.
2.27
143
219
0.82
nonident.
3.85
141
Appendix
Table D.12: The winner’s right-censored record under three treatments: sample size, optimal revenue and reserve of the refit, and at each of the item’s three boundaries the censored fraction, regime, minimax regret (percent of RG∗ ) and minimax reserve (dollars).
ζ
G(λ)
r†
Δ
Plug-in
Auto-selector
Linear compl.
Frontier
0.500
0.78
0.949
2.376%
2.736 (0.038)
2.736 (0.038)
3.737 (0.090)
0%
0.300
0.88
0.735
1.904%
2.454 (0.059)
2.454 (0.059)
2.322 (0.031)
2%
0.200
0.91
0.646
1.236%
1.979 (0.132)
1.979 (0.132)
1.620 (0.030)
23%
0.100
0.93
0.572
0.416%
1.189 (0.119)
1.189 (0.119)
0.616 (0.018)
40%
0.050
0.94
0.537
0.119%
0.880 (0.113)
0.880 (0.113)
0.223 (0.012)
47%
0.020
0.95
0.515
0.020%
0.758 (0.167)
1.558 (0.913)
0.073 (0.006)
47%
Appendix
Table D.13: Worst-case regret of the plug-in, the auto-selector and the linear-completion reserve in percent of RG∗ , mean (Monte-Carlo standard error), for two-level MHR distributions approaching full identification: λ=1.5 , h0=2 , A=h0λ(1−ζ) ( n=3 , N=3000 , 60 paired replications per ζ ). “Frontier” is the percentage of replications in which the raw estimate lies beyond the zero-regret frontier A=λh0 and the plug-in posts the linear-completion reserve at the projected pair, rc(AˉN) . The minimax regret Δ has body-grid error at most 8.3×10−4 percentage points.
n
λh0
A/(λh0)
w
Δ
W(rc)
W(rc)/Δ
bound
W(rc) /bound
2
2
0.50
0.500
0.02177
0.11574
5.3
0.12500
0.93
2
2
0.80
0.320
0.00430
0.00775
1.8
0.02000
0.39
2
2
0.90
0.180
0.00111
0.00150
1.3
0.00500
0.30
2
2
0.95
0.095
0.00028
0.00033
1.2
0.00125
0.26
2
2
0.99
0.020
0.00001
0.00001
1.1
0.00005
0.24
2
4
0.50
1.000
0.02612
0.03417
1.3
0.25000
0.14
Appendix
Table D.14: Worst-case regret W(rc) of the linear-completion reserve and minimax regret Δ at the population level, in revenue units, as A approaches λh0 ( λ=1 ). w=A(λh0−A)/(λh0) is the largest gap between the chord and the backward tangent; Δ uses 1600×1600 bodies (body-grid error at most 5.1×10−6 ); “bound” is k1d2/(2h0) of Proposition 5.8 with d=λh0−A and k1=((n−1)/n)n−1 .
Family
λh0
A/(λh0)
Side
N
Misclass.
Plug-in
Auto-selector
Tail
Linear compl.
Weibull( k=2 )
0.85
0.5
tail-optimal
1,000
22.0 (2.9)
0.139 (0.005)
0.118 (0.011)
0.090 (0.011)
0.120 (0.000)
Weibull( k=2 )
0.85
0.5
tail-optimal
10,000
3.0 (1.2)
0.121 (0.000)
0.027 (0.003)
0.023 (0.003)
0.120 (0.000)
Weibull( k=2 )
0.9
0.5
tail-optimal
1,000
34.5 (3.4)
0.094 (0.008)
0.122 (0.016)
0.091 (0.015)
0.055
Weibull( k=2 )
0.9
0.5
tail-optimal
10,000
11.5 (2.3)
0.056 (0.001)
0.020 (0.002)
0.016 (0.001)
0.055
Weibull( k=2 )
0.95
0.5
tail-optimal
1,000
36.5 (3.4)
0.072 (0.009)
0.127 (0.013)
0.095 (0.012)
0.014 (0.000)
Weibull( k=2 )
0.95
0.5
tail-optimal
10,000
32.5 (3.3)
0.026 (0.003)
0.024 (0.003)
0.013 (0.001)
0.014 (0.000)
Appendix
Table D.15 : Excess worst-case regret WG(r)−Δ of four rules in percent of RG∗ , mean (Monte-Carlo standard error), in cells on both sides of the zero-regret frontiers at λ=1 ( n=3 , 200 paired replications per cell and N ). “Side” is the case of the true boundary pair (A,h0) : tail-optimal or fully identified, both in the identifiable regime, or nonidentifiable; “Misclass.” is the percentage of replications in which the auto-selector’s branch disagrees with that regime (binomial standard error). Tail is the empirical revenue maximizer on [λ,10λ] . In identifiable cells Δ=0 and WG includes the tail gain ( B.13 ). Entries without a standard error are constant across replications.
Figure D.4 : (a) Minimax regret Δ/λ and (b) the position (r†−1/h0)/(u−1/h0) of the minimax reserve in the interval ( 4.6 ), u=min{λ,λ−(A−1)/h0} , over the boundary pair (λh0,A) , n=3 , computed with 200×200 bodies, 401 reserves and three local refinement rounds. No pair lies above the line A=λh0 ; white lines are contours.
Distribution
λ
Plug-in
Auto-selector
Drop
Lognormal (1.6,0.5)
3
0.34 (0.00)
0.03 (0.00)
0.03 (0.00)
Lognormal (1.6,0.5)
5
0.66 (0.02)
0.66 (0.02)
2.40 (0.01)
Lognormal (1.6,0.8)
3
1.69 (0.00)
0.09 (0.01)
0.11 (0.01)
Lognormal (1.6,0.8)
5
0.00 (0.00)
0.06 (0.01)
0.29 (0.01)
Appendix
Table D.16: Mean actual regret in percent of each lognormal distribution’s optimal revenue ( n=3 , N=3000 , 400 paired replications). Parentheses show Monte-Carlo standard errors.
Instance
λ
rMHR†
rreg†
rall†
Rayleigh(1)
1.5
1.088
1.132
0.886
Weibull(1.5,1)
1
0.827
0.797
0.636
Weibull(2.5,1.5)
2
1.128
1.030
0.803
Appendix
Table D.17: Population minimax reserves used in Table 7.1 ( n=3 ).
Instance
λ
Computed minimax regret
Refinement change
Rayleigh(1)
1.5
0.02039
0.000001
Weibull(1.5,1)
1
0.00699
0.000000
Weibull(2.5,1.5)
2
0.04468
0.000002
Appendix
Table D.18: Numerical MHR minimax regret in revenue units ( n=3 ). Refinement change compares 150 with 800 bodies per side and is not an error bound.
Quantity
Bound
Largest measured / bound
LipΘ
3λmax{1,λ}
0.0486
Lipr
max{1,λh0}
0.169
Appendix
Table D.19: Measured finite-difference slopes relative to the proved MHR bounds.
Online auction houses must set reserve prices for highly differentiated products whose unstructured descriptions contain valuable demand information. Existing approaches generally require managers to specify product attributes in advance. We ask whether a language model can instead learn from listing descriptions and historical bids to set item-specific reserve prices before bidding begins. We fine-tune a pretrained text embedding language model using the structural likelihood of auction bids as its training objective. This directs representation learning toward the valuation distribution that determines the reserve decision. We show that consecutive upper bids identify the tail information needed to set an optimal reserve without observing or modeling the number of potential bidders. A split-sample estimator learns the representation on one half of the data and conducts structural estimation and inference conditional on that representation on the other. We apply the method to approximately 65,800 online used-vehicle auctions. With seller values calibrated to match the sale rate among listings, the model implies a counterfactual net increase of $165.8 million, representing a 14.6% increase above the incumbent reserve policy. The framework translates a listing description and a seller's value into an item-specific reserve recommendation without constructing a product taxonomy or requiring a bid history for the item being priced. It provides an approach to adapting language models for operational decisions in marketplaces where products are heterogeneous and demand-relevant information is recorded in text.
We study the problem of learning to bid when the bidder's value is dynamic, i.e., when the current value depends on past outcomes. Specifically, we consider a bidder participating in repeated second-price auctions whose value depends on the time elapsed since their last successful bid, with auctions arriving in continuous time and only aggregated feedback revealed at the end of the horizon. Such a bidder must (1) balance the immediate benefit of winning the current auction against its impact on future values and (2) learn unknown environmental parameters. We derive regret bounds for a class of learning methods that combine plug-in estimators with a differential-equation characterization of the optimal policy, and show that a specific confidence bound algorithm learns the optimal policy with a near optimal regret of O(logN) for piecewise linear primitives, and O(N1/3) for general, smooth primitives, achieving these regrets without explicit randomization. These theoretical results are supported by numerical experiments.
Bidding in the European Frequency Containment Reserve (FCR) market is challenging for flexibility providers because competing offers are hidden and bidders observe only partial feedback form the market, such as, clearing price and awarded quantity. For a participant active in a single country, we show that the multi-country FCR clearing problem can be recast as a repeated multi-unit uniform-price auction against an endogenous vector of opposing bids. This reformulation yields an online learning problem and allows us to adapt a Best-of-Both-Worlds combinatorial semi-bandit algorithm implementable from this standard market feedback. The resulting bidder achieves logarithmic pseudo-regret in stochastic environments and O(T) regret in adversarial ones. Synthetic experiments confirm the expected scaling, and backtests on historical European FCR data show competitive performance in practice: the method performs especially well on stable products, while EXP3-type baselines can be safer under stronger non-stationarity. Overall, the results show that learning-based bidding in FCR markets is theoretically grounded and practically useful when the learning rule matches product-level market stability.
Marius Potfer, Cheng Wan, Pierre Gruet
ENSAE, CREST (Fairplay joint team), 5 Av. Le Chatelier, 91120, Palaiseau, France · EDF Lab Paris-Saclay, FiME (Laboratoire de Finance des Marchés de l’Énergie), 7 Bd Gaspard Monge, 91120, Palaiseau, France