1) The document proposes a new advertising model called "multi-keyword multi-click advertisement option contracts for sponsored search" that allows advertisers to purchase options to target keywords and receive a certain number of clicks over a period of time.
2) It describes the pricing framework for these advertisement options, which is based on a multivariate geometric Brownian motion model of keyword click-through rates over time. The value of an option is calculated using Monte Carlo simulation or closed-form solutions depending on the number of keywords.
3) An empirical example demonstrates pricing a sample multi-keyword option using Monte Carlo simulation and analyzes how setting the exercise price to the expected value of click-through rates can increase search engine revenue compared
Multi-keyword multi-click advertisement option contracts for sponsored search
1. Multi-keyword multi-click advertisement
option contracts for sponsored search1
Bowei Chen†,2, Jun Wang†, Ingemar J. Cox†,‡, Mohan S. Kanhanhalli]
†University College London
‡University of Copenhagen
]National University of Singapore
2015
1
In ACM Transactions on Intelligent Systems and Technology (TIST) Volume 7 Issue 1, October 2015, No. 5.
2
Corresponding author: bowei.chen@cs.ucl.ac.uk
3. Limitations of auction mechanisms in sponsored search
• Uncertainty in payment prices for advertisers;
• Volatility in the search engine’s revenue;
• Weak loyalty between advertiser and search engine.
4. Tailored exotic options to sponsored search
A multi-keyword multi-click ad option contract allows its buyer to:
• Target a set of candidate keywords for a certain number of total clicks;
• Multiply exercise the option to purchase guaranteed clicks at any time on or prior to
the contract expiration date;
• Switch among candidate keywords when exercise the option without paying any cost.
It has the properties of multi-asset option and multi-exercise option.
8. Benefits
Advertiser Search engine
Secure ad service delivery Obtain upfront income in advance
Reduce uncertainty from auctions Have a stable & increased revenue
Cap advertising costs Increase advertisers’ loyalty online
9. Option pricing building blocks
• Underlying stochastic model
• Option payoff formulation
• Option pricing framework and solution
10. Multivariate geometric Brownian motion
The keyword Ki’s spot market CPC can be described as follows
dCi(t) = µiCi(t)dt + σiCi(t)dWi(t), i = 1, . . . , n,
where µi and σi are the drift and volatility, respectively, and Wi(t) is a standard Brownian
motion satisfying the conditions:
E(dWi(t)) = 0,
var(dWi(t)) = E(dWi(t)dWi(t)) = dt,
cov(dWi(t), dWj(t)) = E(dWi(t)dWj(t)) = ρijdt,
where ρij is the correlation coefficient between the ith and jth keywords, such that ρii = 1
and ρij = ρji. The correlation matrix is denoted by Σ, so that the covariance matrix is
MΣM, where M is the matrix with σi along the diagonal and zeros everywhere else.
11. Option value and payoff
The value of an m-click ad option at time t is equal to m number of 1-click ad option:
V(t, C(t); T, F, m) = mV(t, C(t); T, F, 1).
If exercised, we have
V(t, C(t); T, F, 1) = Φ(C(t)) := max{C1(t) − F1, . . . , Cn(t) − Fn, 0},
where F is a vector of exercise prices for candidate keywords, T is the expiration date, and
Φ(C(t)) is the option payoff at time t.
12. No early exercise property
Since e−rtΦ X(t)
is a sub-martingale under the risk-neutral probability measure Q (see
Appendix A), the proposed ad option can be priced as same as its European structure,
focusing on the payoff on the contract expiration date.
13. Option pricing framework
The option price π0 (i.e., the option value at time 0) can be obtained as follows (see
Appendix B for the derivation):
π0 = E[Φ(C(T)) | F0]
= me−rT
2πT
− n
2
|Σ|− 1
2
n
∏
i=1
σi
!−1
×
Z ∞
0
· · ·
Z ∞
0
Φ(e
C)
∏n
i=1
e
Ci
exp
−
1
2
ζT
Σ−1
ζ
de
C,
where F0 is the information history up to time 0, ζ = (ζ1, . . . , ζn)0, and
ζi = 1
σi
√
T
ln{e
Ci/Ci(0)} − (r −
σ2
i
2 )T
, i = 1, · · · , n.
14. Solutions
# of keywords Pricing method Reference
1 Black-Scholes formula for European call option See Appendix C
2 Peter Zhang dual strike European call option See Appendix C
≥ 3 Monte Carlo simulation See Algorithm 1
Algorithm 1:
function OptionPricingMC(K, C(0), Σ, M, m, r, T)
for k ← 1 to # of simulations do
[z1,k, . . . , zn,k] ← GenerateMultivariateNoise(MVN[0, MΣM])
for i ← 1 to n do
Ci,k ← Ci(0) exp
n
(r − 1
2 σ2
i )T + σizi,k
√
T
o
.
end for
Gk ← Φ([C1,k, . . . , Cn,k]).
end for
return π0 ← me−rTE0[Φ(C(T))] ≈ me−rT
1
e
n ∑e
n
k=1 Gk
.
end function
15. Empirical example of ad option pricing using Monte Carlo simulation
K =
‘canon cameras’
‘nikon camera’
‘yahoo web hosting’
, σ =
0.2263
0.4521
0.2136
, Σ =
1.0000 0.2341 0.0242
0.2341 1.0000 −0.0540
0.0242 −0.0540 1.0000
.
16. Revenue analysis for 1-keyword 1-click ad options
Let D(F) be the difference between the expected revenue from an ad option and the
expected revenue from only keyword auctions, we then have
D(F) =
C(0)N [ζ1] − e−rT
FN [ζ2] + e−rT
F
P(EQ
0 [C(T)] ≥ F)
| {z }
= Discounted value of expected revenue from option if EQ
0 [C(T)]≥F
+
C(0)N [ζ1] − e−rT
FN [ζ2] + e−rT
EQ
0 [C(T)]
P(EQ
0 [C(T)] F)
| {z }
= Discounted value of expected revenue from option if EQ
0 [C(T)]F
− e−rT
EQ
0 [C(T)]
| {z }
= Discounted value of expected revenue from auction
.
= C(0)N [ζ1] − e−rT
FN [ζ2] − e−rT
(EQ
0 [C(T)] − F) × P(EQ
0 [C(T)] ≥ F),
where N [·] is the cumulative probability of a standard normal distribution.
17. Revenue analysis for 1-keyword 1-click ad options con’t
• If F = 0, υ0 achieves its maximum value; therefore, D(F) → 0.
• If π0 = 0, F is as large as possible, P(EQ
0 [C(T)] ≥ F) → 0 and D(F) → 0.
• Since ln{C(T)/C(0)} ∼ N (r − σ2/2)T, σ2T
,
P(EQ
0 [C(T)] ≥ F) ≈ N
1
σ
√
T
ln{C(0)/F} + (r −
1
2
σ2
)T
= N [ζ2].
Therefore, D(F) = C(0)N [ζ1](1 − e− 1
2 σ2
T) 0, then
∂D(F)
∂F
23. F=EQ
0 [C(T)]
= −
1
√
2π
e− 1
2 ζ2
2 1
Fσ
√
T
0,
suggests that by setting F = EQ
0 [C(T)], the search engine can increase its profit.
24. Data3
Market Group Training set (31 days) Devetest set (31 days)
US
1 25/01/2012-24/02/2012 24/02/2012-25/03/2012
2 30/03/2012-29/04/2012 29/04/2012-31/05/2012
3 10/06/2012-12/07/2012 12/07/2012-17/08/2012
4 10/11/2012-11/12/2012 11/12/2012-10/01/2013
UK
1 25/01/2012-24/02/2012 24/02/2012-25/03/2012
2 30/03/2012-29/04/2012 29/04/2012-31/05/2012
3 12/06/2012-13/07/2012 13/07/2012-19/08/2012
4 18/10/2012-22/11/2012 22/11/2012-24/12/2012
3
The data was collected from Google AdWords by using its Traffic Estimation Service.
25. Validation conditions of the GBM model
¶ Normality of change rates of log CPCs
Histogram/Q-Q plot and the Shapiro-Wilk test
· Independence from previous data
Autocorrelation function (ACF) and the Ljung-Box statistic
26. Empirical example of the GBM assumption test
Keyword ‘canon 5d’
The Shapiro-Wilk test is with p-value 0.3712
The Ljung-Box test is with p-value 0.4555.
27. GBM assumption tests
1 2 3 4
0
50
100
150
(a) US market
Experimental group ID
Number
of
keywords
Non−GBM
GBM
1 2 3 4
0
50
100
150
(b) UK market
Experimental group ID
Number
of
keywords
Non−GBM
GBM
There are 14.25% and 17.20% of keywords in US and UK markets that satisfy the GBM
assumption, respectively.
28. Non-GBM dynamics and data fitting
Model Stochastic differential equation (SDE)
Constant elasticity of variance (CEV) dCi(t) = µiCi(t)dt + σi(Ci(t))1/2dWi(t)
Mean-reverting drift (MRD) dCi(t) = ki(µi − Ci(t))dt + σi(Ci(t))1/2dWi(t)
Cox-Ingersoll-Ross (CIR) dCi(t) = ki(µi − Ci(t))dt + (σi)1/2Ci(t)dWi(t)
Hull-White/Vasicek (HWV) dCi(t) = ki(µi − Ci(t))dt + σidWi(t)
WilcoxonA−B K−S
0
20
40
60
80
100
120
(a) US market
Testing model
Similarity
(%)
WilcoxonA−B K−S
0
20
40
60
80
100
120
(b) UK market
Testing model
Similarity
(%)
CEV
MRD
CIR
HWV
Wilcoxon test, Ansari-Bradley (A-B) test and Two-sample Kolmogorov-Smirnov (K-S) test
29. Check arbitrage via delta hedging
• Calculate delta
• 1-keyword 1-click ad options
∂V
∂C
= N
1
σ
√
T
ln
C(0)
F
+ (r +
σ2
2
)T
.
• n-keyword 1-click options (calculated by Monte Carlo method)
∂V/∂Ci = EQ
[∂V(T, C(T))/∂Ci(T)]
• The arbitrage detection criteria is
|e
γ −e
r| ≤ ε ? arbitrage does not exist : arbitrage exists,
where e
γ is the rate of returns from constructed hedging strategy, e
r is the equivalent
(discrete) risk-less bank interest rate in the period, and ε is the model variation
threshold (and we set ε = 5% in experiments). The identified arbitrage α is defined
as the excess return, that is
α =
e
γ − (e
r − ε), if e
γ e
r − ε,
e
γ − (e
r + ε), if e
γ e
r + ε.
30. Check arbitrage via delta hedging con’t
−0.1 0 0.1
−0.05
0
0.05
0.1
Invest return
Arbitrage
(a) n=2 group 2
−0.05 0 0.05 0.1
−0.05
0
0.05
Invest return
Arbitrage
(b) n=2 group 3
−0.1 0 0.1 0.2 0.3
−0.05
0
0.05
0.1
0.15
0.2
Invest return
Arbitrage
(c) n=2 group 4
0 0.1 0.2
0
0.05
0.1
Invest return
Arbitrage
(d) n=3 group 4
−0.05 0 0.05
−0.05
0
0.05
Invest return
Arbitrage
(e) n=3 group 2
Priced options
Risk−less return
Arb threshold
Empirical example of arbitrage analysis based on GBM for the US market.
31. Check arbitrage via delta hedging con’t
Testing arbitrage of options under the GBM: n is the number of keywords, N is the number of
options priced in a group, P(α) is % of options in a group identified arbitrage, and the E[α] is the
average arbitrage value of the options identified arbitrage.
n Group
US market UK market
N P(α) E[α] N P(α) E[α]
1
1 94 0.00% 0.00% 76 0.00% 0.00%
2 64 0.00% 0.00% 45 0.00% 0.00%
3 94 1.06% 0.75% 87 0.00% 0.00%
4 112 0.89% -0.37% 53 0.00% 0.00%
2
1 47 4.26% 1.63% 38 0.00% 0.00%
2 32 9.38% 0.42% 22 4.55% 13.41%
3 47 4.26% 0.84% 43 4.65% 0.82%
4 56 5.36% 3.44% 26 23.08% -6.22%
3
1 31 0.00% 0.00% 25 4.00% 0.00%
2 21 4.76% -1.38% 15 0.00% 0.00%
3 31 0.00% 0.00% 29 3.45% -1.12%
4 37 10.81% 3.87% 17 35.29% -2.54%
32. Robust tests of pricing models
n=1 n=2 n=3
0
20
40
60
80
100
Number of keywords in ad option
Fairness
(%)
GBM
CEV
MRD
CIR
HWV
34. Empirical example of revenue analysis for a 2-keyword 1-click ad option
0
10
20
0
0.2
0.4
0.6
0
5
10
15
Fixed CPC F1
(a)
Fixed CPC F2
Option
price
0
10
20
0
0.2
0.4
0.6
−0.2
0
0.2
0.4
0.6
0.8
Fixed CPC F1
(b)
Fixed CPC F2
Revenue
difference
Keywords ‘non profit debt consolidation’ and ‘canon 5d’, where ρ = 0.2247
35. Limitations and future work
• Other sophisticated stochastic processes are worth studying, such as jump-diffusion
models and stochastic volatility models.
• Game-theoretical pricing models for ad options.
• Optimal pricing and allocation of ad options.