Abstract
Sponsored search advertisement allows advertisers to target their messages to appropriate customer segments at low costs. While search engines are interested in auction mechanisms that boost their revenues, advertisers seek optimal bidding strategies to increase their net sale revenues for multiple keywords under strict daily budget constraints in an environment where keyword query arrivals, competitor bid amounts, and user purchases are random. We focus on the advertiser’s question and formulate her optimal intraday dynamic multi-keyword bidding problem as a continuous-time stochastic optimization problem. We solve the problem, characterize an optimal policy, and bring a numerical algorithm for implementation. We also illustrate our optimal bidding policy and its benefits over heuristic solutions on numerical examples.
Similar content being viewed by others
References
Abhishek V, Hosanagar K (2013) Optimal bidding in multi-item multislot sponsored search auctions. Oper Res 61(4):855–873
Abrams Z, Mendelevitch O, Tomlin J (2007) Optimal delivery of sponsored search advertisements subject to budget constraints. In: Proceedings of the 8th ACM Conference on Electronic Commerce, EC ’07, pp 272–278. ACM
Bae J, Kagel JH (2019) An experimental study of the generalized second price auction. Int J Ind Organ 63:44–68
Balseiro SR, Besbes O, Weintraub GY (2015) Repeated auctions with budgets in ad exchanges: Approximations and design. Manage Sci 61(4):864–884
Borgs C, Chayes J, Immorlica N, Jain K, Etesami O, Mahdian M (2007) Dynamics of bid optimization in online advertisement auctions. In: Proceedings of the 16th International Conference on World Wide Web, WWW ’07, pp 531–540. ACM
Boulatov A, Severinov S (2021) Optimal and efficient mechanisms with asymmetrically budget constrained buyers. Games Econom Behav 127:155–178
Brubach B, Sankararaman KA, Srinivasan A, Xu P (2016) New algorithms, better bounds, and a novel model for online stochastic matching. In ESA
Chaitanya N, Narahari Y (2012) Optimal equilibrium bidding strategies for budget constrained bidders in sponsored search auctions. Oper Res Int J 12(3):317–343
Cholette S, Özlük Ö, Parlar M (2012) Optimal keyword bids in search-based advertising with stochastic advertisement positions. J Optim Theory Appl 152(1):225–244
Dayanik S, Parlar M (2013) Dynamic bidding strategies in search-based advertising. Ann Oper Res 211(1):103–136
Devanur N, Jain K, Sivan B, Wilkens C (2019) Near optimal online algorithms and fast approximation algorithms for resource allocation problems. J ACM 66(1):1–41
Devanur NR, Hayes TP (2009) The adwords problem: online keyword matching with budgeted bidders under random permutations. In: Proceedings of the 10th ACM Conference on Electronic Commerce, EC ’09, pp 71–78. ACM
Feldman J, Muthukrishnan S, Pal M, Stein C (2007) Budget optimization in search-based advertising auctions. In: Proceedings of the 8th ACM Conference on Electronic Commerce, EC ’07, pp 40–49. ACM
Feldman J, Henzinger M, Korula N, Mirrokni VS, Stein C (2010) Online stochastic packing applied to display ad allocation. In: De Berg M, Meyer U (eds) Algorithms - ESA 2010. Springer, Berlin, pp 182–194
Goel A, Mahdian M, Nazerzadeh H, Saberi A (2010) Advertisement allocation for generalized second-pricing schemes. Oper Res Lett 38(6):571–576
Goel G, Mehta A (2008) Online budgeted matching in random input models with applications to Adwords. In Proceedings of the nineteenth annual ACM-SIAM Symposium on Discrete Algorithms, SODA ’08, pp 982–991. SIAM
Hummel P (2018) Hybrid mechanisms for Vickrey–Clarke–Groves and generalized second-price bids. Internat J Game Theory 47(1):331–350
Iyer K, Johari R, Sundararajan M (2014) Mean field equilibria of dynamic auctions with learning. Manage Sci 60(12):2949–2970
Jaillet P, Lu X (2014) Online stochastic matching: New algorithms with better bounds. Math Oper Res 39(3):624–646
Jansen B, Mullen T (2008) Sponsored search: An overview of the concept, history, and technology. Int J Electron Bus 6(2):114–131
Kitts B, Leblanc B (2004) Optimal bidding on keyword auctions. Electron Mark 14(3):186–201
Kotowski MH (2020) First-price auctions with budget constraints. Theor Econ 15(1):199–237
Küçükaydın H, Selçuk B, Özlük Ö (2020) Optimal keyword bidding in search-based advertising with budget constraint and stochastic ad position. J Oper Res Soc 71(4):566–578
Lee D, Zioło P, Han W, Powell WB (2017) Optimal online learning in bidding for sponsored search auctions. In 2017 IEEE Symposium Series on Computational Intelligence (SSCI), pp 1–8
Mahdian M, Nazerzadeh H, Saberi A (2012) Online optimization with uncertain information. ACM Trans Algorith 8(1):1–29
Maillé P, Markakis E, Naldi M, Stamoulis GD, Tuffin B (2012) Sponsored search auctions: An overview of research with emphasis on game theoretic aspects. Electron Commer Res 12(3):265–300
Mehta A, Saberi A, Vazirani U, Vazirani V (2007) AdWords and generalized online matching. J ACM 54(5):22–es
Naor J, Wajc D (2015) Near-optimum online ad allocation for targeted advertising. In Roughgarden T, Feldman M, Schwarz M (eds) Proceedings of the sixteenth ACM Conference on Economics and Computation, EC ’15, pp 131–148. ACM
Özlük Ö (2011) Revenue management: A practical pricing perspective, chapter search engine advertising: an overview from a revenue management angle, pp 153–165. Palgrave Macmillan London
Özlük Ö, Cholette S (2007) Allocating expenditures across keywords in search advertising. J Reven Pricing Manage 6(4):347–356
Perlich C, Dalessandro B, Hook R, Stitelman O, Raeder T, Provost F (2012) Bid optimizing and inventory scoring in targeted online advertising. In Proceedings of the 18th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’12, pp 804–812. ACM
Pin F, Key P (2011) Stochastic variability in sponsored search auctions: Observations and models. In: Proceedings of the 12th ACM Conference on Electronic Commerce, EC ’11, pp 61–70. ACM
Qin T, Chen W, Liu T (2015) Sponsored search auctions: Recent advances and future directions. ACM Trans Intell Syst Technol 5(4):1–34
Rusmevichientong P, Williamson DP (2006) An adaptive algorithm for selecting profitable keywords for search-based advertising services. In: Proceedings of the 7th ACM Conference on Electronic Commerce, EC ’06, pp 260–269. ACM
Selçuk B, Özlük Ö (2013) Optimal keyword bidding in search-based advertising with target exposure levels. Eur J Oper Res 226(1):163–172
Skiera B, Abou NN (2013) Prosad: A bidding decision support system for profit optimizing search engine advertising. Market Sci 32(2):213–220
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary Information
Below is the link to the electronic supplementary material.
Appendix: Supplementary proofs
Appendix: Supplementary proofs
Proof of the equality in (7)
For a given admissible bidding policy, since \( \sum _{n\ge 1} b_n R_n Z_n 1_{\{\tau _n \le T\}} \) is bounded by B, we have
where the second line follows by the monotone convergence theorem applied to each expected random sum separately. Note that
due to the conditional independence of \(W_n\), the conditional expectations in (3), and the conditional distributions in (1). Similarly, we write
thanks to (1) and (2). Using now (37-38) in (36) yields
establishing (7). In (39), the second equality is by the monotone convergence theorem (applied to each expectation), and the last equality follows simply by the boundedness of \(\sum _{n\ge 1} 1_{\{\tau _n \le T\}} b_n \rho _{\kappa _n} G_{\kappa _n} (b_n)\) (by B). \(\square \)
Proof of Proposition 1
Monotonicity of \(\mathcal {D}[f]\) in f is obvious, and the non-negativity of \(D[f_1]\) follows after taking \(b_1= 0\) in (10). To prove the upper bound \(\mathcal {D}[f_2] (\cdot , T) \le \gamma \lambda T\), we observe that
in which the second line follows by conditioning on \(\mathcal {F}_{\tau _1}\), the inequality in the third line is by related arguments on the mapping \(b \mapsto (\mu _k - b \rho _k)G_k(b)\) and the definition of \(\gamma \) in Remark 3, and the very last equality is simply by integration (by parts). \(\square \)
Proof of Proposition 2
For notational convenience, let us define for \(1 \le k \le K\)
with \(M_k\) and \(M^*_k \) defined in (12-13). It is easy to verify that the mappings \(B \mapsto L^*_k[f] (B,\cdot )\) and \(T \mapsto L^*_k[f] (\cdot ,T)\) are non-decreasing under the given assumptions (on the monotonicity of f in its arguments). Since
it follows that \(B \mapsto \mathcal {D}[f] (B,\cdot )\) and \(T \mapsto \mathcal {D}[f] (\cdot ,T)\) are again non-decreasing. This proves the non-negativity of the differences in (16) and (17).
The prove the upper bound in (16), we note that, for \(T_1 < T_2\), we have
It is easy to verify that \( L^*_k[f] (B,T) \le \mu _k + \Vert f\Vert \le {\bar{\mu }} + \Vert f\Vert \). Hence, we obtain \( \mathcal {D}[f] (B,T_2) - \mathcal {D}[f] (B,T_1) \le \lambda ({\bar{\mu }} + \Vert f\Vert ) (T_2 - T_1) \).
To establish the second claim, let \(B_1 < B_2\) be two budget levels, and for fixed T, let \(b_2 = b^*_k [f] (B_2,T)\) denote the maximum bid in (14) for a fixed k with the budget level \(B_2\). Note that \(b_1 = \frac{B_1}{B_2} b_2 < b_2\) is a feasible bid for the budget \(B_1\), and we have \(b_2 - b_1 = \frac{(B_2-B_1)}{B_2} b_2 \le B_2 - B_1\). Then, we write
Next, we consider the difference
and using (41), we obtain \(\mathcal {D}[f](B_2,T) - \mathcal {D}[f](B_1,T) \le \)
giving us the upper bound in (17). \(\square \)
Proof of Proposition 3
Because \({\underline{U}}_0 (\cdot ,\cdot ) = 0 \le {\overline{U}}_0(B,T) = \gamma \lambda T\), Proposition 1 implies that \(0 \le {\underline{U}}_1 \le {\overline{U}}_1 \le \gamma \lambda T \). Let us now assume that \({\underline{U}}_{n-1} \le {\underline{U}}_{n} \le {\overline{U}}_{n} \le {\overline{U}}_{n-1} \) for some \(n\ge 1\). Then, again by Proposition 1, we have \(\mathcal {D}[{\underline{U}}_{n-1}] \le \mathcal {D}[{\underline{U}}_{n}] \le \mathcal {D}[ {\overline{U}}_{n}] \le \mathcal {D}[{\overline{U}}_{n-1}] \), and together with the induction hypothesis, this yields \( {\underline{U}}_{n-1} \le {\underline{U}}_{n} \le {\underline{U}}_{n+1} \le {\overline{U}}_{n+1} \le {\overline{U}}_{n} \le {\overline{U}}_{n-1}\). Hence, by induction we conclude that \(( {\underline{U}}_n (\cdot ,\cdot ) )_{n \ge 0}\) and \(( {\overline{U}}_n (\cdot ,\cdot ) )_{n \ge 0}\) are non-decreasing and non-increasing, respectively, and the collection \(( {\underline{U}}_n (\cdot ,\cdot ) )_{n \ge 0}\) is bounded from above by \(( {\overline{U}}_n (\cdot ,\cdot ) )_{n \ge 0}\).
Let \({\underline{U}}_{\infty } \) denote the pointwise limit of the monotone sequence \(({\underline{U}}_{n})_{n \ge 0}\). Because the operator \(\mathcal {D}\) is a contraction mapping, \( \Vert {\underline{U}}_{\ell +1} - {\underline{U}}_{\ell } \Vert = \Vert \mathcal {D}[{\underline{U}}_{\ell }] - \mathcal {D}[{\underline{U}}_{\ell -1}] \Vert \le (1-e^{-\lambda T_{\text {max}}}) \Vert {\underline{U}}_{\ell } - {\underline{U}}_{\ell -1} \Vert \le \ldots \le (1-e^{-\lambda T_{\text {max}}})^\ell \; \Vert {\underline{U}}_1 \Vert , \) using which we obtain
and this gives \( 0 \le {\underline{U}}_{\infty } (B,T) - {\underline{U}}_{n} (B,T) = \lim _{m \rightarrow \infty } {\underline{U}}_{m+n} (B,T) - {\underline{U}}_{n} (B,T) \le \Vert {\underline{U}}_1 \Vert \lim _{m \rightarrow \infty } \sum _{\ell = n}^{m+n-1} (1-e^{-\lambda T_{\text {max}}})^\ell \) \( \le \gamma \lambda T_{\text {max}} \sum _{\ell = n}^{\infty } (1-e^{-\lambda T_{\text {max}}})^\ell = \gamma \lambda T_{\text {max}} (1-e^{-\lambda T_{\text {max}}})^n e^{\lambda T_{\text {max}}}\). Since this is true for all \((B,T) \in \Delta \), we have \(\Vert {\underline{U}}_{\infty } - {\underline{U}}_{n} \Vert \le \gamma \lambda T_{\text {max}} (1-e^{-\lambda T_{\text {max}}})^n e^{\lambda T_{\text {max}}}\), which proves that the convergence of \({\underline{U}}_{n}\) to \({\underline{U}}_{\infty }\) is uniform on \(\Delta \) as \(n \rightarrow \infty \). In the arguments above, if we replace \({\underline{U}}\) with \({\overline{U}}\), then we obtain the same upper bound for \( {\overline{U}}_{n} (B,T) - {\overline{U}}_{\infty } (B,T) \ge 0\). This shows that the \({\overline{U}}_{n} \)’s converge to \( {\overline{U}}_{\infty } (B,T)\) also uniformly with the same error bound. Because \({\overline{U}}_{n}\)’s and \({\underline{U}}_{n}\)’s are continuous, their uniform limits \({\overline{U}}_{\infty }\) and \({\underline{U}}_{\infty }\), respectively, are also continuous on \(\Delta \).
The uniform convergence of \(({\underline{U}}_n)_{n \ge 0}\) and \(({\overline{U}}_n)_{n \ge 0}\) also imply that as \(n \rightarrow \infty \), we have \( M^*_{k} [{\overline{U}}_n] (\cdot ,\cdot )\) \( \rightarrow M^*_{k} [{\overline{U}}_\infty ] (\cdot ,\cdot )\) and \( M^*_{k} [{\underline{U}}_n] (\cdot ,\cdot ) \rightarrow M^*_{k} [{\underline{U}}_\infty ] (\cdot ,\cdot )\) for each \(1 \le k \le K\), and they are bounded from above by \(\max _k M^*_{k} [{\overline{U}}_0] (B_{\text {max}},T_{\text {max}}) < \infty \). Then, by bounded convergence theorem we obtain
which shows that \({\underline{U}}_{\infty }\) is a fixed point of \(\mathcal {D}\). Replicating the arguments above with \({\overline{U}}_{\infty }= \lim _n {\overline{U}}_{n}\), we observe that \({\overline{U}}_{\infty }\) is also a fixed point of the operator \(\mathcal {D}\). Finally, the uniqueness of the fixed point (see Lemma 1) implies that \({\underline{U}}_{\infty } = {\overline{U}}_{\infty } \). \(\square \)
Proof of Lemma 2
The inequality in (26) becomes an equality for \(j=1\). Assume that it holds for some \(1 \le j \le n \), and let us prove it for \(j+1\). Note that the right hand side of (26) can be decomposed as
Conditioning on \(\mathcal {F}_{\tau _{n-j+1}}\) and using the conditional distributions of \(W_{n-j+1}, R_{n-j+1}\), and \(Z_{n-j+1}\), the second expectation in (42) above becomes
where the last line is due to strong Markov property. Because \(\mathcal {D}[{\underline{U}}_{j-1}] = {\underline{U}}_j \), combining (42-43) with the induction hypothesis yields
and this proves the inequality (26) for \(j+1\). Hence, by induction it holds for all \(1\le j \le n+1\). \(\square \)
Proof of Proposition 5
For \(n =0\), we have \( V_0 = {\underline{U}}_0 = 0\) by construction, and the summation in (28) equals zero. Hence, the claim of the proposition is obvious. Therefore, we state the proposition and prove the equalities in (28) for \(n \ge 1\).
By Lemma 2, we have for every admissible policy \((b_i)_{i \ge 1}\)
for \(1 \le j \le n+1\). Evaluating this inequality with \(k = n+1\) gives \( \mathbb {E}\Big [ \sum _{i=1}^n (W_i-b_i R_i) Z_i 1_{\{\tau _i \le T\}} \big ] \le {\underline{U}}_{n} (B,T)\) and this implies \(V_n (B,T) \le {\underline{U}}_n (B,T)\) for all \(n \ge 1\).
We next establish the second equality in (28) for all \(n \ge 1\). Because \({\underline{U}}_0 \equiv 0\), we have
and this gives the second equality in (28) for \(n = 1\). Assume now that the second equality in (28) holds for some \(n\ge 1\). Then we have
Using the induction hypothesis and the strong Markov property, we obtain
Substituting (46) into (45) gives
This proves the second equality in (28) for \(n+1\). Hence, it holds for all \(n \ge 1\) by induction.
Clearly, \(V_n\) is an upper bound for the expected net revenue of any n-bid policy. Hence, combining all the arguments above, we now have, for any \(n \ge 1\),
and this establishes the equalities in (28). \(\square \)
Proof of Proposition 6
To prove the claim, it is sufficient to establish the identity
inductively for all \(n \ge 1\). When we let \(n \rightarrow \infty \), the expectation of the second term converges to zero because V is bounded (see Remark 3) and \(\mathbb {P} (\tau _n \le T) \rightarrow 0 \). Also both \(\mathbb {E}\big [ \sum _{i=1}^n W_i \cdot R_i Z_i 1_{\{\tau _i \le T\}} \big ]\) and \(\mathbb {E}\big [ \sum _{i=1}^n b^{(\infty )}_i R_i Z_i 1_{\{\tau _i \le T\}} \big ]\) convergence by monotone convergence theorem to the expectations of the corresponding infinite sums, and \(\sum _{i=1}^\infty b^{(\infty )}_i R_i Z_i 1_{\{\tau _i \le T\}} \le B\). Hence, as \(n \rightarrow \infty \), the expectation of the summation in (47) converges to the right hand side in (31).
For \(n= 1\), the expectation in (47) becomes \(\mathcal {D}[V] (B,T)\), and the equality holds since V is a fixed point of the operator \(\mathcal {D}\). Assume now that the equality holds for some \(n \ge 1\). Then
On the event \(\{\tau _1 \le T\}\), the conditional expectation above is equal to \(V\big (B^{(\infty )}_{\tau _1}, T-\tau _1 \big )\) by the strong Markov property and the induction hypothesis. Hence, the right hand side of the equality in (48) becomes
which proves (47) for \(n+1\). Hence, (47) holds for all \(n \ge 1\) by induction, and this completes the proof. \(\square \)
Proof of Proposition 7
The identity \(f(B,0) = V(B,0)\) for \(T = 0\) is obvious. Therefore, we only give the proof for \(T> 0\). Note that for a given admissible policy \((b_i)_{i \ge 1}\) and the corresponding budget process \(\{ B_t \}_{t \in [0,T]}\), the chain rule gives
Because \(M^*_{k} [f] (\cdot ,\cdot )\) is continuous on \(\Delta \) and the budget process is an \(\mathbb {F}\)-adapted càdlàg process, it follows that \(\{ M^*_{k} [f] (B_{t-} ,T-t) \}_{t \in [0,T]}\) is bounded and \(\mathbb {F}\)-predictable. Therefore, in terms of the counting process \(N = \{ N_t\}_{t \ge 0}\) with \(N_t = \sum _{i \ge 1} 1_{ \{ \tau _i \le T \} }\), for \(t \ge 0\), we have
Using the conditional distribution of \((Z_i,R_i)\) (with the help of the dominated convergence theorem to interchange the summation and the expectation), we obtain
Inserting (50-51) into (49) yields
and this implies that \(f(B,T) \ge V(B,T)\) because \((b_i)_{i \ge 1}\) was an arbitrary admissible policy. In particular, with the optimal policy \(b^{(\infty )}\) (and its budget process \(B^{(\infty )}\)) given in Lemma 6, the inequalities in (50) and (52) become equalities, and this proves that \(f(B,T) = V(B,T)\). \(\square \)
Rights and permissions
Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.
About this article
Cite this article
Dayanik, S., Sezer, S.O. Optimal dynamic multi-keyword bidding policy of an advertiser in search-based advertising. Math Meth Oper Res 97, 25–56 (2023). https://doi.org/10.1007/s00186-022-00803-y
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00186-022-00803-y