Financial methods in online advertising
-
Upload
jun-wang -
Category
Technology
-
view
1.131 -
download
3
Transcript of Financial methods in online advertising
Financial Methods in Online Adver3sing
Dr. Jun Wang UCL
ACM SIGIR IATP13
Acknowledgements
Bowei Chen, Mathema3cal Finance Shuai Yuan, Computer Science
When online adver3sing goes wrong
h9p://mashable.com/2008/06/19/contextual-‐adverNsing/
Web users were unlikely to click a shoes ad that appeared along side an arNcle about the rather gruesome story about severed feet washing up on shore
What is the (fundamental) problem?
• AutomaNcally place the right ad for the right web page at the right -me for the right user
• The match is scienNfically difficult – because ulNmately ads shouldn’t be matched to the web pages where they are to be allocated,
– but the underlying web users, whose informaNon needs and reacNons are, however, not known a priori
Why financial methods?
• Our observaNons 3+ years ago: – although mulNple parNcipants from ad providers, adver-sers, and web users, the industry and research seem unbalanced • It is unclear who should do that. “search engines” as publishers, ad networks, doing the keyword matching etc
• li9le research has been found from the bidders (the adverNsers)’s perspecNve to help them manage their campaigns
An analogy with Financial markets
• Ads (display opportuniNes) are “traded” based on the dual force of supply (publishers) and demand (adverNsers)
Display OpportuniNes =? “raw materials” like
petroleum and natural gas
Ads prices are vola3le
(a) (b)
(c) (d)
(e) (f)
(g) (h)
Figure 4: Plots of time series analysis of CA-995.
discussing the pricing models for an on-line advertisementderivatives in the paper.
Advertisement derivatives can reduce risk—by enablingtwo parties to fix a price for a future transaction now. Theadvertiser will pay the money to the search engines for ad-vertising. So we assume that the advertiser should have theright while search engines should have the responsibility toshow the advertisement if advertiser proceeds his right. Thissituation could be modelled by options. We, therefore, pro-pose that search engines sell an call option to an advertiser,
for example, an European call option which gives its ownerthe right (without obligation) to buy the position of an ad-vertisement in people’s searching result in the future at astrike price, say K, which is determined at the current time,when the option is bought.
5. CONCLUSION
6. REFERENCES
The price movement of a display opportunity from Yahoo! ads data Under GSP (generalized second price aucNon)
(a) (b)
(c) (d)
(e) (f)
(g) (h)
Figure 4: Plots of time series analysis of CA-995.
discussing the pricing models for an on-line advertisementderivatives in the paper.
Advertisement derivatives can reduce risk—by enablingtwo parties to fix a price for a future transaction now. Theadvertiser will pay the money to the search engines for ad-vertising. So we assume that the advertiser should have theright while search engines should have the responsibility toshow the advertisement if advertiser proceeds his right. Thissituation could be modelled by options. We, therefore, pro-pose that search engines sell an call option to an advertiser,
for example, an European call option which gives its ownerthe right (without obligation) to buy the position of an ad-vertisement in people’s searching result in the future at astrike price, say K, which is determined at the current time,when the option is bought.
5. CONCLUSION
6. REFERENCES
Research opportuni3es • Need Ad’s Futures Contract and Risk-‐reduc4on Capabili4es
– AutomaNon is constrained mainly to “spots” markets, i.e., any transacNon where delivery takes place right away
– No principled technologies to support efficient forward pricing &risk management mechanisms
• If we got Futures Market, adverNsers could lock in the campaign cost and Publishers could lock in a profit
• Need to engineer a Unified Ad Exchange – The ad market is in the hands of a few key players. Each individual player defines its own ad system
– Arbitrage opportuniNes exist. Two display opportuniNes with similar targeted audiences and visit frequency may sell for quite different prices on two different markets.
Current trends in online adver3sing
• Spot markets: Real-‐Time Bidding (RTB) allows selling and buying online display adverNsing in real-‐Nme one ad impression at a Nme • Future markets: ProgrammaNc Premium to use automated procedures to reach agreement of sales between buyers (adverNsers) and sellers (publishers) • With the two trends, a step closer to the financial markets where unificaNon and interconnecNon are strongly promoted
Summary • Spot market: Real-‐Nme Bidding Shuai Yuan, Jun Wang, Real-‐Nme Bidding for Online AdverNsing: Measurement and Analysis, AdKDD’13 h9p://arxiv-‐web3.library.cornell.edu/abs/1306.6542 • Ad opNons Bowei Chen, Jun Wang, Ingemar Cox, and Mohan Kankanhalli, MulN-‐Keyword MulN-‐Click OpNon Contracts for Sponsored Search AdverNsing, under submission, 2013 h9p://arxiv.org/abs/1307.4980
Summary
• Spot market: Real-‐Nme Bidding Shuai Yuan, Jun Wang, Real-‐Nme Bidding for Online AdverNsing: Measurement and Analysis, AdKDD’13 h9p://arxiv-‐web3.library.cornell.edu/abs/1306.6542 • Ad opNons Bowei Chen, Jun Wang, Ingemar Cox, and Mohan Kankanhalli, MulN-‐Keyword MulN-‐Click OpNon Contracts for Sponsored Search AdverNsing, under submission, 2013
advanced approach is to learn a model including various fea-tures of webpages, which could then be used to compute arelevance score of advertisers’ targeting criteria [17, 7, 8].The understanding of user is usually referred to as behaviourtargeting, where ad networks utilise the browsing history ofa specific user to infer his interests, as well as geographicallocation, local time, etc for target matching [32, 23, 30].
In ad networks advertisers largely adopt the cost-per-click(CPC) or cost-per-acquisition (CPA) pricing models wherethey only pay when certain goal is achieved. These choicesreduce their risks thus are good for goal-driven campaigns.But then it is ad networks’ responsibility to optimise to max-imise clicks or conversions. In order to take the measurementof performance into account, ad networks usually employthe generalised second price auction (GSP) [13] which allowthem to apply bid biases (e.g the quality score) that usuallyweight the historical clickthrough rate (CTR) or conversionrate (CVR) heavily.
For the publisher side, an important research topic is toallocate impressions to premium contracts and ad networks.If chooses contracts, the publisher also needs to decide whichspecific contract to fulfil when multiple contracts havingoverlapping targeting rules. It is possible that ad networksbring in more revenue, but if a publisher sends too many im-pressions to this revenue channel and fails to fulfil a contract,he needs to pay the good-will penalty. In [27] the authorsused a two-phase models to sample, compute a compact al-location plan, and assign impressions online to advertiserswho submit contracts with overlapping targeting rules. Thisbalancing problem was discussed in [25] by using a certaintyequivalent control heuristic to show the necessity of usingboth channels to reach the global maximum revenue. In [15]it was generalised as an online stochastic optimisation prob-lem where given a set of resources, demands for resourcearrive online with associated properties. Given a generalpriori about the demands, one has to decide whether andhow to satisfy a demand when it arrives. The goal is tofind a valid assignment (strategy) with the maximum totalpayo↵.
When there were more and more ad networks, a problemled to the birth of ad exchanges: the excessive impressionsin some ad networks. It is preferable to have more demandthan supply because intense competition leads to higher rev-enue of both ad networks and publishers. However, whenthere are plenty of impressions unsold, ad networks try hardto find buyers. Besides, a common practice for advertis-ers was to register with multiple ad networks to find cheapinventories, or at least to find enough impressions withintheir budget constraints. They only found that managingnumerous channels di�cult and ine�cient (e.g how to splitthe budget). Ad exchanges, like Google AdX, Yahoo! RightMedia, Microsoft ad exchange, were created to address thisproblem by connecting hundreds of ad networks together,c.f. Block III in Figure 1. Advertisers now have a higherchance to locate enough impressions with preferred target-ing rules; publishers may received higher profit, too, becauseof more bidders potentially.
There are new research problems introduced by ad ex-changes. In the pioneer work of [20] the author discussedseveral issues including the truthfulness of auctions [19], call-out optimisation [9], arbitrage bidding and risk analysis, etc.The most significant feature introduced by ad exchanges isreal-time bidding, which queries bidders for a bid for every
'LUHFW�VDOHV��SULYDWH�FRQWUDFWV�5HDO�WLPH�ELGGLQJ�IORZ1RW�HVWDEOLVKHG
¼
'DWD�H[FKDQJH
$G�QHWZRUN
3XEOLVKHU$GYHUWLVHU
$G�QHWZRUN
3XEOLVKHU$GYHUWLVHU
3UHPLXP�DGYHUWLVHU
'HPDQG�VLGH�SODWIRUP
6XSSO\�VLGH�SODWIRUP
3UHPLXP�SXEOLVKHU
'LUHFW�VDOHV�VLQFH�����
$JJUHJDWHG�GHPDQG�DQG�VXSSO\�VLQFH�����
&HQWUDOLVHG�PDUNHWSODFH�DQG�UHDO�WLPH�ELGGLQJ�VLQFH�����
¼
¼
½
½
¾¿
$G�H[FKDQJH
$G�H[FKDQJH
,
,,
,,,
Figure 1: A brief illustration of the history and structure of dis-play online advertising. Ad networks were created to aggregateadvertisers and publishers. Ad exchanges were created to resolvethe unbalance of demand and supply in ad networks. Premiumadvertisers and publishers now choose to work with ad exchangesthrough demand-side platform (DSP) and supply-side platform(SSP) to take the advantage of real-time bidding (RTB). The ar-rows with number describe the process of RTB: an impression iscreated then passed to ad exchanges; advertisers are contactedthrough DSPs; advertisers choose to buy 3rd party data option-ally. Then following reversed path, bids are return to the adexchange, then the SSP or ad network, and winner’s ad will bedisplay to the user on the publisher’s website. The link betweenSSP and data exchange, as well as those among ad exchanges, arepotentially useful however not widely adopted in current market-place.
impression. Along with the query, ad exchanges send metadata of the webpage and the user. The exposure of thismeta data enables advertisers to switch from the inventory-centric to user-centric optimisation. There are also notewor-thy attempts to introduce concepts from finance market [28].These attempts will make the ad exchange more mature andattractive.When advertisers want to take advantage of RTB, they
work with ad exchanges through 3rd party platform that areusually referred to as demand-side platform (DSP). DSPsare delegates of advertisers that answer bidding requests andoptimise campaigns at the impression level. Comparing withad networks, the advantages of using DSPs are: 1) advertis-ers do not need to manage their registration with many adnetworks; 2) they can optimise at a finer granularity and ahigher frequency because of local impression logs instead ofaggregated reports from ad networks; 3) DSPs are also morecustomisable to better suit advertisers’ goals. For example,advertisers traditionally set a frequency cap on their cam-paigns (the maximum times to display the ad to the sameuser). Now the cap can be applied to user groups, or even asingle user for optimal e�ciency.At the other end, supply-side platforms (SSP) were cre-
ated to serve publishers. Similarly, SSPs provide a centralmanagement console with various tools for publishers’ ulti-mate goal: the yield optimisation. For example, SSPs allowpublishers to set a reserve price for a specific placement,or even against a specific advertiser. Some SSPs also allowpublishers to have preference over bidders (bid bias).In Figure 1 arrows with number describe the process of
RTB:
Real-‐3me bidding
Real-‐3me bidding
“This is Lawrence from India. I was searching Recommender model in web and found your webpage in search engine. Then, I visited your webpage searching relevant contents and saw unrelevant Google add in "Research Team" page (aRached screenshot). This add might vary from country to country. But I feel it will mislead and give wrong opinion to users who visit your webpage.” -‐ Lawrence from India
The winning bids and hourly average
Figure 7: The time series snippet of winning bids and its hourlyaverage of a single placement. The hourly average series peaksaround 6-8am every day when there are less impressions but morebidders.
less than 1% accepted by either the Shapiro-Wilk test [26]or the Anderson-Darling test [4], regardless of the numberof bidders or impressions in that hour.
2.2.2 The Daily Pacing
The daily pacing refers to the way that advertisers spendtheir budget in a single day. Usually an advertiser submitsa daily budget for his campaign, and choose from spend-ing it evenly throughout a day (uniform pacing) and as fastas possible (no pacing). The no pacing setup may lead topremature stop easily, which means the budget depletes tooquickly so advertisers cannot capture tra�c later in the day,that may have high quality impressions. An instance of pre-mature stop is illustrated in Figure 9 (day 1). The uniformpacing also su↵ers from the tra�c problem: if high qual-ity impressions appear in early part of the day, the pacingsetup would not be able to capture all of them; if there isnot enough tra�c in the late part of the day, the pacingsetup would not be able to spend all the budget (usuallycalled under-delivery). In sum they are not good daily pac-ing strategies although being widely used in DSPs.
In Figure 8 we plot the hourly mean of number of biddersand number of impressions, that were normalised respec-tively. There is a clear lag of when these two series reachtheir maximum in a day. Generally speaking, the number ofbidders peaks in the morning but the number of impressionspeaks afterwards. This lag indicates the unbalance of supplyand demand of the market in certain hours. For example,there are more bidders competing over limited impressionsin the morning, resulting in high winning bids as shown inFigure 7, which in turn costs more of advertisers. However,this higher cost does not necessarily lead to higher perfor-mance. From Figure 4 we can see that both post-view andpost-clicks CVR peak in the evening, which argues that in-tense bidding activities in the early hours are not reasonable.
This distribution of number of bidders throughout a daymay be due to the mixture of hour-of-day targeting and nodaily pacing setup. Most of advertisers wish to skip the lastnight hours because of the low CVR. Some of them may usethe no daily pacing setup to avoid the risk of under-deliveryas we discussed before. When these bidders start to bid in
Figure 8: The distribution of number of bidders and impressionsagainst hour-of-day. Their correlation plot shows a clear lag ofwhen they reach the maximum in a day. This lag indicates theunbalance of supply and demand of the market in certain hours.Besides, the fact that there are more bidders in the morning maybe due to the mixture of hour-of-day targeting and no daily pacingsetup. The plots used 3 months worth of data sampled from asingle placement. Note for some placements the lag was not veryclear.
Figure 9: An interesting instance we found in the dataset: anadvertiser switched from no pacing to even daily pacing. He wasbidding at a flat CPM. With ad exchanges, the large amount ofavailable inventories could deplete a daily budget quickly. How-ever, even daily pacing is far from optimal because it does notconsider the performance (e.g ROI) in di↵erent time slots. Notethat in practice the pacing engine would spend slightly more inthe beginning of the day to learn the available impressions and tocalculate the spending speed, especially for campaigns with smallbudgets.
the morning, they win lots of impressions with high bids,and they quit as their budgets depletes quickly. We leavethe validation of this hypothesis to the future works.A reasonable alternative is the dynamic pacing against the
performance. It is intuitively correct to spend more budgetin hours that generate more clicks or conversions, and lessin low performing hours.Note that this problem is not the same as a typical bud-
geted multi-arm bandit problem that has been discussed
Daily Pacing
Figure 7: The time series snippet of winning bids and its hourlyaverage of a single placement. The hourly average series peaksaround 6-8am every day when there are less impressions but morebidders.
less than 1% accepted by either the Shapiro-Wilk test [26]or the Anderson-Darling test [4], regardless of the numberof bidders or impressions in that hour.
2.2.2 The Daily Pacing
The daily pacing refers to the way that advertisers spendtheir budget in a single day. Usually an advertiser submitsa daily budget for his campaign, and choose from spend-ing it evenly throughout a day (uniform pacing) and as fastas possible (no pacing). The no pacing setup may lead topremature stop easily, which means the budget depletes tooquickly so advertisers cannot capture tra�c later in the day,that may have high quality impressions. An instance of pre-mature stop is illustrated in Figure 9 (day 1). The uniformpacing also su↵ers from the tra�c problem: if high qual-ity impressions appear in early part of the day, the pacingsetup would not be able to capture all of them; if there isnot enough tra�c in the late part of the day, the pacingsetup would not be able to spend all the budget (usuallycalled under-delivery). In sum they are not good daily pac-ing strategies although being widely used in DSPs.
In Figure 8 we plot the hourly mean of number of biddersand number of impressions, that were normalised respec-tively. There is a clear lag of when these two series reachtheir maximum in a day. Generally speaking, the number ofbidders peaks in the morning but the number of impressionspeaks afterwards. This lag indicates the unbalance of supplyand demand of the market in certain hours. For example,there are more bidders competing over limited impressionsin the morning, resulting in high winning bids as shown inFigure 7, which in turn costs more of advertisers. However,this higher cost does not necessarily lead to higher perfor-mance. From Figure 4 we can see that both post-view andpost-clicks CVR peak in the evening, which argues that in-tense bidding activities in the early hours are not reasonable.
This distribution of number of bidders throughout a daymay be due to the mixture of hour-of-day targeting and nodaily pacing setup. Most of advertisers wish to skip the lastnight hours because of the low CVR. Some of them may usethe no daily pacing setup to avoid the risk of under-deliveryas we discussed before. When these bidders start to bid in
Figure 8: The distribution of number of bidders and impressionsagainst hour-of-day. Their correlation plot shows a clear lag ofwhen they reach the maximum in a day. This lag indicates theunbalance of supply and demand of the market in certain hours.Besides, the fact that there are more bidders in the morning maybe due to the mixture of hour-of-day targeting and no daily pacingsetup. The plots used 3 months worth of data sampled from asingle placement. Note for some placements the lag was not veryclear.
Figure 9: An interesting instance we found in the dataset: anadvertiser switched from no pacing to even daily pacing. He wasbidding at a flat CPM. With ad exchanges, the large amount ofavailable inventories could deplete a daily budget quickly. How-ever, even daily pacing is far from optimal because it does notconsider the performance (e.g ROI) in di↵erent time slots. Notethat in practice the pacing engine would spend slightly more inthe beginning of the day to learn the available impressions and tocalculate the spending speed, especially for campaigns with smallbudgets.
the morning, they win lots of impressions with high bids,and they quit as their budgets depletes quickly. We leavethe validation of this hypothesis to the future works.A reasonable alternative is the dynamic pacing against the
performance. It is intuitively correct to spend more budgetin hours that generate more clicks or conversions, and lessin low performing hours.Note that this problem is not the same as a typical bud-
geted multi-arm bandit problem that has been discussed
choose from spending it evenly throughout a day (uniform pacing) and as fast as possible (no pacing).
The frequency factor
• How many Nmes ads (a.k.a. creaNves) would be displayed to a single user.
extensively in the literature [18, 12, 10]: 1) In this allo-cation problem the advertiser can only explore the currenttime unit; 2) There is potential over or under-delivery in atime unit due to the latency in the practical implementa-tion. Therefore revising the remaining budget is requiredafter every time step.
2.3 Conversion Rates and Selective BiddingConsidering the maintenance cost (e.g servers and band-
width) it is not wise for an advertiser to submit a bid everytime he receives a bidding request. Among various factorsthat he uses to decide to bid or not, the frequency and re-cency factor are important ones that have not been givenenough attention to.
2.3.1 The Frequency Factor
The frequency factor (or frequency cap, FC) defines howmany times ads (a.k.a. creatives) would be displayed to asingle user. It can be applied to campaign groups, cam-paigns, and creatives separately. For example, given a cam-paign with 3 ads, an advertiser could set FC(campaign) =6, FC(ad1) = 2, FC(ad2) = 3, and FC(ad3) = 4 and thesesettings would work together. The same user would at mostsee ad1 twice, ad2 three times, ad3 four times, and all adsdisplayed to him together would not exceed six times.
The frequency factor is normally set based on historicaldata and needs to be adjusted constantly during the flighttime of the campaign, because di↵erent campaigns ask forvery di↵erent FCs, as illustrated in Figure 10. The campaign1 from the left plot received the highest CVR with 6-10 im-pressions, which is also true for the cumulative CVR metrics.However the campaign 2 from the right plot received thehighest CVR with 2-5 impressions. If the campaign 1 sets afrequency cap of 2-5 impressions, most of conversions wouldnot be achieved. If the campaign 2 sets a frequency cap of6-10, nearly half of impressions could be wasted. Therefore,a good FC setting is crucial to the e�ciency of advertising.
Before setting an optimal FC, we need to find the rightmetrics to measure the e�ciency of di↵erent FCs. See theexample in Table 1 where we compare popular metrics. As-sume there are 100 users; the CPM is fixed at 10; the ad-vertiser’s conversion goal is worth 500. From the table wecan see using di↵erent metrics could lead to very di↵erentdecision. If the advertiser uses CVR as most of advertisersdo, he would go for FC=3; if the advertiser cares more aboutthe total number of conversions, he would use FC=5; if theadvertiser cares more about CPA he would use FC=3, too;if the advertiser measures the performance against the profitand cost, he would go for FC=2.
Using FC = 2 seems the most profitable. However, choos-ing FC = 3 is reasonable, too, especially when consideringthe long term impact: FC = 3 gives more conversions at lowcost, and once these users are attracted they have a higherchance of converting again in future.
2.3.2 The Recency Factor
The recency factor (or recency cap, RC) helps to decideto bid or not based on how recently the ad was displayedto the same user. It also works at di↵erent levels includingcampaign groups, campaigns, and creatives. For example,an advertiser can set RC(campaign) = (1 hour) so that allads from this campaign would be displayed to the same useronly once in every hour. Similar to the frequency factor,
Figure 10: The frequency against CVR plot from two di↵erentcampaigns. The campaign 1 from the left plot received the high-est CVR with 6-10 impressions, which is also true for the cumu-lative CVR metrics. However the campaign 2 from the right plotreceived the highest CVR with 2-5 impressions. If the campaign1 sets a frequency cap of 2-5 impressions, most of conversionswould not be achieved. If the campaign 2 sets a frequency cap of6-10, nearly half of impressions could be wasted.
fc cvr (cumulative) cvr convs cpa roi
1 0.0000 0.0000 0 - 0.00
2 0.0150 0.0150 3 667 1.50
3 0.0067 0.0167 2 600 1.25
4 0.0025 0.0150 1 667 1.00
5 0.0020 0.0140 1 714 0.88
6 0.0000 0.0117 0 857 0.70
7 0.0000 0.0100 0 1000 0.58
Table 1: The comparison of di↵erent metrics against frequencycaps (FC). Note cvr and convs are fc specific, i.e. the extra valuecould be gained by increasing the frequency cap from the previouslevel to the current one. The cumulative cvr is the CVR adver-tisers use: total conversions divided by total impressions. Thecpa gives the cost-per-acquisition. The roi gives the return-on-investment based on the advertiser’s conversion valuation. Thesevalues are calculated by assuming there are 100 users; the CPMis fixed at 10; the advertiser’s conversion goal is worth 500.
RCs are useful to achieve the best advertising e�ciency. Forexample, displaying ads intensively incurs high cost but littlee↵ect (or even getting people annoyed) for some campaigns(e.g financial services). Users need to think, compare, thenmake decisions. A better strategy is to display the samead right after the thinking time to remind users. However,for some campaigns (e.g booking flight) users would convertvery quickly or not convert at all, which requires relativelyintense advertising.Figure 11 plots the RC against CVR of two di↵erent cam-
paigns in the dataset. Both campaigns show the highestCVR at the 1-5 minutes level. However for the campaign 2the CVR is still not negligible after a long time (14-30 days).If the advertiser uses a more strict RC setting (e.g do notdisplay ads to users who were first exposed 14 days ago) hecould lose potential conversions. On the other hand, using aloose RC setting for the campaign 1 will only waste budgetsince the CVR is very low after 14 days.Another interesting observation is given in Figure 12. Sim-
ilarly to the frequency factor, the analysis of e�ciency ofRCs requires metrics based on the understanding of adver-tising goal, which we do not repeat here.
The Recency Factor • helps to decide to bid or not based on how recently the ad was displayed to the same user.
Figure 11: The recency factor against CVR plot from two di↵erentcampaigns. Both campaigns show the highest CVR at the 1-5minutes level. However for the campaign 2 the CVR is still notnegligible after a long time (14-30 days). If the advertiser uses amore strict RC setting (e.g do not display ads to users who werefirst exposed 14 days ago) he could lose potential conversions. Onthe other hand, using a loose RC setting for the campaign 1 willonly waste budget since the CVR is very low after 14 days.
The above analysis shows the importance of setting upproper FCs and RCs. At present these settings are at mostat the creative level. With real-time bidding, advertisers canpush them to a finer granularity: setting FCs and RCs forindividual users. The individual FCs and RCs can then beused to decide to bid or not on a specific impression.
3. CONCLUSIONIn this paper we introduced the history of real-time bid-
ding and discussed research issues related to the demandside. In fact, RTB, ad exchanges, and DSP are concep-tual ideas; what we really want to explore are the behaviourof advertisers and their delegates in the market, and thechallenges brought by the impression-level bidding and user-centric bidding, distinguished from bulk buying and inventory-centric buying. Through analysis of datasets acquired froma production ad exchange, we discovered that floor pricedetection, daily pacing, and frequency/recency setting areproblems not addressed. We explained their importance toadvertisers however leave the development and evaluation ofalgorithms to the future works.
4. REFERENCES[1] Internet advertising revenue report.
www.iab.net/insights_research/industry_data_
and_landscape/adrevenuereport (last visited27/6/2013).
[2] Abramson, M. Toward the attribution of webbehavior. In 2012 IEEE Symposium on CISDA (2012).
[3] Anagnostopoulos, A., Broder, A. Z.,Gabrilovich, E., Josifovski, V., and Riedel, L.Just-in-time contextual advertising. In Proceedings ofthe ACM CIKM 2007.
[4] Anderson, T. W., and Darling, D. A. Asymptotictheory of certain “goodness of fit” criteria based onstochastic processes. The Annals of MathematicalStatistics 23, 2 (1952), 193–212.
[5] Bharadwaj, V., Ma, W., Schwarz, M.,
Figure 12: The histogram of conversion window lengths since thefirst exposure. Post-view and post-click conversions are drawnseparately but are from the same campaign. The plot for thepost-view conversions roughly distinguish people into two groups:impulse purchaser who converted quickly and rational purchaser
who took time to consider after seeing the first ad. In order tomaximise conversions, the advertiser would consider to set upRCs to target these two types of users. Interestingly the plot ofpost-click conversions suggests that the time needed to fill theform and complete the purchase varied a lot for di↵erent users(or they could leave the page open for a long while).
Shanmugasundaram, J., Vee, E., Xie, J., andYang, J. Pricing guaranteed contracts in onlinedisplay advertising. In Proceedings of the ACM CIKM2010.
[6] Bilenko, M., and Richardson, M. Predictiveclient-side profiles for personalized advertising. InProceedings of the ACM SIGKDD 2011 (2011).
[7] Broder, A., Fontoura, M., Josifovski, V., andRiedel, L. A semantic approach to contextualadvertising. In Proceedings of the ACM SIGIR 2007.
[8] Chakrabarti, D., Agarwal, D., and Josifovski,V. Contextual advertising by combining relevancewith click feedback. In Proceedings of the ACM WWW2008.
[9] Chakraborty, T., Even-Dar, E., Guha, S.,Mansour, Y., and Muthukrishnan, S. Selectivecall out and real time bidding. Internet and NetworkEconomics (2010), 145–157.
[10] Chapman, A., Rogers, A., and Jennings, N. R.Knapsack based optimal policies for budget–limitedmulti–armed bandits.
[11] De Bock, K., and Van den Poel, D. Predictingwebsite audience demographics forweb advertisingtargeting using multi-website clickstream data.Fundamenta Informaticae 98, 1 (2010), 49–70.
[12] Deng, K., Bourke, C., Scott, S., Sunderman, J.,and Zheng, Y. Bandit-based algorithms for budgetedlearning. In Proceedings of the ICDM 2007.
[13] Edelman, B., Ostrovsky, M., and Schwarz, M.Internet advertising and the generalized second priceauction: Selling billions of dollars worth of keywords.Tech. rep., National Bureau of Economic Research,2005.
[14] Edelman, B., and Schwarz, M. Optimal auctiondesign in a multi-unit environment: The case of
Summary • Spot market: Real-‐Nme Bidding Shuai Yuan, Jun Wang, Real-‐Nme Bidding for Online AdverNsing: Measurement and Analysis, AdKDD’13 • Ad opNons Bowei Chen, Jun Wang, Ingemar Cox, and Mohan Kankanhalli, MulN-‐Keyword MulN-‐Click OpNon Contracts for Sponsored Search AdverNsing, under submission, 2013 h9p://arxiv.org/abs/1307.4980
Why Ad Futures (1) • Suppose there is a travel insurance company whose major customers are found through online adverNsing
Why Ad Futures (2) • In March the company plans an adverNsement campaign in three months Nme as they think there will be more opportuniNes to sell their travel insurance products in the summer
Why Ad Futures (3) • If the company worries that the future price of the impressions will go up, they could hedge the risk (lock in the campaign cost) by agreeing to buy (taking a long posiNon) the display impressions in 3 months Nme for an agreed price (taking a long posiNon in a 3-‐month forwarding market).
Why Ad Futures (4) • Equally, search engines and large publishers could agree to sell (taking a short posiNon) the display impressions in the future if worry the price will go down (lock in a profit).
• IntuiNvely, also useful for inventory management
However, online adver3sing is different
• A. Non-‐storability: unlike stocks or other common communiNes (petroleum and natural gas ), we cannot buy and keep an impression (thus ad display slot) for a period of Nme and sell it later in order to gain the profit by its price moment – This is in fact more similar to electricity/energy markets
• B. Not just about the price movements: the risk also lies in the uncertainty of the number of impressions (number of visits) and click-‐through rate in the future
�������������������
��������������
����������������
����������������������
�������������
�����������
�������������
�����������
�����������������������
��������
������������������������
���������������
������������� ������������������ ������������� ����������
���������������������� ��������� ������������
�������������
���������������
������� ���
�����������
���������������
��������
�����������
��������������
Figure 1: The classification of Internet advertising trading.
2.4 Competition methodsThere will always be competitions in trading. In order to
solve the competitions and generate winners, several waysare used as shown in Figure-1.
• 1st price negotiation, is normally operated by hu-man. For example, the publisher reviews contractssubmitted by advertisers and get in contact with thepreferred one;
• 1st price auction, was created in 1996 by Open Textand then Goto.com in 1998, for their cost-per-click pro-grams and gradually abandoned after the invention ofGoogle AdWords in 2000. Now some ad network star-tups propose the idea of ad futures (delivery impres-sions in future but not marketable), with which adver-tisers could place 1st price bids if they do not want too�er outright buy order [2].
• Some ad networks choose 1st price reservation todeal with ad futures with a first-come-first-serve ba-sis. The case could be considered as advertisers alwaysplace outright buy orders (assigning a contract at thespecified price immediately).
• 2nd price auction, or generalized second price auc-tion (GSP) [4] is the most popular in today’s ad net-works. Instead of paying for the bid o�ered, the ad-vertiser pays the price calculated from the next highestone and quality scores. For example, if advertiser Aplaces a bid of $10 with a quality score of 5, advertiserB places a bid of $8 with a quality score of 8, and theyboth target exactly the same audience. Then adver-tiser B wins because of $8 � 8 > $10 � 5, and the actualprice paid will be $10 � 5/8 = $6.25.Additionally the advertisers and publishers could choosebetween pre-set bidding (PSB) or real-time bidding (RTB)if they go with GSP. With PSB the advertisers setthe campaign, desired target, and bids before auctions;however with RTB the advertisers could bid every time
before an impression is delivered, adjusting their bidsaccording to user profiles and other factors as well.
2.5 Pricing modelsThe two sides of demand and supply must agree on a
pricing model as well, i.e. how much to pay for a unit good.Some popular models used are,
• Flat-rated (or cost-per-time, CPT) and cost-per-mille (CPM) are most selected when delivery is in thefuture. The first one indicates that the cost (of somead slot) will be constant (for some time) regardlessof actual tra⇥c (number of impressions) delivered bypublishers. The second one take actual tra⇥c into ac-count, however the CPM price will be fixed due to thefact that no auction is held against other competitors;
• In spot markets the preferred pricing models are cost-per-mille, cost-per-click, cost-per-action, cost-per-lead, cost-per-view, and cost-per-complete-view, as explained in Section-1.Additionally, if the advertiser choose to use RTB, thepossible pricing model at present will be CPM only.The total cost will receive greater variance due to thefact that the advertiser has to place the bid every timean impression is delivered. As for ad exchanges andpublishers, they tend to receive higher revenue sinceadvertisers are willing to bid more, given that everyimpression is better matched with visitors.
Regardless of the choice of pricing models, in Internet ad-vertising the goods traded are always impressions, even ifthe advertisers choose to use CPC, CPA, or others. By pro-viding various charging models in ad networks, more adver-tisers will be attracted while they are alleviated from thepain of calculating how many impressions are needed to geta click or a conversion, thus how much should bid. The adnetworks take the calculation and relevance matching chal-lenges over; they have enough data to solve the problem as
C. There are many pricing schemes Based on cost per click, cost per impression etc.
However, online adver3sing is different
Ad Op3ons Pricing is the Core • Due to the differences menNoned, financial models and theories cannot be directly employed – need to rethink what is the “fair” price for ads
• We have developed a novel ad opNon pricing model to understand a “fair” price in various senngs, e.g., sponsored search, contextual adverNsing, and banner ads
GSP-Based Keyword Auction
Search engine
Online adver3sers
1st ad slot £ 0.2
£ 0.32
£ 0.15
£ 0.22
2nd ad slot
…
£ 0.2
£ 0.22
Choose the adverNser with the highest bid, but normally charge him based on the second highest bid price, called Generalised Second Price (GSP) auc3on model.
• VolaNlity in revenue.
• Uncertainty in the bidding and charged prices for adverNsers' keywords.
• Weak brand loyalty between the adverNser and the search engine.
Problems of Auction Mechanism
Multi-Keyword Multi-Click
Advertisement Option Contract
for Sponsored Search [B.Chen et al. 2013]
An opNon is a contract in which the opNon seller grants the opNon buyer the right but not the obligaNon to enter into a transacNon with the seller to either buy or sell an underlying asset at a specified price on or before a specified date. The specified price is called strike price and the specified date is called expiraNon date. The opNon seller grants this right in exchange for a certain amount of money at the current Nme is called opNon price.
Option Contract
online adver3ser search engine
sell a list of ad keywords via a mulN-‐keyword mulN-‐click opNon
t = 0
mul3-‐keyword mul3-‐click op3on (3 month term)
upfront fee (m = 100) keywords list fixed CPCs
£5
‘MSc Web Science’ £1.80
‘MSc Big Data AnalyNcs’ £6.25
‘Data Mining’ £8.67
t = T
Timeline
submit a request of guaranteed ad delivery for the keywords ‘MSc Web Science’, ‘MSc Big Data AnalyNcs’ and ‘Data Mining’ for the future 3 month term [0, T], where T = 0.25.
t = 0 pay £5 upfront opNon price to obtain the opNon.
Selling & Buying An Option
32 32
online adver3ser search engine
t = T
Timeline
exercise 100 clicks of ‘MSc Web Science’ via opNon.
t = 0
pay £1.80 to the search engine for each click unNl the requested 100 clicks are fully clicked by Internet users.
t = t1
reserve an ad slot of the keyword ‘MSc Web Science’ for the adverNser for 100 clicks unNl all the 100 clicks are fully clicked by Internet users..
t = t1c
Exercising the Option
33 33
online adver3ser search engine
t = T
Timeline
if the adverNser thinks the fixed CPC £8.67 of the keyword ‘Data Mining’ is expensive, he/she can a9end keyword aucNons to bid for the keyword as other bidders, say £8.
t = 0
pay the GSP-‐based CPC for each click if winning the bid.
t = …
select the winning bidder for the keyword ‘Data Mining’ according to the GSP-‐based aucNon model.
Not Exercising the Option
Benefits of Ad Option
Adver3ser Search Engine
§ secure ad service delivery § reduce uncertainty in aucNons § caps ad cost.
§ selling the inventory in advance; § having a more stable and predictable
revenue over a long-‐term period;
§ Increasing adverNsers’ loyalty
35 35
§ No-‐arbitrage [F.Black and M.Scholes1973; H.Varian1994]
§ StochasNc underlying keyword CPC [P. Samuelson1965]
§ Terminal value formulaNon
Bowei Chen, Jun Wang, Ingemar Cox, and Mohan Kankanhalli, MulN-‐Keyword MulN-‐Click OpNon Contracts for Sponsored Search AdverNsing, under submission, 2013 h9p://arxiv.org/abs/1307.4980
Ad Option Pricing: Building Blocks
Ad Option Pricing: Formula
§ n=1, Black-‐Scholes-‐Merton European call
§ n=2, Peter Zhang dual strike European call
§ n>=3, Monte Carlo method
Bowei Chen, Jun Wang, Ingemar Cox, and Mohan Kankanhalli, MulN-‐Keyword MulN-‐Click OpNon Contracts for Sponsored Search AdverNsing, under submission, 2013 h9p://arxiv.org/abs/1307.4980
Ad op3on pricing for the keywords `canon cameras', `nikon camera‘ and `yahoo web hos3ng‘.
Experimental Results: Monte Carlo Simulation
Fig. empirical example for the keyword ‘lawyer’, where the Shapiro-‐Wilk test is with p-‐value 0.1603 and the Ljung-‐Box test is with p-‐value 0.6370.
39 39
About 15.73% of keywords that can be effectively priced into an option contract under the GBM.
The test of Arbitrage
41 41
42 42
§ n=1, analyNcal proof + numerical valuaNon
§ n>=2, numerical valuaNon
Bowei Chen, Jun Wang, Ingemar Cox, and Mohan Kankanhalli, MulN-‐Keyword MulN-‐Click OpNon Contracts for Sponsored Search AdverNsing, under submission, 2013 h9p://arxiv.org/abs/1307.4980
Effects of Ad Options on Search Engine’s Revenue
43 43
An empirical example of search engine's revenue for the keyword `equity loans'.
:20 B. Chen et al.
4 4.5 50
0.1
0.2
0.3
0.4
0.5
0.6
Fixed CPC
Revenueofsearchengine
(a) GBM
Reve differenceOption priceExp spot CPC
4 4.5 50
0.1
0.2
0.3
0.4
0.5
0.6
Fixed CPC
Revenueofsearchengine
(b) CEV
Reve differenceOption priceExp spot CPC
4 4.5 50
0.1
0.2
0.3
0.4
0.5
0.6
Fixed CPC
Revenueofsearchengine
(c) MRD
Reve differenceOption priceExp spot CPC
4 4.5 50
0.1
0.2
0.3
0.4
0.5
0.6
Fixed CPC
Revenueofsearchengine
(d) CIR
Reve differenceOption priceExp spot CPC
4 4.5 50
0.1
0.2
0.3
0.4
0.5
0.6
Fixed CPC
Revenueofsearchengine
(e) HWV
Reve differenceOption priceExp spot CPC
4 4.5 50
0.1
0.2
0.3
0.4
0.5
0.6
Fixed CPC
Revenueofsearchengine
(a) GBM
Reve differenceOption priceExp spot CPC
4 4.5 50
0.1
0.2
0.3
0.4
0.5
0.6
Fixed CPC
Revenueofsearchengine
(b) CEV
Reve differenceOption priceExp spot CPC
4 4.5 50
0.1
0.2
0.3
0.4
0.5
0.6
Fixed CPC
Revenueofsearchengine
(c) MRD
Reve differenceOption priceExp spot CPC
4 4.5 50
0.1
0.2
0.3
0.4
0.5
0.6
Fixed CPC
Revenueofsearchengine
(d) CIR
Reve differenceOption priceExp spot CPC
4 4.5 50
0.1
0.2
0.3
0.4
0.5
0.6
Fixed CPC
Revenueofsearchengine
(e) HWV
Reve differenceOption priceExp spot CPC
Fig. 8. Empirical example of search engine’s revenue for the keyword ‘equity loans’.
Next, Figure 9 illustrates the case where we have 2 candidate keywords: ‘iphone4’and ‘dar martens’, where r = 5% and ⇢ = 0.0259. In Figure 9(a), we see that thehigher the fixed CPCs the lower is the option price. This property is the same as forthe single-keyword options. Also, the calculated option price achieves maximum whenall the fixed CPCs are zeros. Figure 9(b) then shows the revenue difference curve ofthe search engine, where the red star represents the value when F1 = EQ
t
[C1(T )] andF2 = EQ
t
[C2(T )]. The expected revenue differences are all above zero, showing thatthis 2-keyword ad option is beneficial to the search engine’s revenue. However, aninteresting point to discuss is that the red star point is not the maximum differencerevenue, which is different from single-keyword options. This may be due to the factthat the underlying CPCs move in a correlated manner and the advertiser switcheshis/her exercising from one to another. The revenues’ difference curve in Figure 9(b)is very smooth while Figure 10(b) shows a bit volatile pattern because the underlyingcorrelation increases. Above all, the properties of the revenue difference are similar tothose of single-keyword options and they are all positive.
It would be impossible to graphically examine the revenue difference for higher di-mensional ad options (i.e., n � 3). However, based on the earlier discussions, we cansummarize two properties. First, there are boundary values of the revenue differences.If every F
i
! 0, D(F) ! 0; if every F
i
! 1, D(F) ! 0. Second, there exists a maxi-mum revenue difference value even though this may not at the point F
i
= EQt
[C
i
(T )].Overall, we are able to say that a proper setting of fixed CPCs by a search engine canincrease the ad revenue compared to keyword auctions.
5. CONCLUDING REMARKSIn this paper, we proposed a new ad selling mechanism for sponsored search that bene-fits both advertisers and search engine. On the one hand, advertisers are able to secure
† This manuscript is under submission to a journal.
Concluding remarks: “Recipe for Disaster”-‐ Model That “Killed” Wall Street?
• Wall Street’s math models for minNng money worked brilliantly... unNl one of them devastated the global economy
• David X. Li's Gaussian copula funcNon as first published in 2000. Investors exploited it as a quick way to assess risk (the chance a company is likely to default)
• As Li himself said of his own model: "The most dangerous part is when people believe everything coming out of it.”
44 h9p://www.wired.com/techbiz/it/magazine/17-‐03/wp_quant?currentPage=all
For more informa3on, please refer to
Shuai Yuan, Jun Wang, Real-‐Nme Bidding for Online AdverNsing: Measurement and Analysis, AdKDD’13 h9p://arxiv-‐web3.library.cornell.edu/abs/1306.6542 Bowei Chen, Jun Wang, Ingemar Cox, and Mohan Kankanhalli, MulN-‐Keyword MulN-‐Click OpNon Contracts for Sponsored Search AdverNsing, under submission, 2013 h9p://arxiv.org/abs/1307.4980
Thanks for your aRen3on