Optimal Bidding Strategies for Keyword Auctions
In the earlier blog post we discussed about sponsored search marketing and the mechanics of auction design. In this post we shall look deeper into the challenges of online keyword advertising auctions among multiple bidders with limited budgets, and try to come up with the bidding strategy that will increase the expected utility of the advertiser.
Data Extraction and Classification:
Data was collected from one of the clients running campaigns for adwords. Data was collected on a daily frequency and had information regarding the performance and the characteristics of the keyword.
Below you can find the description of some of the attributes that were considered:
|Quality Score||Quality Score is the measure of quality and relevance of the keyword|
|CTR||Click through rate|
|CPC||Cost per click|
|Keyword cluster||Feature that was created on basis of average position and quality score obtained by keyword for past some days.|
|Average Position||A statistic that describes how your ad typically ranks against other ads. This rank determines in which order ads appear on the page.|
|Revenue per click||Evaluation of a click by an advertiser.|
Below we discuss some of the engagement metrics likes that capture the interaction of user with our ad.
- Total clicks.
- CTR ( Click Through Rate ).
One of the engagement metrics is total number of clicks, to see if there is any trend in the total number of clicks that we receive for a keyword we plot the activity for a time period. We take the rolling mean and plot that curve which is much more smooth, we can see that there is no clear trend which suggests that naive methods of prediction would not work so well.
Below we discuss some of the optimization metrics that capture the sales etc. for a domain.
- Conversion Rate
- Return on Investment
We wanted to find out how frequently do we encounter updates for revenue on our platform for a given keyword. So we plotted this graph to find out days which generated revenue and the frequency of the event. So as we can see that this is a very rare event so any model that would directly try to predict revenue should have a balanced set of examples to be of any use.
In the above figure we plot the relationship between average position and total number of clicks for a given keyword for a given date range. This shows that there is some kind of negative linear relationship between the two variables, as we go about achieving the higher average positions our chances of getting a click also tend to take a hit.
Next we tried to model bids against total number of clicks, so that we can consider different bid values in a range which depends on the keyword cluster that we discussed earlier. Some keyword clusters require us to explore a wider range of bid values and for some we can get away with not so many values as well.
We created models for estimating total clicks and bid values based on past performance and for every keyword we evaluated the expected utility, expected revenue, expected cost for every bid value in the range and the idea is to maximize the expected utility for a given keyword. So this gives us an indication of what bid values would work well for a given keyword.
Now we were able to get some sense of what bid values for an individual keyword works but a portfolio is termed as a success when we are able to achieve the maximum utility as a whole while respecting the budget constraint. The ideal scenario would be to pick the best bid values for each of the keyword and still not overshooting the total budget. But for the other more frequent scenarios we can solve this problem by choosing the best possible combination of bid values for all of the keywords so that maximize the objective. We are calling the routine as Equalizer.
In this work we have tried to find a optimal bidding strategy using the relationship between bid values and clicks that would increase the utility of a keyword and tried to find with right combination of bid values across the portfolio to maximize the objective as a whole.