The following is an exclusive guest contributed post to MMW from Rajiv Bhat, Senior Vice President of Data Sciences and Marketplace at InMobi.
Mobile usage has surpassed desktop usage with over 65 percent of all digital time being spent on the phone. The mobile advertising landscape has evolved in two ways:
- The volume of advertising inventory has exploded, and there are more ways to reach the user on the phone than ever before;
- There is more data available and at the astute marketer’s disposal.
The former make it seem that media has become cheaper, but that it has also become easier to spend a lot of money without learning or getting much in return. The latter trend has created a data-driven movement towards greater efficiency, and the sheer volumes of data have rendered simple spreadsheet based optimizations untenable. This article will explore signals and strategies that marketers can effectively deploy data to maximize returns on their spend.
Signals and Strategies to Effectively Deploy Data
There are broadly six types of raw data signals available to advertisers on mobile; handset, demographic (age, gender, income), geographic (location), advertiser relationships, app engagement, and transactions. While the first two are straightforward, the rest merit a brief description. Geographic data refers to location data, typically available with a timestamp. This can be at granularities varying from real-time location information accurate by a few meters to zip or country level. Advertiser relationships capture historical interactions between the user and the advertiser. The user may have completed multiple transactions with the advertiser or are yet to see a single advertisement from the advertiser. App engagement describes the amount of time a user spends on various applications, websites and categories. Transaction data is about understanding the kinds of goods and services the user has purchased (not limited to transactions with the advertiser). While extremely valuable, this is also the most difficult information to come by for most advertisers.
A useful way to think about audience segments is to consider them as meaningful abstractions of one or more raw data signals described above. For instance, it’s useful to describe users playing Candy Crush, a popular game where gamers get points for striking through connected gems, as ‘Casual Gamers’. Similarly, users who are consistently seen between the same two locations everyday can safely be labelled as ‘commuters’. Feature aggregations are powerful both for targeting and optimization. For games monetizing through in-app purchases, it’s useful to create segments that indicate high disposable income, perhaps users with high priced handsets? Storing as many features, and abstractions of features, as available is invaluable for modelling. Robust machine learning techniques exist for picking only the most relevant features.
Once appropriately equipped, a trifecta of data strategies can be deployed to suitable gain:
- Discard the Chaff
For brand new campaigns, exploration refers to finding new segments that are meaningful to the campaign. Exploration is more expensive than one would think. Consider the case where a developer is advertising a game where good target segments are unclear. Here, a segment could be defined by an app, country or operating system tuple. Let us assume canonical click through and conversion rates of one percent each, and inventory costs corresponding to a dollar for a thousand impressions. Consider a modest spread of 10,000 segments for which we want to drive an average of one download per segment in order to build some semblance of a distribution that works. One download would take on average 10,000 impressions and would cost $10. While 10,000 downloads would cost $100,000. Even at these numbers, the learning would be suboptimal. For a broader range of segments, the learning would be more diluted.
Campaigns can be quickly re-optimized by appropriately weighing events down the funnel and evaluating performance at aggregations. A learning mechanism such as a multi-armed bandit is handy here. A simple strategy could be to spend 90 percent of the budget on the highest performing segments, while exploring new segments with the other 10 percent. Frequency capping can also be leveraged to restrict impressions served to a single user facilitate exploration.
- Price Correct
Almost all mobile ad inventory (unless cross promotion is happening across apps) is obtained via participation in an auction. Auctions help determine the market price for a particular slice of inventory, but the market price may deviate significantly from the buyer price. For first price auctions, if the marketer started off at the maximum bid, she will need to constantly trim bids till there is no drop in the win rate. As difficult as it sounds, it’s made tougher by a constantly changing market landscape. Cross subsidization is a meaningful strategy in first price scenarios, since scale is obtained by using gains from cheap inventory to buy inventory in more expensive segments.
Most auctions today are second price, where the marketer puts down a maximum willingness to pay and is only charged the second highest bid or the minimum price. Unfortunately, a vast number of auctions are sold at a minimum price, and often the highest bid is a factor in determining the minimum price. So even in a second price setup, one could find the price of the inventory creeping up to the bid if not constantly adjusted. A common price optimization tactic in second price scenarios is to find elasticity by probing the bid landscape, and finding the best value for the incremental dollar. A lot of audience strategies fail to deliver return on investment (ROI) due to poor price optimization and hence, targeting and pricing should be considered holistically.
- Go All In
Advertising against ROI targets requires aggressive modelling. This means leaning towards false positives as against false negatives in modelling – it’s better to pay for events that don’t convert rather than lose out on potential users by not bidding. Another implication is the need for strong machine learning techniques. One technique that works reliably is combining the outputs of multiple machine learning models to make decisions. In an ensemble of models, each model might predict the same quantity (e.g. a transaction) or a different parameter in the sales funnel. Another data technique is the use of lookalike modelling. In this method, machine learning models predict the likelihood of prospects being similar to a cohort of highly engaged users. A third method entails the use of feedback loops to dynamically change predictions and bids. Often there is a significant time delay between the purchase of inventory and the actual transaction. Typically, conversions will exponentially taper off, but factoring this delay into prediction models provides additional accuracy.
In summary, the key aspects of a data driven scaling strategy include collecting as many signals as available, storing them as many meaningful abstractions, using exploration strategies wisely, using pricing levers to improve costs and fine tuning using aggressive models.