• by Mert Bay , January 18, 2016

Game theory just sounds fun. It rings of intriguing calculations done in the interest of entertaining or worthwhile pursuits: How to choose your March Madness bracket. How to get your spouse to do the dishes.How to make kidney donation more efficient. And of course, for us marketers, how do we measure cross-channel campaigns, or the effect of multiple touchpoints on a conversion?

Game theory seems like a logical solution to this last challenge. A mathematical model that measures interaction effects between multiple “players” -- in our case, advertising channels -- it purports to quantify how marketing actions work together to impact sales. And it works! Collaborative game theory can provide definite insights beyond last-click attribution. Drop your data into the formula, and it returns which channels are contributing how much to the final result. Great.

Here’s the catch: The game theory model only works for some data, some of the time. Like all single models, it has strengths and weaknesses.  Game theory works best when every player is observed in isolation and in various combinations with other players. In our case, players are marketing channels.  The total utility (conversions) is attributed between the channels based on those observations, so that each touchpoint is theoretically given its due.

The problem with this approach is that it is inherently observational, not predictive.  It anticipates outcomes based on the premise that we know how every “player,” or channel, will react under a certain set of circumstances based on previous observations, and that each player is only affected by the other players in the game. But in real life, that’s rarely the case.  There are innumerable market forces that can change the equation.

Game theory attempts to solve this challenge by accounting for all of the different combinations of “players,” but each assumption leads to a different unstable equilibrium. In comparison to other predictive models, game theory is more susceptible to performance degradation when input data is noisy -- with affiliate channels, for instance. It can also be difficult to establish an accurate baseline with an observational model. Game theory may produce accurate results some or even most of the time, but it can’t adapt quickly or effectively enough to guarantee precise, and ongoing, attribution insights.

To be fair, most algorithmic models suffer from similar challenges. No one algorithm can generate predictions that are 100% accurate, 100% of the time. But there are alternatives to a single algorithm model that bring greater breadth and depth to your attribution data. Data scientists are now combining models, pulling from the “wisdom of crowds” theory to take the best from each model and leave the rest. It’s an evolving science, and one that I’m excited about.

Game theory is fun. It’s interesting and effective, too, from sports to marriage to, yes, marketing attribution. But it’s simply one model: a one-size-fits-all approach to the very diverse world of business and advertising. Game theory is a great tool to have in your attribution toolbox. I just don’t think it should be the only one.