"I have mixed feelings when I hear attribution modeling come up on a panel," Hamman said, adding, "There is no 100% right answer."
To illustrate the point, Hamman gave a personal anecdote about how his children did numerous searches – on their own browsers – to research the best hotel to stay at on a family trip. Hamman then booked the hotel via his browser.
"Put an attribution model around that?," he said, adding, "There is no 100% answer. Don't kill yourself over it."
We partly agree with Dax’s comment. It is important to understand that there are no “100%” correct answers in any business setting. George Box, one of the greatest statisticians, had a famous quote “All models are wrong but some are useful”. Models should be built to help better predict future conversions for advertisers. Attribution models are not an end in itself but are just one way of addressing the allocation problem. Hence, the real problem we should be solving for advertisers is optimal budget allocation across multiple online channels.
Attribution models with additional modeling approaches such as modeling cross-channel demand elasticity and temporal effects of event realization will help allocate resources optimally. It is imperative to have out-of-sample validation of any allocation model that one builds. Also, using the exploration-exploitation framework that is common in machine learning literature, one can in fact setup online experiments that provide a clean control group. Finally, we should realize that while comparing models, we should evaluate them against each other rather than discarding all predictive approaches as they may not predict with 100% accuracy.