I am exploring ways of using adaptive models to predict the best day of the week and hour of the day to send an outbound action (specifically email) to a given customer. The rationale for doing this is that many businesses experiment to find out when their emails get the most engagement, and will often run campaigns during the days/hours when engagement is highest. Our hypothesis is that the "best" day or hour varies by individual as much as the best action, and optimizing for day/hour can increase engagement (clicks in this case).

I've come up with two approaches, and I'm interested to know the product team's thoughts.

Using Send Day and Send Hour as parameterized predictors

In this approach, the DateTime of the email send is transformed into two data points: Send Day (Monday, Saturday) and Send Hour (0 for 00:00-00:59, 7 for 07:00-07:59, 14 for 14:00-14:59, etc.). Send Day and Send Hour would be included as parameterized predictors in the OOTB email treatment model.

The rationale behind this approach is that it provides a quick time to value, and it allows the business to validate the assumption that day and hour are relevant with minimal configuration. However, it does not prevent an email from being sent until the best day or time. The relevance of day/hour will be included in the calculation of overall propensity, and the highest propensity action will be selected for a customer in a given run, regardless of whether it's the right day or time. It could have some impact, however, if Send Day or Send Hour have high predictive value.

Creating New Models with Day and Hour as Additional Model Context Dimensions

In this approach, new predictions are built based on adaptive model rules that include Send Day and/or Send Hour in the model context. The rationale behind this approach is that customer behavior varies by day and hour, therefore, Send Day and Send Hour really belong in the model context. With this approach, an email would not be sent unless the propensity from these models (or one single model?) is above a specified threshold.

I think this would create many (potentially hundreds or thousands) of models, and intuitively I think this means it would take quite awhile for them to learn. Additionally, reasonable threshold values wouldn't be known until the models had matured.

Operationally, taking either of these approaches would mean running the outbound schedule on an hourly basis, probably within a specified window. For example, it wouldn't make much sense to run the schedule in the wee hours of the morning, since most customers of a regional bank would be asleep during that time.

Could either of these approaches work? Do you have other ideas? Could they be implemented iteratively, using the first approach to validate the relevance of day and hour, and if so, using the second approach to optimize?