sending time optimization
Emarsys in numbers
2.5B contacts
1.2B unique contacts
first idea
every day usage
~ 1 year
1. what & why
2. how
3. how much
truly personal
optimize within a day
vs.
available data?
-
historical sends
-
opens, clicks
-
mobile / desktop
can we optimize
based on past data?
opportunity for improvement?
differences in open rate
for one contact
systematic differences between campaigns by send hour?
differences in open rate for all contacts in reality
| 06:00 | 16:00 | 20:00 | SD | |
|---|---|---|---|---|
| contact_A | 0.5 | 0.33 | 0.4 | 0.085 |
| contact_B | 0 | 0.25 | 0.25 | 0.144 |
reality vs simulation
how to choose personal sending times?
agile data
at time of last open
bandit learning algorithm
algo trained for 3 years
bandit algorithm
bandit algorithm
-
arms: possible send times
-
winning: open of sent message
personal send hour for one contact over time
exploration ...
... exploitation
algo adaptation
-
priors
-
history of 12 months
-
2 hour intervals
piloting starts!
what to measure?
present performance
and
future potential
50%
50%
?
uplift in first 2 hours
uplift in open rate
uplift in click rate
in customer engagement
2-12% increase
handle recent campaigns
distorting priors
priors from STO emails
more piloting
non-random A/B testing
weekend effect with shifted sending for control group
-
send time opt. worth it
-
bayesian bandit algo
-
algo is only one component of success
startup safary send time optimization
By Czeller Ildi
startup safary send time optimization
Slides for 2017 Startup Safary about send time optimization at Emarsys.
