At MailChimp, we retrain our models offline once a quarter, test them, and then promote them into production.

You name it; if there's good historical data, a trained AI model can help.

best model is one that strikes the right balance between functionality and maintainability. If an analytics model is never used, it's worthless.

just cleverly turns past data into a formula or set of rules that it uses to predict a future case. As we saw in the case of naïve Bayes in Chapter 3, it's the AI model's ability to recall this data and associated decision rules, probabilities, or coefficients that make it so effective.

Identify the analytics opportunities you want to tackle in as much detail as possible before acquiring tools.

And quantifying the forecast uncertainty by creating prediction intervals is invaluable and often ignored in the forecasting world.

This story about Target is just one of many that have peppered the press recently. Watson won Jeopardy!. Netflix offered a million dollar prize to improve its recommendation system. The Obama re-election campaign used artificial intelligence to help direct ground, online, and on the air media and fundraising operations. And then there's Kaggle.com,

Just when you think you have a good demand forecast, a flood disrupts your supply chain, limiting your supply, forcing you to raise prices, and throwing your sales completely out of whack.

a non-pregnant customer bought maternity clothing and stopped buying cigarettes, and the model gave them a score of 0.76. It's clear then that if you are going to use these predictions in real marketing efforts, you need to set a score threshold for when you can assume someone is pregnant and reach out to that person with marketing materials.