My unwritten actuary spec-fic involves a secret NSA project to build a massive predictive model trained on the internet text emails, sms, transcribed phone calls, purchases, health records, etc.
(This idea is 5 years old… Now of course it would just be an LLM.)
The goal is to detect terrorists or whatever, but it can’t of course because the sample size is too small. So when prototyping they pick a more common event – death-- and a life actuary is brought on to test its predictions against industry models.
It’s not super great at predicting deaths either, unless people go to the hospital or explicitly talk about their symptoms. But combined with existing models there is a small improvement that’s hard to explain. The actuary digs deeper into the test data, he finds that many of the new positives are divorcees for some reason.
At this point he pulls together a team to test for errors (data contamination is the obvious answer), but instead they find that the model is internally predicting who gets divorced, and divorce just happens to be a death covariate. And it turns out the model is actually really good at predicting divorce-- decades in advance.
Because whether people are truly compatible in the long run, is a reflection of their personalities, and how they feel about each other, which is unconsciously revealed in how they talk to one another. Not just what they say, but how they say it. Eventually, all of the main characters are driven to look themselves up, to find out whether they are probabilistically destined for divorce.
Throughout the story, I want to incorporate little emails and texts between the protagonist and his spouse. Nothing shocking, but enough to get a sense of their marriage.
The thesis of the story is that you if you could just look really, really closely at those same little missives, you could know if your love is true.