Your lost customers, whether they make particularly high-value purchases or not, are very valuable to you in terms of what you know about them.
Considering a customer as “lost” if they haven't made a purchase in, say, 90 days makes it very easy to send emails to all of your lost customers. This is fine if you sell a limited number of products and your customers shop on relatively similar frequencies, but what if you have some customers who shop weekly and some who only shop around the holidays? At the 90 day mark, your weekly customers have been sitting idle for almost three months. Meanwhile, you would expect your holiday customers to not shop with you once every 90 days.
RFM is a method of segmenting customers using three data points: how recently the customer purchased, how often they purchase, and how much they spend. You might create buckets for each of these attributes, and each segment is comprised of a unique combination of buckets (e.g., a segment might be “customers who last purchased within the last 90 days, who buy every two weeks and spend under $100 each time”). While this method takes into account the shortcomings of a “Days Since Last Purchase” approach, it doesn't factor in where a customer is in their lifecycle. Lifecycle transitions can have a large effect on likelihood of response. In addition, RFM doesn't try to predict behavior; it segments your customer base but doesn't make any assessment of how valuable those segments are.
Some types of advanced probabilistic models include pattern recognition and machine learning algorithms. There are many approaches, but the general idea is that these approaches look at individual customer stories and whether their types of ordering behavior are indicative of customers who drop off the map. With these types of models, you will be able to reach the ideal where you're alerted about a customer who orders weekly not ordering for two weeks and a customer who orders monthly not ordering in the last 45 days. While these types of models may seem ideal, they are complicated and often require a lot of data science help and computational horsepower to get up and running. The trade-off between complexity and predictive power you choose to make really depends on the resources at your disposal.
If you’re interested in learning more about probabilistic models and techniques, take a look at Bruce Hardie's site, and in particular, the “Forecasting Repeat Buying for New Products and Services” tutorial and associated Excel spreadsheets (Part 1, Part 2).