Recency Signals: You’re Only As Good as the Last Thing You Did

It took hundreds of experiments to invent the light bulb. But Edison became famous for the final experiment, the one that worked.

The saying “you’re only as good as the last thing you did” is a common expression that suggests a person’s value or reputation is based on their most recent performance or achievement. This saying is often used in various fields, including sports, entertainment, and business, to emphasize the importance of consistently delivering high-quality work and maintaining a good reputation. The idea behind this saying is that past successes or achievements are less relevant than a person’s current performance and that each new project or task represents an opportunity to demonstrate one’s abilities and build a new reputation for excellence.

In feature engineering, a recency signal is derived from the latest value. It could be the time since the latest event, the latest event type, or the latest magnitude (e.g. the monetary amount). 

Recent data is often the most relevant. When optimizing the timing of when to send the next marketing campaign content to a sales prospect, it is the time since the last contact that determines whether the person feels spammed or forgotten. In healthcare, it is the patient’s latest temperature that determines their prognosis. When estimating credit risk, the borrower’s current income and loan balance are more relevant than historical values.

Here are a few tips to get the most out of recency features:

• Ignore old events. Set a maximum time period to search for recent events, and ignore all events before that period.
• Experiment with duration, type, and magnitude metrics to discover the most effective recency signals
• Use domain knowledge to filter the data to list the latest bellwether events

Experimenting with these features and serving them in production can take a lot of time and energy. We’ve built an open-source feature engineering library that makes it easy to create these recency signal types and more! Click here to sign up for a free download, with worked examples in Python: https://docs.featurebyte.com/