Recency Weighting

Anyone wanting to say something about the future from past data must decide how much each individual data point counts. The simplest answer is: all equally. Recency weighting chooses a different answer. It gives fresh observations more weight than old ones. A sales figure from yesterday then influences the forecast more strongly than one from three years ago. The term “recency” simply means currency or newness.

Why old data can be misleading

The world rarely stands still. Prices rise, fashion tastes shift, user habits change after every app update. Experts call this “drift”: the rules by which the data is generated shift over time. A model that treats all years equally then learns a mixture of the old and the new world. The result doesn’t quite fit either period.

An example from retail makes this tangible. A chain wants to estimate demand for winter jackets. If figures from five years flow in with equal weight, the average dominates. A warm winter last year and new competition around the corner get lost in it. With recency weighting, exactly these most recent signals carry more weight.

But there is also the opposite direction, and it is often underestimated. Anyone who weights the most recent data too heavily confuses chance with trend. A single crazy week is then taken for the new normal. In behavioral research, this cognitive error in humans is called recency bias. In software, it doesn’t arise from psychology, but from a poorly chosen setting.

Half-life and time windows

Technically, every data point gets a number as its weight. Current points get about 1.0, older ones correspondingly less. Exponential decay is very common: the weight halves after a fixed period. This period is called the half-life, similar to physics. With a half-life of 30 days, a data point from 60 days ago only counts a quarter as much.

The simpler variant is the sliding time window. Here you only use data from the last 90 days and discard everything older. This is easy to implement, but acts like a hard cut. A data point from day 89 counts fully, the one from day 91 not at all. The exponential variant proceeds more gently and is therefore chosen more often.

The decisive question is always how fast the weight should decline. If it falls too slowly, the system reacts sluggishly to real changes. If it falls too quickly, the forecast jitters with every outlier. That’s why several half-lives are tested on past data to check which would have delivered the best forecasts. In large AI models, the same principle appears when more recent texts occur more frequently in training than very old ones.

From credit checks to the feed

You encounter this principle constantly in financial news, often without the name. Credit risk models weight payment behavior from recent months more heavily than that from eight years ago. Price forecasts work with moving averages that weight more recent prices higher. Fraud detection also relies mainly on fresh patterns, because fraudsters constantly change their methods.

In everyday life, recency weighting is embedded in almost every recommendation system. A music service is more likely to suggest what you listened to this week, not what you liked three years ago. Social networks also sort posts by how new they are. That’s exactly why such feeds sometimes feel short-winded: what is two days old is already considered done.

It’s important to distinguish this from a related misconception. Recency weighting is not a statement about whether data is true. Old data can be completely correct and still reveal little about today. Conversely, currency does not make a figure more reliable. Weighting only decides relevance, not quality.