The first time I heard the term ‘recency’ was during an appraisal at work. My manager had convinced me that they weren’t applying ‘recency bias’ during a performance meeting. In this particular case, I was going through a challenging period and it had become the sole focus instead of the relatively good performance I’d had previously. As it happens, it didn’t affect the overall outcome but the term ‘recency’ stuck with me and I’ve noticed it more ever since. Recency bias means ‘to favor the most recent occurrence or information’, even if it isn’t indicative of the big picture.
What is Recency Bias?
The human tendency of ‘recency’ means to take into account the most recent memory instead of older, sometimes more accurate data. It turns out to be a common trait. We tend to forget easily and maybe through a desire to short cut a decision, we only consider what just happened. It could be that our memories become less potent over time and so we turn to the newer and fresher experiences. We can make some logical assumptions and look at the science of the brain to propose why it happens. First, a bit more about where we see it occur and ask if it seems to be getting worse.
We see it at Election Time…
Election time naturally brings about a drive for new, insightful data that a lobbying party hopes might influence a voter. Every day there is a breaking story urging voters to swing one way or another. In the 2016 US election, the late breaking news of Hilary Clinton’s e-mails that dominated the media reporting just 11 days before the election was bound to have had a strong influence. We all now what happened and this time around you can see similar tactics being attempted. Last minute information coming to light, just at the time that undecided voters maybe searching for signals on which way to cast their vote. It’s unclear at the time of writing this, what (or if) a big hitting recent headline will have an impact on the 2020 US election.
…And we see it in Social Media and the Stock Market
Recency bias shows up strongly in the news, with Facebook and Twitter feeds being all about the ‘timeline’ and ‘trends’ i.e. what happened in the recent past. It can make a big difference to the way you view a person or an event. Every time a natural disaster strikes, our sympathies are strongest just after but over time, our concern can fade fast. We’ll reserve our empathy for the event that just happened even if the need for help continues long after. Another place it shows up strongly is the stock market. A place where decision making can make or lose a huge amount of money in seconds. Faced with a time critical decision on whether to buy or sell a stock, it turns out that recency bias is common place. There is only also much data that is to hand in the world of trading and the most recent data is the one that skews the decision.
We’ve become recency junkies these days. The rate at which we can get new information is accelerating. Our appetites for it is only getting larger. News from a week ago can feel like a lifetime ago.
Why does it happen?
We know our memory is selective. We tend to forget in unpredictable ways but the freshest memories in our mind take precedence. In Kahneman’s ‘Thinking Fast and Slow’ (essential reading), the two main systems of thinking are identified. System 1 that involves fast reactions and System 2 for more reasoned deliberation. Our System 1 thinking kicks in and given a restricted amount of time and recent events top of mind, we automatically over index on that. Having the luxury of a long period of time to reflect and dig deeper into the past, may lead to a different conclusion.
It makes sense that we turn to the most recent information to make decisions. There is good reason to apply recency bias in some circumstances. For example, when making a decision about whether to travel during Covid, we want to know what the current situation is, not average numbers over the last few months. In this case, the most recent knowledge is the most important because it is an indicator of a trend. It is a warning to the future and the deeper past doesn’t tell us the critical information we need.
Computers apply recency too
Machines can produce amazing predictions given the right data. Most of the AI that we know is built using machine learning whose main purpose is just that. Figuring out what may happen based on the past. However, they have an advantage that they are capable of taking in vastly bigger data sets than we as humans can handle. They can assess every single memory if required and calculate a best decision in extraordinary fast timing. But it’s different to what we do and there are plenty of incidences where having all this data to hand still leads to an incorrect (and sometimes humorous) result. The common sense element of thinking that we humans usually have, could be the missing piece. It’s within the common sense trait that recency bias could play a major role.
Perhaps it be considered more within machine learning. By including a bias on new information, could it help to produce more human like, common sense responses? Being able to focus on a smaller, recent data set could streamline processing. Less data needs to be analyzed and it could save massive processing power without compromising accuracy. In fact by making it closer to the human way of thinking, it could lead to more human like predictions. The challenge of course it knowing which recent data to select. Why and how humans choose memories to reference remains a mystery of the brain.
Could it help us design better AI?
We have some clues to the way the brain selects certain memories and over indexes on recent events, but it still remains largely subjective. We all make decisions in a different way and with sometimes with conflicting perspectives. Which recent data and how much we factor it in, is unique to us. As we learn more about the brain and try to apply this to build better ‘thinking machines’ we may need to revisit the recency bias factor and see if we can use this to our advantage. By bringing the human trait of recency bias and combining it with machine learning’s ability to analyze big data, we may find a path that utilizes the best of both.