Many of us use algorithms every day for decision making.
We don’t always trust them, however, and tend to use them less if they have shown themselves to be imperfect in the past.
We tend to judge algorithms by how well they do at meeting a performance goal of some kind, rather than working out whether they will do better than the method we currently use.
This usually results in worse outcomes.
For example, if you drive, the chances are that you use a satellite navigation system often.
Whether it is a standalone system with built in maps or a connected system with traffic feedback like Google maps, how often have you decided to ignore the guidance and decided you know best?
The chances are that you will do better more often by following the guidance.
Algorithms work better in some situations than others.
Broadly, there are three kinds of situations or environments you could face.
In the image above these are categorised into learnable situations, where you can improve through practice, and the predictability of situations – whether you know what could happen next or not.
A zero validity situation is one where you can’t learn through practice and you don’t know what could happen next. A career path for a baby, for example, or the direction of world policy with Trump and Brexit.
A high validity situation is one where you can get better with practice and you can tell what is going to happen next.
Learning to play tennis for example, or learning to drive a car.
You know that a ball is going to arrive in your direction in the near future, or that you will need to drive in a straight line, or around a curve or slow down or speed up.
Between these two extremes is a wide range of low validity situations characterised by uncertainty and unpredictability.
The nobel prize winning economist Daniel Kahneman writes about how algorithms perform best in such low-validity environments.
These cover a wide range of situations including medicine, recruitment, finance, logistics and so on.
In study after study we find that simple rules outperform experts.
For example, a simple six point model outperformed doctors in judging the probability of cancers in a patient.
A stock market index fund that that simply follows the top 500 companies will outperform the vast majority of expert stock pickers.
Using a few measures to score applicants will select candidates who will perform better than those selected by “gut instinct”.
In fact, selecting applicants purely based on the information in CVs can produce a better result than selecting after an interview.
Algorithms don’t have to be complex. They can be based on simple rules based on existing statistics or common sense.
What algorithms do is help cut through the “noise” and focus on a few factors that can make a difference.
When used well, algorithms can help experts make much better decisions by helping them bypass their own cognitive biases.
We should all be using them much more in our work and lives.