What is an algorithm?

BarcelonaThe term algorithm is used much more now than when the 9th century Persian mathematician from whom it derives, Muhammad ibn Musa al-Khwarizmi, devised methods for solving first and second degree equations. The reasons are clear: now there is more data and more computing power. But there have always been algorithms, which are nothing more than a set of instructions aimed at executing a task. The recipe for a chocolate cake or a Valencian paella are algorithms. The way to light a fire with two sticks is an algorithm. The hunting technique used by some groups of killer whales, when they expel air while spinning under their prey to catch them in a cylinder of bubbles, is an algorithm, in this case generated and transmitted in a pre-linguistic way. So are the nuptial dances of the bird of paradise or lullabies.

Today there are algorithms everywhere. When we press the button for the floor we want to go to in an elevator, an algorithm is executed: the elevator uses previous information to move up or down and to stop at the right time. As you know, almost everything we do on the Internet is mediated by algorithms that can become very sophisticated. However, all algorithms must meet a number of conditions, the most important of which is that the instructions can be carried out in a finite number of steps. Otherwise, the algorithm is useless. If the time it takes to make a chocolate cake exceeds the age of the universe, perhaps we'd better choose another recipe.

Today's complex algorithms, which work thanks to large amounts of data, are built from models, which are nothing more than abstractions and simplifications of the functioning of a real system. There are simple ones, which work on the basis of data that are easy to obtain, and there are complicated ones, and often incomplete ones, which work on the basis of the data they can obtain. The case that Michael Lewis explains in the book Moneyball is one of the first: using historical baseball player data to optimize a team's performance. The data exists and is relevant to predict the performance of a player and, incidentally, the whole team.

However, when you want to predict the ability to repay a loan, there is a lot of relevant data that does not exist. The models are then more approximate, because they use data from similar situations that have occurred in the past. To answer the question of whether the store the applicant wants to open with the loan money will do well, it would be ideal to have information from many of the same stores that the same person would have opened so far in similar positions, as with baseball players. But what if this is the first store he or she wants to open?

The models on which algorithms that work with large amounts of data are based can have holes that lead to uncertainty in the predictions. And uncertainty means that the probability of error when making a specific prediction is higher. It is therefore especially important to know how algorithms are designed and what data they are fed with when refining them.