Machine learning

Machine learning has transformed image recognition and translation. In finance, its successes are more uneven than people think.

There are four fields. Predicting prices, executing orders, detecting fraud, and deciding instead of a human.

First field, price prediction. It is the dream. And it is the most disappointing field.

A photo of a cat stays a photo of a cat. A market changes as soon as its rule is found, because others exploit it at once.

The signal is weak, the history is short, and the past does not repeat. A powerful model mostly learns the noise.

Second field, order execution. Less glamorous, but far more profitable.

Buying one million shares at once pushes the price against you. The order must be cut up, at the right pace.

For the example: on a one hundred million euro order, saving a tenth of a per cent in cost is one hundred thousand euros saved.

Repeated every day across thousands of orders, it is a fortune. And here, the model has millions of examples to learn from.

In the app, this chapter is animated and narrated, card by card. Start free trial.

Download on theApp Store

Third field, fraud detection. Where it really works.

Every card payment is scored in a fraction of a second. Place, amount, time, habits: everything is compared.

Classic trap: say one fraud per thousand payments. A model that always answers "no fraud" is right ninety-nine point nine per cent of the time.

And it is useless. What counts is how many frauds it catches, and how many honest customers it blocks.

For the example: one million payments, one thousand frauds. The model catches nine hundred, and is wrong on zero point five per cent of honest payments.

That makes four thousand nine hundred and ninety-five false alerts. Barely one alert in six is a real fraud.

Fourth field, the black box. A model refuses a loan. Why? Often, even its creators cannot say simply.

Yet in many countries, the law requires that a loan refusal can be explained. And a regulator wants to understand a decision before accepting it.

A model can also reproduce the biases in its data. If it learns from unfair decisions, it repeats them, at high speed.

So banks make a trade-off: a slightly less accurate model that can be explained, rather than a silent oracle.

Remember: the machine excels where examples are many and rules are stable. It disappoints where the future does not look like the past.

So beware of any artificial intelligence that promises to predict the stock market. That is precisely the field where it shines least.