Quantitative finance: Looking for a whisper in a stadium

Quantitative finance replaces intuition with data and calculation. It looks for patterns in mountains of numbers.

Its first enemy is noise. Prices move mostly at random. The signal, the small predictable part, is tiny.

For the example: a share moves about one per cent a day. A good signal adds perhaps two hundredths of a per cent. One in fifty.

It is like listening for a whisper in the middle of a stadium. You need thousands of observations to hear it.

Second tool: the backtest. You test a rule on the past, to see what it would have earned.

The problem: everything works on the past, if you search long enough. The past cannot defend itself.

Test twenty worthless rules. With the usual five per cent statistical threshold, one, on average, will seem to work.

And you have about a sixty-four per cent chance of finding at least one. Then you only need to show that one.

That is overfitting: finding a pattern in a game of dice. The rule sticks to the past, and fails on the future.

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

The more settings a rule has, the closer it sticks to the past. And the less chance it has of working afterwards.

Researchers themselves have fallen for it. A study from two thousand sixteen lists more than three hundred factors supposed to predict returns.

Another shows that after publication, the return of these anomalies falls on average by more than half.

Good quants protect themselves. They keep data aside, never touched, for the final test.

They demand an economic explanation: why would this effect exist? And who would accept losing on the other side?

They count real fees, and distrust a curve that is too pretty. A strategy that has never lost has mostly been fitted too well.

To find a signal others do not have, some look for new data. It is called alternative data.

Satellite images to count cars in supermarket car parks. Or to measure oil tanks, seen from the sky.

Anonymised till receipts, to estimate sales before the official announcement. A website's traffic, to sense a trend.

But an edge is quickly shared. As soon as a dataset is sold to everyone, its signal wears out.

Remember: a result that looks too good on the past should worry you, not reassure you.

Before believing a backtest, always ask how many rules were tried to find this one.