Past performance is a small sample and a noisy one

The warning is not a formality

Every promotional document carries a line saying past performance does not predict future results, and almost everyone reads it as regulatory throat-clearing before the interesting numbers. It is not. It is a compressed statement about how little can be inferred from a short record of a noisy process, and the reasoning behind it is worth having.

The core problem is that returns vary enormously from period to period for reasons unrelated to skill or structure. A record of several years therefore contains a mixture of whatever is durable about an approach and a large quantity of noise, and separating the two requires far more data than a few years provides.

This is not scepticism about whether the numbers are accurate. They generally are. It is scepticism about what they support as a conclusion.

Chance produces convincing records on its own

Consider a field with a large number of participants all making decisions with uncertain outcomes. Purely by the arithmetic of chance, some will produce a string of good years in succession. Their records will look remarkable, they will be the ones written about, and nothing about the process that generated them was different from the process that generated the unremarkable ones.

This is not a claim that skill does not exist. It is a claim that a good record is consistent with both explanations, and that distinguishing between them from the record alone is much harder than it appears. The stronger the record, the more attention it attracts, which is precisely why the sample being examined is not a random one.

The effect is compounded by which records survive to be examined. Arrangements that performed badly are frequently closed or merged, so the set available for comparison is systematically better than the set that originally existed. Any average calculated across survivors is flattering by construction.

What the number describes is a period, not a method

A return over a stretch of years is inseparable from the conditions of those years. An approach that suits a particular environment will look excellent while that environment persists and unremarkable afterwards, without anything about the approach having changed.

This means a record is partly a description of the past decade rather than of the thing being measured. Two arrangements with different records may have differed mainly in which conditions happened to occur, and the ranking between them can reverse when conditions do.

It also means that the periods being compared have to match exactly for a comparison to mean anything. A record starting immediately after a large fall looks quite different from one starting immediately before it, and start dates in promotional material are not chosen at random.

What does carry information

Some things about an arrangement are considerably more predictable than its returns. What it costs is knowable in advance and applies every year. What it actually holds is knowable. How concentrated it is, and therefore how much damage a single failure could do, is knowable. So is whether the approach is one you could still describe accurately in five years.

These are less exciting than a return figure and they have the substantial advantage of being facts rather than samples. The general observation that costs are among the more reliable predictors of long-run outcomes follows directly from the fact that they are certain while returns are not.

A long record is more informative than a short one, and consistency of approach is more informative than consistency of result. But even a long record supports weaker conclusions than most readers take from it.

How to read a record without over-reading it

A performance figure is worth looking at, and it belongs in the same category as any other single data point: suggestive, not conclusive. What it can do is prompt a question about why, which occasionally has an answer that is more informative than the number.

What it cannot do is settle a choice on its own, and the confidence with which it is usually presented is inversely related to how much it supports. That mismatch between presentation and evidential weight is the reason the warning exists at all.

None of this is guidance about what to select. Choices of this kind depend on the horizon, the purpose, the tax and regulatory environment where you live and the rest of the household position, and anything of consequence belongs with a regulated adviser rather than with a table of past returns.