Annualized return

What it is

Annualized return is a return converted into a yearly rate, so that track records of different lengths can be compared.

The idea is sound. The execution is where most of the inflation happens, because “annualized” is not one calculation — it is at least three, and they disagree.

Three calculations, three answers

Method	How it works	When it flatters
Simple	total return ÷ number of years	Short records
Compound (CAGR)	(1 + total)^(1/years) − 1	Long records, steady gains
Partial-period	scales a fraction of a year up to twelve months	Very short records

The third is where the worst numbers come from. A strategy up 8% in one month annualises to roughly 150% by simple scaling. Nothing about it is wrong arithmetically; everything about it is wrong as a description of the strategy.

Why it matters for AI quant

A model will produce whichever annualisation you ask for, and will not warn you that a 40-day sample annualises into a fantasy. The number looks like a property of the strategy; it is a property of the sample length.

This is the same trap as backtest window selection, one layer up: the period you chose decides the number you report.

How to check

Do not accept an annualized figure on its own. Ask for three things:

1. The equity curve — not a return number
2. The exact date range — start and end
3. Whether costs are included — transaction costs and slippage turn many double-digit annualised returns negative

Then compute the number yourself from the curve. If your calculation and theirs disagree, the difference is the answer.