Search the docs⌘K
Glossario
Intermediate
ENIT

Selection bias

UPDATED 2026-07-25

Selection bias is the error of drawing a general conclusion from a sample of instruments or periods that is not representative of the population you mean to describe — typically a small, favorable subset chosen with the benefit of hindsight. The resulting figures are arithmetically correct; they simply describe that sample rather than the strategy. Selection bias is the bias of the sampling step: it enters before any rule is applied, and once the sample is chosen no analysis performed on it can undo the distortion.

Also seen as: Sample selection bias, sampling bias

What makes a sample unrepresentative?

A sample is unrepresentative when membership in it correlates with the outcome being measured, which in backtesting happens in three ways:

  • Hindsight-informed instruments. You already know which names did well over the period, and those are the names in the test. The knowledge that shaped the sample was not available at the start of the period being tested.
  • Hindsight-informed periods. The test covers a stretch of market history that suited the strategy's style, chosen because it suited it. A trend-following rule tested only across a long uninterrupted trend is a sample of one favorable regime. Choosing that stretch and then reading its result as support for the idea is where selection bias shades into confirmation bias.
  • Too small a sample. A handful of instruments or a couple of years cannot distinguish a durable effect from noise, so the result is dominated by whichever few positions happened to dominate.

The clearest symptom is that the conclusion does not survive a change of sample: run the same rules on a wider list of instruments, or on a different span of years, and the effect disappears.

Worked example: hand-picking the assets that happened to work

Consider a strategy tested on five instruments. You know, in 2026, that those five were among the strongest performers of the previous decade, and you chose them for that reason. The backtest reports a high return — but the return is a property of the five names, not of the rules, and the same rules applied to five names chosen without hindsight would have produced something ordinary. The test cannot tell you which of the two it measured, because it only ever saw the favorable sample. Running the identical rules over a broad universe instead of the five separates the two questions: if the effect holds across hundreds of names it is at least a property of the rules, and if it collapses it was a property of the five.

What does Fincanva do about selection bias?

Fincanva lets a strategy draw from a broad universe narrowed by explicit filters rather than a hand-typed list of names, and lets a backtest run from an early simulation start year across many market regimes. The start-date sensitivity view re-runs the same strategy across many entry dates and holding windows, which reveals directly whether a result depended on one favorable starting point.

None of this chooses a sample for you. A universe you narrow down to the instruments you already know worked, or a start year you moved because the earlier years looked bad, carries selection bias whatever the tool does. See the nine biases Fincanva helps you avoid for the product behavior in context.

How is selection bias different from survivorship bias and cherry-picking?

The three biases attach to three different steps, and each one can occur without the others:

BiasWhere it entersWho introduces it
Survivorship biasthe instrument list itself, before any choice is made — the failures are already gonethe data
Selection biaschoosing which instruments or which period to testthe researcher, at test time
Cherry-picking biaschoosing which of the results already produced to reportthe researcher, at reporting time

Read together: survivorship bias means the losers were never on the list; selection bias means you picked the sample that suited you; cherry-picking means you ran the full test and then quoted only the good part of it.

Backtests show what would have happened — not what will. Fincanva provides no financial advice — see Is this financial advice?.

Where this term is used

Generated · 1 pages

The pages that reference this term — so a term page is somewhere you pass through, not somewhere you land and stop.

Fincanva provides no financial advice. Backtests show what would have happened — not what will.

GLOSSARY · 193 TERMS