---
title: "Confirmation bias"
description: "Confirmation bias is reading evidence in favor of a conclusion you already hold — accepting the runs that agree and explaining away the ones that don't."
canonical_url: "https://fincanva.com/glossary/confirmation-bias"
last_updated: "2026-10-06"
md_url: "https://fincanva.com/glossary/confirmation-bias.md"
---

# Confirmation bias

Confirmation bias is the tendency to read evidence in favor of a conclusion you already hold — accepting the results that agree with your idea at face value while finding reasons to discount the ones that disagree. It is an error of **interpretation**, so it survives every fix applied to the data and the rules.

**Also seen as:** Confirmatory bias, my-side bias

It is the most human of the backtesting biases, because it needs no bad data, no broken test, and no intent to mislead: the same person can run a technically flawless backtest and still come away believing something the run did not say.

## How does Fincanva handle it?

A [backtest](/glossary/backtest) applies your rules mechanically across the whole history regardless of what you hoped would happen, and reports the result whole — the return and headline figures like [CAGR](/glossary/cagr), the worst peak-to-trough [drawdown](/glossary/max-drawdown), the losing months, and the [benchmark](/glossary/benchmark) run over the identical period beside it. That gives the rules a standing chance to disagree with you, and puts the disagreeing evidence on the same screen as the agreeing evidence rather than one search away.

A backtest's Capital Growth page opens on four tiles side by side: Final value, Vs benchmark, Max drawdown and CAGR — the gap to the benchmark and the deepest fall on the same row as the gain.

It cannot make you read it. Nothing in a backtest stops you from dismissing an inconvenient run, and no tool can supply the intent to be proved wrong.

## How does confirmation bias show up when you read a backtest?

It rarely feels like bias from the inside. It feels like judgment. The recognizable patterns are:

- **Stopping when it agrees.** The first run that supports the idea ends the investigation; a run that contradicted it would have prompted three more.
- **Asymmetric scrutiny.** A confirming result is accepted as-is. A disagreeing one gets audited — wrong universe, wrong period, "something's off with the data" — and the audit stops as soon as a reason is found.
- **Reasons produced after the result.** The objection to a run is invented once its number is known. If a disagreeing run had come out well, the same objection would never have been raised.
- **Reading only the flattering metric.** The return is read and the worst [drawdown](/glossary/max-drawdown) is skipped, or the [strategy](/glossary/strategy)'s own figure is read without the [benchmark](/glossary/benchmark) beside it.
- **Remembering the run that agreed.** Weeks later the memory is "it worked" — the version of the test that agreed, not the version that didn't.

A useful check is a single question asked *before* a run finishes: what result would make me abandon this idea? An idea with no such result is not being tested.

## How do the run you kept and the run you explained away compare?

You believe a twelve-month momentum rule works, and you test it over 2015–2024.

- **Run 1** applies the rule to twenty large, familiar companies. It returns **9.1% a year**. This matches what you expected, so you save it.
- **Run 2** applies the identical rule to a broad four-hundred-name [universe](/glossary/universe). It returns **3.4% a year**. You conclude the wider universe has too many low-quality names in it, and you set the run aside.

Now put the [benchmark](/glossary/benchmark) next to both. Over the same period it returned **7.8% a year**, so run 1's [excess return](/glossary/excess-return) is +1.3pp and run 2's is −4.4pp. Two things follow. First, even the run you kept beat its benchmark by a much thinner margin than "9.1%" suggested on its own. Second, and more important: the reason you gave for discarding run 2 was produced *after* you saw its number. Had run 2 returned 12%, the four-hundred-name universe would not have been "too junky" — it would have been "a broader, fairer test". That asymmetry, not the numbers, is the bias.

## What separates confirmation bias from cherry-picking bias?

The difference is the audience. [Cherry-picking bias](/glossary/cherry-picking-bias) is selective reporting **outwards** — you have the full result and you show someone else a favorable part of it. Confirmation bias is the same selection turned **inwards** — you show it to yourself, usually without noticing, and there is no moment of deciding to omit anything.

The two feed each other. Confirmation bias decides which run you believe; cherry-picking decides which run you present. And a disagreeing result blamed on "bad data" without anyone checking the data is confirmation bias borrowing the language of [data-quality bias](/glossary/data-quality-bias) — as is switching costs off because the net curve "looks wrong" (see [cost-ignoring bias](/glossary/cost-ignoring-bias)).

Fincanva is for education and illustration only. It is not personalised financial advice, and past or simulated results do not predict future ones. [Read the Terms Addendum](https://fincanva.com/terms/addendum#section-3)