Mean vs Median Retirement Savings: Why Headlines Feel Impossible

Updated August 2, 2026. Quick answer: when an article says the average American aged 55–64 has $306,404 saved for retirement, it is not lying. It is using the mean. The median for that same group is $16,600 — about 18 times smaller. Half of those households have less than that. Both numbers come from the same survey; only one describes a typical household.

The gap, cohort by cohort

Why the two diverge so violently

A mean adds everything up and divides by the number of households. A median lines every household up and takes the one in the middle. When a distribution is roughly symmetric — adult heights, say — the two land close together and either will do.

Retirement wealth is not remotely symmetric. It is bounded below at zero, and unbounded above. A single household with $20 million lifts the mean of a thousand households by $20,000 while moving the median not at all. So the mean answers “how much retirement money exists per household”, which is a real question about an economy. It does not answer “how much does a household like mine have”, which is the question the reader is asking.

The clearest evidence is the zero column. Among 55–64s, 43.0% have no retirement account at all. A mean of $306,404 in a group where more than two in five have nothing is arithmetically correct and descriptively useless.

What to do with a headline

Three questions settle almost any published savings figure.

Is it a mean or a median? If the article does not say, assume mean, because it is the larger number and larger numbers travel further.

Is it conditional on having an account? Excluding people with zero raises the figure enormously. Across all families, the median holding is $4,260 counting everyone, and $86,600 counting only those who have an account. Same survey, same year, and roughly twentyfold apart — purely a definitional choice about who is in the denominator.

Whose balances, and from when? Household or individual, retirement accounts only or all wealth, and which survey year.

Where you actually stand

Rank yourself against your own age cohort with the full distribution rather than a single number. And then the question that matters more than the ranking: whether your own number works for your own life.

How these numbers were produced

They are our own computation from the Federal Reserve’s 2022 Survey of Consumer Finances summary extract — 4,595 households — not a figure copied from another article. The method was written down before the computation ran, and the script is deterministic.

The part most summaries get wrong. The SCF is published as five multiply-imputed replicates of every household. Treating the file as one dataset of five times as many households inflates the sample and understates uncertainty. We compute every statistic separately within each of the five, then average — and say so, because a summary that does not mention implicates usually has not handled them.

We checked ourselves against the Board. Against the Federal Reserve Bulletin’s own published table for the same wave: our share of families holding a retirement account matches exactly (54.3% against their 54.3%), and our conditional mean matches to 0.027% ($334,090 against $334,000).

Our conditional median comes to $86,600 against their $86,900 — a 0.345% difference, and we can say exactly where it comes from. It is the weighted-median convention, not the data: two defensible conventions bracket their figure (averaging the five implicates gives $86,600, pooling them gives $87,000), and the five implicate medians themselves span $83,000 to $90,000. A median on this data is genuinely sensitive to how you compute it, which is worth knowing about every published figure of this kind, including ours.

Source: Board of Governors of the Federal Reserve System, Survey of Consumer Finances 2022, Summary Extract Public Data. Variable: RETQLIQ (quasi-liquid retirement assets), whose composition we verified arithmetically against its four components on all 22,975 records. Figures cover all families in each cohort, including those with nothing. Archive SHA-256 recorded in our staged dataset.