Glossary
Glossary
The words this site leans on, written for somebody who is smart and does not work in finance. Each entry answers in one line, then explains, then says why the word would change what you do. A term is here because a page elsewhere rests on it.
Block bootstrap
Resampling in chunks, so the resamples keep the clustering the real data has.
Ordinary resampling shuffles observations one at a time, which destroys the way market returns cluster and produces intervals that are far too tight. A block bootstrap draws contiguous stretches instead. When several correlated series are pooled, the block indices have to be drawn once and applied to all of them at the same time.
Certainty equivalent
The guaranteed return you would swap a risky one for.
A risky portfolio and a certain return are equivalent to you when you would genuinely accept either. The gap between a portfolio's average return and its certainty equivalent is the price you put on its risk, and it depends on a risk preference somebody has to declare. Two experiments here declared a risk-aversion setting of 3 for their own comparison.
Contango and the funding basis
The built-in cost of holding exposure through futures rather than owning the asset.
A futures contract embeds a financing rate, and that rate has been measured above cash: 58.70 basis points on five-year Treasury note futures over 1991–2018, positive in all 28 years. Equity futures are similar and more variable, and their sign is not even constant. It is a stable cost rather than a crisis artefact.
Deflated Sharpe ratio
A Sharpe ratio marked down for how many strategies were tried before this one.
Search enough variations and one will look good by chance. Deflation adjusts a Sharpe ratio for the number of trials, the length of the sample and the non-normality of the returns. The number of trials it needs is the number of distinct specifications searched, which is why the experiment ledger records every attempt including the failures.
Drawdown, and time under water
The worst peak-to-trough fall, and how long it took to get back.
The US total market returned 10.80% a year over 1963-07 to 2025-12 and fell 50.3% along the way, spending 72 months below its previous peak. Drawdown deepens mechanically with sample length, so two drawdowns from windows of different lengths cannot be compared directly.
Effective sample size
How many independent looks a set of correlated series is actually worth.
Three regions of one factor are not three independent tests, because they share global risk factors, construction and accounting definitions. The effective count is measured from the realised sample rather than assumed: three regions of value were worth 1.49, and three of momentum only 1.33. In the tail it is worse still, because that is when they move together.
Error bars widened for clustered returns
Error bars that still hold when quiet and stormy months arrive in runs rather than at random.
Ordinary error bars assume each month is independent of the last and equally noisy, and monthly returns are neither: calm stretches and violent stretches come in runs. The standard repair, named after Newey and West, widens the bars to allow for both, typically by 7% to 18% on the series here.
Factor exposure
How much of a factor's behaviour a fund actually delivers.
Regress a fund's returns on a factor and the coefficient is its exposure to that factor. An exposure of 0.41 to value means the fund moves 0.41 units for each unit the value factor moves. It measures manufacturing rather than returns: a fund can deliver its exposure perfectly and still be a poor thing to own.
Foreign tax credit
Credit for tax a foreign government already withheld on your foreign dividends.
A US fund pays foreign withholding and may elect to pass it through, after which you credit it against your US tax. Inside an IRA or a Roth there is no US tax to credit against, so the withholding is paid and permanently lost: 15.78 bp a year on a developed-markets holding and 20.00 on emerging. Below $300 of creditable tax ($600 joint) you claim it without Form 1116 or its limitation.
Geometric vs arithmetic return
What you actually compounded, against the simple average of the yearly numbers.
Lose 50% then gain 50% and the arithmetic average is zero while you are down 25%. The geometric return is the one your wealth actually followed. The gap widens with volatility, which is why a volatile strategy's advertised average return overstates what a holder received.
Kelly, or growth-optimal sizing
The bet size that maximises long-run compound growth, given a known edge.
Betting the log-optimal fraction beats any other strategy asymptotically, which is a theorem rather than a preference. It is also unusable without an edge and its uncertainty, and the estimate has to be shrunk by its own standard error first. Omitting that shrinkage is the most likely catastrophic sizing error in a system like this.
Minimum detectable effect
The smallest true effect a window could have found, if one were there.
It answers a different question from a p-value. A p-value asks whether a result could be zero; the minimum detectable effect asks whether the window could have found something worth having. When a measured premium is smaller than its own detection threshold, an interval that excludes zero is not evidence the window can carry.
Model-misfit pedestal
What a factor model charges a fund that is definitionally the market.
A total-market fund is the market portfolio, so under a correctly specified model its alpha should be about minus its three-basis-point fee. Under the standard six-factor model over 2020–2025 it came out at −0.55 percentage points a year, on error bars widened for clustered returns, with a t of −3.41. Every fund priced by the same model over the same window carries that offset.
Multiple testing and the Holm correction
Adjusting for the fact that testing twenty things guarantees one looks significant.
Test twenty independent true nulls at 5% and you expect one false positive. Benjamini-Hochberg controls the share of discoveries that are false and assumes the tests are independent; Holm-Bonferroni controls the chance of any false positive and stays valid under arbitrary dependence. The tests in this record are heavily dependent — nested specifications, shared factors, shared months — so Holm is the defensible one and the BH count is an optimistic bound.
Purged walk-forward
Testing on data the rule has never seen, with the overlapping bits cut out.
Fit on an early window, test on the next one, roll forward, and never let information from the test period leak backwards. Purging removes observations whose outcomes overlap the boundary, and embargoing leaves a gap after it. A final holdout stays untouched, because looking at it once turns it into training data.
Qualified dividend
A dividend taxed at long-term capital-gain rates instead of ordinary rates.
The US schedule offers 0%, 15%, 18.8% and 23.8%. A dividend qualifies only if the stock was held more than 60 days inside the 121-day window around the ex-dividend date, so a fund only 70% qualified on a 2% yield loses about 10.2 bp a year to the difference.
Securities lending
Funds lend their holdings to short sellers and pass most of the fee back to you.
It is small and it is real: about 1.01 bp a year for a US total-market fund, 0.07 for an S&P 500 fund, and 9 to 10 for a core emerging-markets one. The premium is international and emerging lending demand rather than a size effect — US small-cap earns the same as large-cap developed international. Unit investment trusts such as SPY and QQQ cannot lend at all.
Sequence risk
The order returns arrive in, which matters only when money is going in or out.
Without external cash flows, permuting the order of returns leaves terminal wealth unchanged — it is a multiplication and multiplication commutes. Contributions and withdrawals break that identity, because a bad early year is applied to a different amount of money than a bad late one.
Specific identification
Telling your broker exactly which shares to sell, rather than letting it pick.
Regulation requires only that you specify the particular stock at the time of sale, and it accepts a standing instruction. The default without it is first-in-first-out, which realises the most gain available. Switching is free, needs no form, and is not a method of accounting.
Step-up in basis
Death resets an asset's cost basis to its market value, forgiving the gain outright.
An unrealised gain is an interest-free loan from the government whose principal compounds with the position. Under §1014 that loan is forgiven at death, and a gift of appreciated long-term property to a public charity does the same while you are alive. Deferral is worth 84 bp a year at thirty years and the step-up a further 78, summing to a horizon-free 162.
Tracking error
How far your returns wander from the thing you are measuring yourself against.
It is the standard deviation of the difference between your return and the benchmark's, per year. A small edge against a small tracking error is near-certain quickly; the same edge against a large one may never be provable. Time to any confidence level scales with the square of tracking error divided by edge.