Glossary
Terms, in plain English
The notes on this site don't simplify the method — they explain the terms instead. Every piece of jargon used in the research is defined here, in the language I'd use out loud. Nothing here assumes you've taken the course.
-
ACM (Adrian–Crump–Moench)
The NY Fed's model that splits a yield into expectations and term premium.
A term-structure model from the Federal Reserve Bank of New York that decomposes a Treasury yield into an expected-rate (risk-neutral) component and a term premium. Widely used because the NY Fed publishes it free. Worth remembering that it is an estimate, not a measurement — its output carries its own model error.
See also: Term premium
-
ADF test (Augmented Dickey–Fuller)
The standard test for whether a series wanders or reverts to a mean.
A hypothesis test for non-stationarity. A low p-value says the series reverts to a stable level; a high one says you can't rule out that it wanders. It's the routine first check before regressing two time series on each other — and skipping it is how spurious regressions get published.
See also: Stationary / non-stationary · Spurious regression
-
Basis point (bp)
One hundredth of a percentage point.
One hundredth of a percentage point: 100 bp = 1%. Rates people use it because saying "rates rose 0.25%" is ambiguous — a quarter of a percentage point, or a quarter of one percent of the current rate? "25 bp" can only mean one thing.
-
Chow test
Tests whether a relationship broke at a specific date.
A test for whether a regression's coefficients are the same before and after a chosen break point. Useful for asking whether a relationship survived 2008, or the zero-lower-bound years, or 2020 — and if the answer is no, a full-sample estimate is an average of regimes rather than a description of any of them.
-
Cointegration
The test for whether two wandering series are genuinely tied together long-run.
Two series can each wander indefinitely and still be bound together, so the gap between them stays stable — like a dog and its owner on a lead. That's cointegration, and it's what separates a real long-run relationship from a spurious one. If two series are cointegrated, a regression of one on the other means something; if they aren't, it usually doesn't.
See also: Spurious regression · Stationary / non-stationary
-
Constant maturity (GS2, GS10)
A yield series held at a fixed maturity, so it's comparable over time.
An individual bond's remaining life shrinks every day, so tracking one bond doesn't give a clean time series. A constant-maturity series interpolates the curve to a fixed point — 2 years, 10 years — so today's 10-year yield is comparable to last year's. FRED publishes these as GS2, GS10, and so on.
-
Durbin–Watson statistic
A quick read on serial correlation. ~2 is healthy; near 0 is a red flag.
A one-number check for serial correlation in a model's residuals. Around 2 means the errors look independent; near 0 means each error closely tracks the last one, and the reported precision of the model can't be believed as printed.
See also: Serial correlation (autocorrelation) · HAC / Newey–West standard errors
-
Expectations hypothesis
A long rate ≈ the average short rate people expect, plus a premium.
The idea that a longer-term interest rate is roughly the average of the short-term rates the market expects over that period, plus a term premium for bearing the risk. It's why a 2-year Treasury is often read as the market's forecast of the average policy rate over two years — and why the Fed's grip is tight at the short end and loose at the long end.
See also: Term premium · Federal funds rate
-
Federal funds rate
The overnight rate the Fed targets — its main policy lever.
The interest rate banks charge each other for overnight loans of reserves. The Fed sets a target range for it, and it is the lever nearly everything people mean by "the Fed raised rates" refers to. It is an overnight rate — which is precisely why its grip on a 30-year loan is a question rather than an assumption.
See also: Expectations hypothesis
-
Greenspan's conundrum
2004–06: the Fed hiked 396 bp and long rates barely moved.
Between June 2004 and June 2006 the FOMC raised the federal funds rate by 396 basis points and long-term rates barely responded — the 30-year mortgage moved 39 bp. Greenspan called it a conundrum in 2005. It is the cleanest historical demonstration that the short end and the long end are not the same lever.
See also: Federal funds rate · Term premium
-
HAC / Newey–West standard errors
A correction that fixes overstated precision without changing the estimate.
Heteroskedasticity- and autocorrelation-consistent standard errors — a correction for when errors are neither independent nor equally sized. It leaves the estimate alone and widens the uncertainty around it, which is often enough to cut a headline t-statistic in half. Named for Whitney Newey and Kenneth West.
See also: Serial correlation (autocorrelation) · Heteroskedasticity · t-statistic
-
Heteroskedasticity
Errors that are bigger in some periods than others.
When the size of a model's errors varies with conditions — small in calm periods, large in volatile ones. Common in anything financial. Like serial correlation, it doesn't bias the estimate but it breaks the standard errors, so significance tests can't be taken at face value.
See also: HAC / Newey–West standard errors
-
MBS spread
The gap between mortgage rates and Treasuries — its own moving part.
Mortgages are bundled into mortgage-backed securities, which yield more than Treasuries. That gap compensates for prepayment risk (borrowers refinance when it suits them, not you), plus bank funding conditions and origination economics. It moves for reasons of its own, which is part of why a mortgage rate isn't the policy rate plus a constant.
See also: Prepayment risk
-
No-change benchmark (random walk)
Predict that tomorrow equals today. Embarrassingly hard to beat.
The most naive forecast available: next period's value equals this period's. It sounds trivial and is notoriously difficult to beat for financial series. Any forecasting model that can't outperform it has not demonstrated anything, however good its in-sample fit looks.
See also: Out-of-sample R²
-
Out-of-sample R²
Does the model beat a naive benchmark on data it never saw? Can be negative.
R² computed on data the model wasn't fitted to, measured against a benchmark. Unlike ordinary R² it can be negative — meaning you'd have done better with the naive guess. It's the honest test, because fitting history well is easy and predicting is not.
See also: No-change benchmark (random walk) · R² (R-squared)
-
Prepayment risk
Borrowers refinance when rates fall — so you get your money back at the worst time.
A US mortgage borrower can repay early, and does so exactly when rates drop and reinvesting is least attractive. That optionality is the borrower's and the lender bears it, so mortgage yields carry compensation for it — a component that has nothing to do with the Fed's target rate.
See also: MBS spread
-
R² (R-squared)
Share of a variable's movement the model accounts for. Easy to overread.
The fraction of the variation in the thing you're explaining that your model accounts for, from 0 to 1. A high R² feels like proof and often isn't: two series that both trend upward over 25 years will produce a high R² whether or not they have anything to do with each other. It measures fit, not truth.
See also: Spurious regression · Out-of-sample R²
-
Serial correlation (autocorrelation)
Today's error resembles yesterday's — so you have less information than you think.
When a model's errors are correlated across time rather than independent. It doesn't bias the estimate itself, but it means your 300 monthly observations carry far less independent information than 300 truly separate ones — so the standard errors come out too small and everything looks more certain than it is.
See also: Durbin–Watson statistic · HAC / Newey–West standard errors
-
Spurious regression
Two drifting series look related because both drift, not because they're linked.
Granger and Newbold's 1974 result: regress one wandering series on another and you reliably get a high R² and a huge t-statistic even when the two are entirely unrelated. The regression is measuring shared trend, not a relationship. It is the single most common way a time-series result can be technically correct and completely meaningless.
See also: Cointegration · Stationary / non-stationary · R² (R-squared)
-
Stationary / non-stationary
Whether a series returns to a stable average or wanders indefinitely.
A stationary series has a stable mean and variance — knock it away and it comes back. A non-stationary one wanders with no particular level to return to. The distinction matters because most standard regression results assume stationarity, and applying them to wandering series produces confident-looking nonsense.
See also: ADF test (Augmented Dickey–Fuller) · Spurious regression
-
t-statistic
How many standard errors an estimate sits from zero. Above ~2 is 'significant'.
An estimate divided by its standard error — roughly, how confident you can be that a relationship isn't zero. Above about 2 is conventionally "significant." It depends entirely on the standard error being right, so a t-statistic computed on autocorrelated data can be wildly overstated while looking authoritative.
See also: HAC / Newey–West standard errors · Serial correlation (autocorrelation)
-
Zero lower bound (ZLB)
When the policy rate is pinned near zero and can't fall further.
The period when a central bank's policy rate sits at or near zero and conventional cuts are unavailable — in the US, roughly 2009–2015 and 2020–2021. It distorts any statistical relationship involving the policy rate, because a variable that cannot move cannot explain anything that does.
See also: Federal funds rate