The Hodrick-Prescott Filter

GDP data wiggles up and down every quarter, but economists usually want to talk about two separate things: the underlying long-run growth path (the trend), and how far the economy currently sits above or below that path (the cycle — is it a boom or a recession?). Simply drawing a straight line through the data is too rigid for a growth path that itself bends over decades; simply smoothing with a short moving average lets too much short-term noise leak into what should be a clean trend. The Hodrick-Prescott (HP) filter is a specific, mathematically defined way to split a series into these two pieces, with one dial that controls exactly how flexible the trend line is allowed to be.

An analogy: bending a flexible ruler through the data

Imagine physically bending a flexible plastic ruler so it passes as close as possible to a scatter of data points, but the ruler resists being bent too sharply — it wants to stay smooth and straight where it can. Pull it too rigidly straight and it ignores real, gradual shifts in the data; let it bend too freely and it just tracks every bump, ceasing to be a "trend" at all and becoming indistinguishable from the noisy data itself. The HP filter is exactly this flexible ruler, with a stiffness dial: turning the dial up makes the ruler more rigid (a smoother, slower-moving trend); turning it down lets the ruler bend more freely (a trend that hugs the data more closely, leaving less for the cycle to explain).

The mechanics, one symbol at a time

The HP filter splits the observed series $y_t$ into a trend $\tau_t$ and a cyclical component $c_t = y_t - \tau_t$, choosing the trend to minimize:

$$\sum_{t=1}^{T} (y_t - \tau_t)^2 + \lambda \sum_{t=2}^{T-1} \big[(\tau_{t+1} - \tau_t) - (\tau_t - \tau_{t-1})\big]^2.$$

The first sum is the total squared gap between the actual data and the trend — how well the trend fits the data. The second sum penalizes the trend's curvature — how much the trend's slope changes from one period to the next — so a trend that suddenly bends sharply gets penalized. $\lambda$ (lambda) is the single smoothness dial: at $\lambda = 0$ there's no curvature penalty at all, so the trend simply equals the data exactly (every wiggle becomes "trend," leaving zero cycle); as $\lambda \to \infty$, curvature is punished infinitely, forcing the trend toward a single straight line. In plain English: the filter finds the smoothest possible trend line that still stays reasonably close to the actual data, with $\lambda$ setting the exchange rate between "closeness to data" and "smoothness of trend." The commonly cited default is $\lambda = 1600$ for quarterly data, though this specific number is itself a source of ongoing debate.

Worked example 1: two lambdas on the same series

Suppose quarterly GDP index values run $\{100, 102, 99, 105, 104, 108\}$. With a very small $\lambda$ (say 1), the fitted trend tracks the data closely, producing something like $\{100.5, 101.8, 100.2, 103.9, 104.6, 107.2\}$ — the cyclical component (data minus trend) stays small, roughly $\pm 1$ to $\pm 2$ points, meaning most of the series' movement is classified as "trend," leaving little for the cycle to explain. With a large $\lambda$ (say 100,000), the trend flattens to something close to a straight line, roughly $\{100.5, 102.0, 103.5, 105.0, 106.5, 108.0\}$ — now the cyclical component swings much more, for instance quarter 3's actual (99) minus this flatter trend (103.5) gives a cycle reading of $-4.5$, a much larger "below trend" signal than the low-$\lambda$ version gave. The same raw data can support very different "how far below trend is the economy right now" conclusions purely from the choice of $\lambda$.

Worked example 2: reading a cycle signal for a trading decision

Suppose a commodity-linked equity index's HP-filtered cyclical component currently reads $c_t = -3.2\%$ (the index sits 3.2% below its smooth long-run trend) with a historical standard deviation of cyclical deviations of about 2%. A mean-reversion strategy interpreting this as "1.6 standard deviations below trend" might read this as an oversold signal worth a long position, betting on reversion toward the trend line — but only if the trend itself is trustworthy, which depends heavily on the $\lambda$ chosen and, as the warning below explains, on whether the most recent data points are stable.

What this means in practice

The HP filter is a standard, quick way to extract a business-cycle or trend-cycle decomposition from macro or price series without fitting a full structural model, and it's widely used to construct "gap" measures (output gap, valuation gap) for regime and mean-reversion signals. It sits alongside changepoint detection (see Changepoint Detection and CUSUM) as a way to characterize how a series' level is behaving over time, though the two answer different questions — smooth decomposition versus discrete break detection.