"Why You Should Never Use the Hodrick-Prescott Filter"

Academic paper titles are usually cautious. "Some Evidence On." "A Reconsideration Of." "Toward a Theory Of."

James Hamilton published a paper in 2018 called "Why You Should Never Use the Hodrick-Prescott Filter." There is no hedge in it. The title is the conclusion, and it is aimed at a tool that at the time was used in a substantial fraction of all empirical macroeconomics papers and taught to every graduate student in the field.

The paper is worth reading not just for the argument about one filter, but because it is a masterclass in a much more general failure mode: a tool that reliably produces plausible-looking output, adopted by everyone, that turns out to be generating its own findings.

The problem: everybody was using it, almost nobody had checked it

The Hodrick-Prescott filter splits a series into a smooth trend and a wiggly cycle. You choose how stiff the trend is allowed to be, the filter bends it through the data, and whatever the trend fails to capture becomes "the cycle."

It is fast. It is simple. It always works. It produces charts that look exactly like what you expect a business cycle to look like. And that last property is the trap, because a tool that always produces plausible-looking output gives you no signal when it is producing nonsense.

Hamilton's paper makes four charges. Any one of them would be serious. Together they are close to fatal.

Charge one: the cycles are made up

This is the big one, and it is the one every quant needs to internalise.

Take a pure random walk. By construction, a random walk has no cycles at all. It has no trend, no periodicity, no oscillation. It is the purest possible example of a series with no cyclical structure whatsoever.

Run it through the HP filter.

The filter dutifully produces a "trend" and a "cycle." And the cycle it produces has apparent structure: it oscillates, it has characteristic periodicities, it has autocorrelation patterns that look meaningful. It looks exactly like a business cycle.

None of it is in the data. All of it was created by the filter.

The analogy is a machine that is supposed to sift gold from river silt, but which happens to shed tiny flakes of gold-coloured paint from its own casing. Every batch you run through it produces a satisfying glint of gold. You conclude the river is rich. In fact you are just recovering paint from the machine.

Hamilton showed that the "stylised facts" of the business cycle, patterns that had been measured on HP-filtered data and taught for thirty years, are partly properties of the HP filter rather than properties of the economy. That is a devastating thing to be able to say about a field's core empirical findings.

Charge two: the end of the sample behaves differently from the middle

The HP filter estimates the trend at any given date using data from both before and after that date. This is fine in the middle of your sample where both are available.

At the end of the sample it is not fine, because there is no "after." The filter is forced to behave differently there, and the trend estimate at the final observations is both noisier and systematically different in character from the estimates in the middle.

This has an ugly practical consequence. The filter revises history. The "output gap" you computed for the most recent quarter will change when next quarter's data arrives, and change again the quarter after. The number that policymakers most need, where are we relative to trend right now, is precisely the number the filter is worst at producing.

For quantitative traders this is worse than an inconvenience, it is an outright trap. If you HP-filter a price series across your entire backtest sample and trade the resulting deviations from trend, you have used the future to compute the past. Your model, at every historical date, knew what was going to happen next. The backtest will look magnificent. It is fiction. This is one of the easiest and most seductive lookahead biases to commit, and it happens constantly.

Charge three: the famous smoothing parameter is not justified

Everyone uses a stiffness of 1600 for quarterly data. It is recited like a constant of nature.

Hamilton points out that if you take the filter seriously as a statistical procedure and ask what value the data would actually imply, you get an answer wildly different from 1600. The number is not calibrated to anything, it is a convention, and the resulting cycles depend heavily on a choice nobody defends.

Charge four, and the constructive part: there is a simpler alternative

A critique without an alternative is easy to ignore, so Hamilton offers one, and its simplicity is almost insulting to the machinery it replaces.

He asks: what does "the cyclical component" actually mean? A reasonable answer is: the part of today's value that you could not have predicted a while ago. If you knew everything about the economy two years ago, some of today's output is what you would have forecast. That part is the trend, the predictable path. What you could not have forecast is the cyclical surprise.

So Hamilton's proposal is: regress today's value on a handful of values from a couple of years ago, and call the regression residual the cycle. That is it. A simple regression on lagged values.

This has several virtues that the HP filter lacks entirely:

• It has an interpretation. The cycle is "what was not forecastable," which is a meaningful economic concept rather than "what a springy ruler failed to catch."
• It uses no future data. At every point, the estimate depends only on the past. There is no revision problem and no lookahead bias. This alone makes it vastly safer for anyone building a trading signal.
• It does not invent cycles. Apply it to a random walk and it does not manufacture a business cycle out of nothing.
• You do not have to choose a stiffness parameter.

Why it mattered

• It forced a field to re-examine its own foundations. When a top journal publishes a paper telling everyone that a tool underlying thirty years of empirical work is broken, people notice. Usage of the HP filter has declined, and the "Hamilton filter" has become a standard alternative or robustness check.
• It is a permanent lesson about tooling. The HP filter survived for decades not because it was right, but because it was convenient and its output always looked correct. Convenience plus plausible-looking output is a genuinely dangerous combination in any research process, and it describes an enormous number of tools in the quant toolkit.
• The lookahead lesson is directly transferable. Any two-sided filter, any centred moving average, any smoothing procedure that uses future data to estimate the present, is contaminated in exactly the same way. If you can name three such things you have used in a backtest, this paper is about you.

The honest limitations

This is a polemic, and polemics overreach. It is worth being fair to the other side.

• The title is stronger than the evidence supports, and the response literature says so. Several economists have pushed back, arguing that the HP filter, used carefully and with awareness of its properties, remains informative, and that Hamilton's alternative has its own issues. There is now a genuine "Comment on Hamilton" literature, and the debate is not one-sided.
• Hamilton's alternative has its own parameter choices. How far back is "a couple of years"? How many lags? He proposes defaults, but they are defaults, not derivations. The arbitrariness has been reduced, not eliminated.
• The regression approach produces a noisier cycle. Because it does not smooth, the Hamilton filter's cyclical component can be considerably choppier than the HP version. Whether that is a bug or an honest reflection of the data depends on your view.
• Spurious cycles are a property of many filters, not just this one. The general result that filtering can create apparent periodicity where none exists is old, and applies to band-pass filters and other detrending methods too. Hamilton's argument is really an argument about detrending in general, and the HP filter is simply the most popular target.

The one-line takeaway

Hamilton showed that the most widely used detrending tool in macroeconomics manufactures cycles out of series that contain none, revises its own estimate of the present every time new data arrives, and rests on a smoothing constant nobody can defend, and his proposed replacement, a plain regression on lagged values, has the enormous virtue of never using data from the future.