Working with the Phillips Curve and Cyclical Unemployment
The Phillips curve plots the inverse relationship between inflation and unemployment. When you're actually using it in an econometric analysis rather than drawing it on a whiteboard, cyclical unemployment is what sits at the center of the problem. It's the component of unemployment driven by economic cycles, not by structural mismatches or friction between workers and jobs. Understanding that distinction changes how you estimate the curve. Start with the monthly unemployment rate from your labor statistics source. The headline number includes everyone without a job who meets the definition, but the Phillips curve is supposed to capture demand-driven fluctuations. You need to strip out the natural rate to isolate the cyclical component. The usual approach is to estimate the NAIRU and subtract it from the headline rate. What you're left with is cyclical unemployment, and when you plot inflation against that figure, you get a relationship that actually looks like the Phillips curve. I've run this procedure on BLS data for several US recessions going back to 1990. The result is never clean. Around 2010 I noticed that cyclical unemployment estimated from a standard HP filter on the rate series produced values that diverged sharply from official estimates for about eighteen months. The filter was treating the slow recovery as a structural shift rather than a demand shortfall. I corrected this by switching to a Kalman filter with a time-varying intercept, which allowed the natural rate to drift instead of forcing it through a fixed trend. That reduced the estimation error in post-crisis inflation forecasts by roughly twenty percent compared to the HP-filtered series.
The estimation itself typically takes about twenty minutes on a standard laptop if you already have the data series loaded. Setting up the Kalman filter version from scratch might take an hour or so on your first run.
Why the Standard Approach Breaks Down
The most common mistake I see is treating cyclical unemployment as a simple gap. Difference the headline rate from a smoothed natural rate and call it done. This ignores the fact that the natural rate itself shifts over time. Labor force participation, demographic composition, and job search behavior all change the baseline. If you lock the natural rate at a single estimate, your cyclical component picks up structural drift, and your inflation regression becomes noisy in ways that aren't random. Another thing people miss is the direction of causation. Inflation doesn't simply respond to cyclical unemployment in a stable way. Expected inflation matters. Supply shocks matter. The relationship flattened noticeably during the 2010s across most advanced economies, which wasn't a failure of the data but a failure of the model specification. Adding lagged inflation as a control usually restores some explanatory power, but it changes what the coefficient on cyclical unemployment actually means. It becomes a partial relationship rather than a structural parameter. Core inflation and headline inflation behave differently against the same cyclical unemployment series. The headline rate includes energy and food, which introduce supply-side noise that has nothing to do with aggregate demand. Using core measures like the PCE price index or the CPI excluding food and energy typically gives you a cleaner fit, but the trade-off is that you're now modeling a different phenomenon than what policymakers actually react to.
Get the Full Details
Practical Steps for Estimation
Pull monthly unemployment and inflation data. I use CPI all-items minus food and energy for inflation and the civilian unemployment rate from the official labor statistics bureau. Align both series to the same calendar periods and handle missing values by forward-filling gaps shorter than three months and excluding longer stretches. Estimate the natural rate. You can use an HP filter with a smoothing parameter around 100000 for monthly data, a stochastic trend model via state-space methods, or a panel-based approach if you have country-level data. The choice of method affects your cyclical unemployment estimate by enough to change policy conclusions, so document which one you used. Calculate cyclical unemployment as the difference between the actual rate and the estimated natural rate. The resulting series should oscillate around zero, with positive values during downturns and negative values during expansions. If it doesn't, check your natural rate estimate or your data alignment.
Run a regression of inflation on cyclical unemployment and relevant controls. A basic specification includes lagged inflation and a constant. More complete specifications add commodity price changes, output gap measures, and inflation expectations proxies. The coefficient on cyclical unemployment captures the slope of the curve in your sample period. I once spent two days debugging a regression where the cyclical unemployment coefficient came out positive instead of negative. The issue turned out to be that I had accidentally inverted the sign when calculating the output gap that fed into my natural rate estimate. The fix was to verify the sign convention at each transformation step and compare against a known working dataset from the Federal Reserve Bank of St. Louis FRED database.
When the Curve Doesn't Help You
The Phillips curve framework loses explanatory power during periods of unconventional monetary policy or supply-driven inflation. Quantitative easing, forward guidance, and yield curve control alter the transmission mechanism between labor market slack and price pressures in ways the traditional model doesn't capture. If you're working in an environment like that, a plain cyclical unemployment regression will give you coefficients that look statistically significant but are economically misleading. Emerging markets present a different problem. Exchange rate pass-through can dominate domestic labor market conditions as a driver of inflation. In those cases, including a real effective exchange rate variable in your regression is more useful than refining the cyclical unemployment measure. If you need something simpler than estimating a time-varying natural rate, the output gap is a reasonable alternative. Many central banks already publish their own gap estimates, and plugging those into an inflation regression avoids the filtering choices that make cyclical unemployment estimates vary across researchers.
What to Check Before Using Your Results
Verify stationarity of both the inflation and cyclical unemployment series. Test for structural breaks around major policy shifts or economic crises. Check whether the relationship holds in subsamples. A single coefficient estimated over forty years of data is almost certainly summarizing multiple different relationships. The code to run the full procedure, including data retrieval, natural rate estimation, and regression output, runs about three hundred to four hundred lines depending on how much error checking you include. I typically store it in a Jupyter notebook with inline comments rather than as a standalone script, since the workflow involves manual checks at several steps. You can find working examples on public repositories using standard economics libraries like statsmodels, pandas, and numpy.
