Economics · Model

Extracting Technology Shocks

A Real Business Cycle model needs two numbers to describe technology: how long a deviation from its long-run path lasts (persistence, \(\rho\)) and how large the surprises are (the standard deviation of the shock, \(\sigma\)). This page measures them step by step from US data, following King and Rebelo (1999, section 4.1).

Productivity
Technology

Loading data…

Step 1 Start from measured productivity

Total factor productivity (TFP) is the part of output growth that is not explained by the growth of capital and labour, each weighted by its share of income: the Solow (1957) residual. John Fernald publishes it each quarter for the US business sector as a growth rate, \(g_{A,t}\), in percent at an annual rate (400 times the change in the logarithm). Adding up these growth rates gives the level of TFP, shown here as an index equal to 100 on average in 2017.

TFP rises over time, but not at a steady pace: it grows quickly in some periods and stalls or falls in others.

TFP level

Loading chart…

Step 2 Take logarithms and fit a linear trend

TFP follows the law of motion

\[ A_t = A_{t-1}\, e^{g_{A,t}/400} , \]

since Fernald’s growth rates are 400 times the change in the logarithm. Taking logarithms, the logarithm of TFP is the sum of past growth rates,

\[ \ln A_t = \ln A_{t-1} + \frac{g_{A,t}}{400} = \sum_{s=1}^{t} \frac{g_{A,s}}{400} , \]

starting from \(\ln A_0 = 0\). Working in logarithms turns proportional changes into differences, and a constant growth rate into a straight line; the chart shows \(100 \times \ln A_t\), so a change of 1 is a change of about 1 percent.

A straight line is fitted by ordinary least squares,

\[ \ln A_t = \alpha + \beta\, t + c_t , \]

where \(\alpha\) is the constant, the slope \(\beta\) is trend growth per quarter (\(4\beta\) per year), and the error \(c_t\) is the cyclical component of TFP. The line uses the full history of the series, from 1947.

Log TFP and its linear trend

Loading chart…

Step 3 Keep the deviation from trend

The cyclical component, the error of the regression in step 2,

\[ c_t = \ln A_t - (\alpha + \beta\, t), \]

is how far technology is above (\(c_t > 0\)) or below (\(c_t < 0\)) its trend, in percent. It wanders slowly: once technology is above trend, it tends to stay above trend for many quarters.

Deviation of log TFP from its linear trend

Loading chart…

Step 4 Measure persistence

Persistence is measured by regressing \(c_t\) on its value one quarter earlier, without a constant:

\[ c_t = \rho\, c_{t-1} + \varepsilon^A_t, \qquad \hat{\rho} = \frac{\sum_t c_{t-1}\, c_t}{\sum_t c_{t-1}^2} . \]

Each point in the chart is one quarter, with last quarter’s deviation on the horizontal axis and this quarter’s on the vertical axis; the line through the origin has slope \(\hat{\rho}\).

A slope close to 1 means that a deviation from trend dies out slowly. After \(k\) quarters, a fraction \(\rho^k\) of it remains, so half of it is gone after \(\ln 0.5 / \ln \rho\) quarters.

Persistence: deviation from trend, \(c_t\), against its previous value, \(c_{t-1}\)

Loading chart…

Step 5 Recover the technology shocks

The technology shock is the part of this quarter’s deviation that could not be predicted from last quarter’s,

\[ \varepsilon^A_t = c_t - \hat{\rho}\, c_{t-1}, \]

the vertical distance between each point in step 4 and the line. Its standard deviation, \(\sigma\), measures the typical size of a shock. The shocks are written \(\varepsilon^A_t\) for measured TFP and \(\varepsilon^{A^u}_t\) for utilization-adjusted TFP (\(\varepsilon^Z_t\) and \(\varepsilon^{Z^u}_t\) for Harrod-neutral technology \(Z\)).

Technology shocks

Loading chart…

Estimates

The table collects the numbers estimated in steps 2 to 5, for each version of technology; the row for the version selected at the top of the page is highlighted. Persistence \(\hat{\rho}\) comes from step 4, and the half-life, \(\ln 0.5 / \ln \hat{\rho}\), is the number of quarters after which half of a deviation from trend has gone. The standard deviation of the shock, \(\sigma\), comes from step 5. Trend growth is the slope of the line in step 2, \(4\beta\), in percent a year. The share of quarters with falling technology counts the quarters in which the growth rate of step 1 is negative; in a model driven only by these shocks, every one of those quarters is a technological regress (King and Rebelo, 1999, section 5). All are estimated over the sample shown in the last column.

Step 6 Feed the model

In the model, technology is written as its percent deviation from trend, \(\hat{A}_t\), and follows the process estimated in steps 4 and 5:

\[ \hat{A}_{t+1} = \rho\, \hat{A}_t + \varepsilon^A_{t+1}, \qquad \operatorname{sd}(\varepsilon^A) = \sigma . \]

Feeding the model the measured shocks gives its technology path. The starting value is taken from the data: technology starts at its deviation from trend in the first quarter, \(\hat{A}_0 = c_0\) (the first quarter of 1947), and the shocks enter from the next quarter on. Equivalently, the quarter before is set so that \(\hat{\rho}\,\hat{A}_{-1} + \varepsilon^A_0 = c_0\):

\[ \hat{A}_0 = c_0, \qquad \hat{A}_t = \hat{\rho}\, \hat{A}_{t-1} + \varepsilon^A_t \quad (t \geq 1). \]

The first chart compares this path with the deviation of log TFP from its linear trend, \(c_t\), from step 3. The shocks were defined in step 5 as \(\varepsilon^A_t = c_t - \hat{\rho}\, c_{t-1}\), so with the same starting value the recursion rebuilds \(c_t\) exactly, quarter after quarter: the two lines coincide (the orange line lies exactly beneath the blue one), which checks that the shocks and the persistence have been extracted correctly. Starting instead from \(\hat{A}_0 = \varepsilon^A_0\), or from the steady state, would leave a gap equal to the initial deviation multiplied by \(\hat{\rho}\) each quarter, which with \(\hat{\rho}\) close to 1 takes many years to fade. Both lines still contain slow, long-lasting swings, because a linear trend leaves them in. King and Rebelo (1999), with data for 1947–1996, find \(\rho = 0.979\) and \(\sigma = 0.72\) percent.

Model technology and the deviation of log TFP from trend

Loading chart…

Comparing the model with the data: the HP filter

The slow swings are not a problem for the comparison, because King and Rebelo compare model and data only after filtering both in the same way. Their simulated business cycles (1999, section 4.3.1 and Figure 7) feed the model the shocks \(\varepsilon^A_t\) from step 5, and then, in the words of the figure’s note, “all variables are detrended using the Hodrick–Prescott filter”: output, consumption, investment and hours from the model and from the data, with \(\lambda = 1600\). Their model statistics (Table 3) are likewise computed on the HP cyclical components of the simulated series, so they can be set against the data statistics of the Real Business Cycles page. The shocks themselves do not change; only the final comparison uses the HP filter.

The second chart applies this to technology itself. It shows the HP cyclical component of the model’s technology, \(\hat{A}_t\), and the HP cyclical component of log TFP in the data. The HP filter removes a linear trend completely, so filtering \(c_t\) or log TFP gives the same cycle; since \(\hat{A}_t = c_t\), the two lines coincide. The same procedure, applied to the model’s output, consumption, investment and hours, produces the series that will be compared with the cycles of the data on this site.

After the HP filter: model technology and TFP in the data

Loading chart…

Variants and background

A shortcut: HP-filtered technology shocks

The shortcut. Instead of simulating the model with the shocks of step 5 and filtering its output afterwards, the shocks can be extracted from the HP cyclical component of log TFP, \(c^{HP}_t\), using the persistence \(\hat{\rho}\) estimated in step 4:

\[ \tilde{\varepsilon}^A_t = c^{HP}_t - \hat{\rho}\, c^{HP}_{t-1} . \]

Fed into the model, these HP-filtered technology shocks produce series that are already HP cyclical components and can be compared directly with the cycles of the data, with no filtering of the model’s output.

Why it works. Each step is a linear operation that does not depend on the date, and such operations can be applied in any order. Write \(H\) for the HP filter and \(L\) for the lag operator (\(L x_t = x_{t-1}\)):

  1. The HP filter removes a linear trend completely, so the HP cycle of log TFP equals the HP cycle of its deviation from the linear trend: \(c^{HP}_t = H \ln A_t = H c_t\).
  2. The shocks of step 5 are \(\varepsilon^A_t = (1-\hat{\rho} L)\, c_t\). Changing the order of the two operations gives \(\tilde{\varepsilon}^A_t = (1-\hat{\rho} L)\, H c_t = H (1-\hat{\rho} L)\, c_t = H \varepsilon^A_t\): the shortcut shocks are the HP-filtered shocks of step 5.
  3. The solution of a log-linear model is also linear: output is a weighted sum of current and past shocks, \(\hat{y}_t = G(L)\, \varepsilon^A_t\). Feeding it the filtered shocks gives \(G(L)\, H \varepsilon^A_t = H\, G(L)\, \varepsilon^A_t\), which is the HP cycle of the model’s output. The same holds for consumption, investment, hours and technology itself.

Filtering at the start, on the shocks, or at the end, on the model’s output, therefore gives the same result: the shortcut is King and Rebelo’s procedure with the steps in a different order.

Where it fails.

  • At the ends of the sample. The HP filter treats the first and last few years differently from the middle: it is two-sided and, at the ends, has no future or past data to use. There the order of operations no longer cancels out, and the simulated series differ from King and Rebelo’s, most at the end of the sample, where each new quarter of data also revises the latest shocks. For the most recent quarters, and for anything forward-looking such as forecasts, simulate with the shocks of step 5 and filter afterwards.
  • Persistence must be the linear-trend estimate. The model’s \(\rho\) and the \(\hat{\rho}\) used to extract the shocks must be the same, from step 4. Re-estimating \(\rho\) on the HP cycle itself would break the equivalence and give a different, less persistent technology process.
  • The shortcut shocks are not true shocks. They are not unpredictable from one quarter to the next (they are serially correlated), and their standard deviation is not the model’s \(\sigma\). Use the shocks of step 5 for calibrating \(\sigma\), for computing model statistics, and for random simulations.
  • The model must be linear. The argument holds for a log-linear (first-order) solution. With a second-order or fully non-linear solution, filtering and simulating no longer commute.
  • Starting point. Technology should start from its value in the data, as in step 6; the model’s capital stock also needs a starting value, and the effect of that choice fades only over the first years of the simulation.
Utilization-adjusted TFP

When firms work their machines and employees harder in booms, measured TFP rises even if technology has not changed. Fernald’s utilization-adjusted series removes this effect (Basu, Fernald and Kimball, 2006), so its shocks are smaller and less tied to the business cycle. King and Rebelo (1999, sections 7 and 8) show why this matters: once utilization is accounted for, technology shocks are much smaller, and a model needs stronger amplification to generate realistic cycles. The steps are the same; use the switch at the top of the page.

Hicks- or Harrod-neutral technology

The steps above treat technology as Hicks-neutral: it multiplies the whole production function, \(Y = A K^{\alpha} N^{1-\alpha}\), as in King and Rebelo, so \(A\) is TFP itself. If technology instead multiplies labour (Harrod-neutral or labour-augmenting, \(Y = K^{\alpha}(ZN)^{1-\alpha}\)), then \(A = Z^{1-\alpha}\), and

\[ \Delta \ln Z_t = \frac{\Delta \ln A_t}{1-\alpha_t} , \]

the growth of TFP divided by labour’s share of income. The page does this quarter by quarter with Fernald’s capital share \(\alpha_t\), then applies steps 2 to 5: the shocks are about one and a half times larger and as persistent. The two versions describe the same data; the one to use is the one that matches where technology enters the model’s production function.

Prescott (1986)

Prescott introduced the idea of measuring technology shocks with the Solow residual and of choosing the model’s technology process to match it. He treated the residual as close to a random walk and measured the size of the shocks from its quarter-to-quarter changes, rather than from a linear trend and an estimated first-order autoregression; the step-by-step method on this page is King and Rebelo’s.

References
  • Basu, S., J. G. Fernald and M. S. Kimball (2006). “Are Technology Improvements Contractionary?” American Economic Review 96(5): 1418–1448.
  • Fernald, J. G. (2014). “A Quarterly, Utilization-Adjusted Series on Total Factor Productivity.” Federal Reserve Bank of San Francisco Working Paper 2012-19 (updated March 2014).
  • King, R. G. and S. T. Rebelo (1999). “Resuscitating Real Business Cycles.” In J. B. Taylor and M. Woodford (eds.), Handbook of Macroeconomics, Vol. 1B, Chapter 14: 927–1007. Amsterdam: Elsevier.
  • Prescott, E. C. (1986). “Theory Ahead of Business Cycle Measurement.” Federal Reserve Bank of Minneapolis Quarterly Review 10(4): 9–22.
  • Solow, R. M. (1957). “Technical Change and the Aggregate Production Function.” Review of Economics and Statistics 39(3): 312–320.

Page created on 07/11/2026.

© Arthur Galichère.

  • Errors: Every effort has been made to ensure the data and calculations are accurate; any remaining errors are mine. If you believe you have found a mistake, or have a comment or suggestion, please let me know.
  • Reuse and citation: Total factor productivity and capital’s share of income are from John Fernald at the Federal Reserve Bank of San Francisco. The estimates, charts and text are the author’s; the source data remain subject to the terms of their providers. When reusing these figures, please cite the original data providers and this website: Arthur Galichère, Extracting Technology Shocks.