For once my forecasts concerning forecasts weren’t false

Hostages to fortune

Recently, I wrote a note on expected inflation in the USA. One of the two sources of data I used was the Livingston survey of expert forecasters. I used the median forecasts of the CPI 6 and 12 months ahead from the start of the survey in 1946 through 2003, when the survey format was changed. Forecasts from June 2004 though December 2013 are stored in a separate file which I did not download. The survey is conducted each June and each december, so there are twenty data points outside of the sample which I analyzed. This note discusses 40 numbers, forecasts CPI inflation over the following 6 and 12 months made in June and in December of 2004 through 2013.

Based on forecasts made through 2003 and on TIPS break evens, I made some predictions.

“The next step is to use Livingston forecasts from 2003 though 2013. I hazard some predictions based on the results reported here. First, I expect lagged inflation, especially lagged PCE core inflation, to fit more recent Livingston forecasts too. I do not expect to find a significant increase in expected inflation (after controlling for lagged inflation) when Bernanke was Fed chairman compared to Greenspan in spite of the fact that Bernanke is the first Fed chairman to declare an inflation target higher than lagged inflation (and in spite of huge efforts at unconventional monetary policy). Finally I do not expect the new median Livingston forecasts to markedly outperform simple autoregressive forecasts.”

Extraordinarily, my predictions were correct.

First lagged personal consumption expenditure deflator inflation excluding food and energy is highly correlated with forecast CPI inflation – the median Livingston inflation expectation is not anchored.
Here is a simple regression of forecast inflation over the next 6 months (finf6) on PCE core inflation over the preceding 6 months estimated using only the new data points.

. reg finf6 lpcecinf6 if dat>2003.6

Number of obs = 20
R-squared = 0.2824
Root MSE = .78778

finf6 | Coef. Std. Err. t
lpcecinf6 | .775133 .2912517 2.66
_cons | 1.021819 .5886572 1.74

Here is a regression of CPI inflation forecast for the next 12 months finf12 on lagged PCE core inflation over the preceding 12 months (lpcecinf). Since the forecasts overlap, I report Newey-West standard errors assuming ma1 disturbances

Regression with Newey-West standard errors Number of obs = 20
maximum lag: 1 F( 1, 18) = 12.91
Prob > F = 0.0021

finf12 | Coef. Std. Err. t
lpcecinf | .6182174 .1720291 3.59
_cons | 1.172926 .3310734 3.54

In fact, expected inflation was actually lower when Bernanke was chairman than when Greenspan was chairman in a regression controlling for lpcecinf6 (the indicator for the Greenspan periods is excluded – the indicators name the chairman at the time of the realization of inflation).

. newey finf12 lpcecinf martin burns miller volcker bernanke,lag(1)

Regression with Newey-West standard errors Number of obs = 110
maximum lag: 1

| Newey-West
finf12 | Coef. Std. Err. t
lpcecinf | .6810097 .0498081 13.67
martin | -1.287283 .1595411 -8.07
burns | .1574747 .3685694 0.43
miller | 1.582327 .2260505 7.00
volcker | 1.125648 .3491428 3.22
bernanke | -.4251683 .1592886 -2.67
_cons | 1.451395 .1725814 8.41

This remains true when the unemployment rate at the time the forecasts were made is included in the regression

. newey finf12 lpcecinf martin burns miller volcker bernanke lunem,lag(1)

Regression with Newey-West standard errors Number of obs = 110
maximum lag: 1
| Newey-West
finf12 | Coef. Std. Err. t
lpcecinf | .6769935 .0495356 13.67
martin | -1.328768 .1664241 -7.98
burns | .2100739 .3899154 0.54
miller | 1.633938 .2342526 6.98
volcker | 1.284622 .4076517 3.15
bernanke | -.3208668 .1771306 -1.81
lunem | -.0669514 .0638862 -1.05
_cons | 1.838317 .4379232 4.20

Finally, the last prediction included the weasel word “markedly”. In fact the median Livingston forecast slightly outperforms forecasts based only on lagged PCE core inflation. As in the earlier note, I estimated the constant and coefficient using realizations from June 1973 and before – that is before the oil shocks.

The root means squared error of the median Livinston 12 month forecast made 2004 or later (and referring to annual CPI inflation realized June 2005 or later) is 1.3390063 % . The root mean square forecast using two parameters estimated with data at least thirty years old is 1.4425952 % not “markedly” higher.

Livingston survey participants did not have information on the deflator for the month when they made their forecasts. Also there are preliminary estimates and revisions. Therefore I also look at forecasts which are a constant plus a constant times annual core PCE inflation lagged six months (lohpcecinf). Again the parameters were estimated with data from before the oil shocks. The root mean squared forecast error is greater but only 1.4759438 % which is not markedly higher than 1.3390063 %.

The three predictions are confirmed by analysis of out of sample data.