Forecasting unemployment: a small business-survey model; Small business plans and expectations provide accurate and timely forecasts of national employment

Business Economics, Oct, 2004 by William C. Dunkelberg, Jonathan A. Scott, William J. Dennis, Jr.

Conclusion

Overall, the NFIB employment variables do an excellent job of anticipating changes in the national unemployment rate (with a one quarter lag). The performance is driven primarily by the percent of firms with at least one "hard to fill" job opening, with minor contributions from plans to create new jobs and expectations about real sales growth for the company in the three months following the survey. When the model is estimated over the full 30 years of data, however, it does not perform well after 2001. This appears to be due to the unusual buildup of employment in the late 1990s through the first half of 2000, supported by unusually high growth in the labor force. This created excess labor supply that was released reluctantly due to widespread belief that the recession would be short, the recovery strong, and the difficulty of obtaining employees during the previous expansion. (16) The changing composition of the business sector (e.g., more service firms) may also have been a factor contributing to the deterioration in forecasting performance, as evidenced by the superior performance of the equation estimated using data only from 1993 to 2003. The 1995-2000 period may have been very unusual and unlikely to be repeated, possibly restoring the performance of the equation based on all 30 years of data. An alternative interpretation is that structural changes in the economy may have made past empirical relationships less reliable, tilting the forecasting edge in favor of equations estimated over more recent data.

TABLE 1 U.S. UNEMPLOYMENT RATE REGRESSION EQUATION

Equation             1               1a               1b
Time Period      1974-2003        1974-1983        1984-1993
                   Coeff     t     Coeff      t     Coeff      t

HIREPLAN-1        -0.054     3.1   0.034      1.4   0.012      0.3
JOBOPEN-1         -0.182    11.6  -0.257     10.8  -0.207      9.4
ERSALE-1          -0.001     0.2  -0.013      1.7  -0.016      0.8
Constant          10.62     34.1  12.29      26.5  10.28      20.6
SSQ              245.8            61.0             63.0
R-squared (adj)    0.81            0.86             0.85
N                120              40               40

Equation            1c
Time Period      1994-2003
                  Coeff      t

HIREPLAN-1       -0.004      0.2
JOBOPEN-1        -0.161     10.2
ERSALE-1         -0.008      1.2
Constant          9.48      40.5
SSQ              31.9
R-squared (adj)   0.94
N                40

TABLE 2 ACTUAL AND PREDICTED UNEMPLOYMENT RATES

               FULL MODEL (EQ 1)       PERIOD 3 MODEL (EQ 1c)
Time Period  Actual  Predicted  Error  Predicted  Error

Q1-2001       4.23     4.03      0.20    3.90      0.33
Q2-2001       4.40     4.10      0.30    4.00      0.40
Q3-2001       4.83     5.23     -0.39    4.89     -0.06
Q4-2001       5.57     5.37      0.19    5.18      0.39
Q1-2002       5.67     6.40     -0.73    5.76     -0.10
Q2-2002       5.83     6.49     -0.66    5.93     -0.10
Q3-2002       5.73     6.13     -0.40    5.65      0.08
Q4-2002       5.87     6.06     -0.19    5.57      0.30
Q1-2003       5.83     6.55     -0.71    6.00     -0.16
Q2-2003       6.13     6.71     -0.58    6.08      0.06
Q3-2003       6.13     7.07     -0.94    6.39     -0.26
Q4-2003       5.87     7.21     -1.35    6.55     -0.69
Q1-2004       6.61                       5.97
AVERAGE ABSOLUTE ERROR           0.59              0.24

TABLE 3 BETA COEFFICIENTS

            Standard   Regression       Beta
Predictor   Deviation  Coefficients  Coefficients

HIREPLAN-1     5.4       -0.054         -0.2
JOBOPEN-1      5.9       -0.182         -0.75
ERSALE-1      10.2       -0.001         -0.01
STD. ERROR OF UNE = 1.43