Industrial Production

To evaluate the forecast performance of different PI and PF schemes, we choose monthly data of IPI as reference series to assess the state of the economy. The data is obtained from the Federal Statistical Office (Destatis). As described in section 1.3.1, the weights used by Destatis to aggregate the branch level production to the sectoral IPI series are used for the construction of the sectoral Ifo series as given in Table 1.1. Although the share of the industrial sector in total gross–value added amounts to not more than one forth, IPI mirrors the state of the economy comparably well and shows a high correlation to quarterly values of real GDP. Hinze (2003) shows 18_{See the Ifo Employment Barometer for employment as target and the Ifo Credit Constraint}

that the relative performance of business–cycle indicators is mostly independent of the chosen measure for aggregate output. In fact, due to the higher variance of

production, the industrial sector is frequently referred to as the cyclemaker of the

German economy.19 _{Its monthly publication is a major advantage of choosing IPI}

as reference since the Ifo business survey series – as most business–cycle indicators – are published on a monthly frequency as well. The monthly set–up prevents any information loss due to quarterly aggregation.

Since the Ifo business survey series fluctuate around stationary values, we detrend the target series to satisfy stationarity conditions. As we aim at evaluating the quan- titative forecast performance of several PI and PF strategies, IPI is transformed to

growth rates.20 _{As shown in Figure 1.6, the Ifo business survey series for manufac-}

turing industry exhibit a high correlation to the annual growth rates of IPI.

1992 1996 2000 2004 2008 80 100 Business Climate 1992 1996 2000 2004 2008 80 100 Business Situation 1992 1996 2000 2004 2008 80 100 Business Expectations −10 0 10 −10 0 10 −10 0 10

Notes: The thick lines represent the Ifo Business surveys (LS) and the thin line gives the annual growth rate of IPI (RS).

Figure 1.6: Ifo indices for manufacturing industry and annual growth rate of IPI Hence, we follow Hüfner and Schröder (2002), Dreger and Schumacher (2005), Fritsche and Kuzin (2005) and the analysis of the Sachverständigenrat (2005) who choose annual growth of IPI as target series.

In contrast to symmetric statistical filter approaches as the HP filter or other band– pass filters, growth rates can be interpreted as asymmetric filters, implicating that

19_{See e.g. Sachverständigenrat (2005).}

20_{To evaluate the performance of indicators for the detection of turning points in the business}

cycle, the reference series are frequently detrended by means of symmetric filter techniques, see e.g. Abberger and Nierhaus (2008) for an application with Ifo survey data.

the resulting series do not change when new observations are added. This comes along with the drawback that asymmetric filters lead to phase shifts of the reference time series towards the beginning, i.e. peaks and troughs are dated earlier. Thus, leading indicators potentially loose their lead character. Additionally, as the cyclical component of a reference series can be regarded as the combination of various ideal cycles of different cycle length, calculating growth rates increases the relative size of

shorter cyclical components.21

Forecasting month–to–month growth rates of IPI via month–to–month changes of the Ifo business survey series removes the phase shift problem and the predictor variables maintain their potential qualities as leading indicators. This comes along with the drawback that the resulting series are noisier and the correlation between the indices and the reference series is considerably lower. As it is a priori not clear which of the transformations is optimal, we present the corresponding results for month–to–month changes in Appendix 1.C.2.

In the present study, we exclusively use latest–available (final) data of IPI instead

of real–time data.22 _{We thus circumvent the problem of which vintage to choose as}

target series. Figure 1.7 shows the overlapping time series of the monthly vintages from March 2000 to March 2009 for annual and monthly growth rates of IPI and displays the correlation coefficients of the first vintage with the subsequent vintages. Obviously, the revisions regarding IPI in Germany are only minor such that the results of our forecast experiment should hold whatever vintage is used as estimation 21_{Let the complex cyclical function be represented by a weighted sum ofisine waves of varying}

cycle length

c(y) =X i

aisin(kiy)

with ki determining the frequency and ai the weight in the total cycle of subcycle i. Thus, the greaterki, the shorter is the length of subcyclei. The derivative of the cycle corresponds to the short term growth rate:

c0(y) =X i

aikicos(kiy)

As the derivatives of sine waves are cosine waves, the transformation corresponds to a shift to the left. After the differentiation, the kis modify the original weights ai of the subcycles, over– weighting the shorter cycles.

22_{Schumacher and Breitung (2008) show that data revisions do not affect the forecast perfor-}

and realization base. In fact, the bivariate correlation decreases with an increasing

distance between the vintages but remains above 0.99 for the annual growth rate

and 0.92 for the more volatile month–on–month rate.

1992 1996 2000 2004 2008 −10 −5 0 5 10

IPI annual rate − vintage data

1992 1996 2000 2004 2008

−5 0 5

IPI monthly % change − vintage data

1 30 60 90 0.92 0.94 0.96 0.98 1

IPI annual rate − revision

correlation coefficient number of vintage 1 30 60 90 0.92 0.94 0.96 0.98 1

IPI monthly % change − revision

correlation coefficient

number of vintage

Notes: The figures show monthly overlapping vintages of y–y and m–m growth of IPI from March 2000 to March 2009 and the corresponding bivariate correlation coefficients of each vintage to the first vintage.

Figure 1.7: Real–time data of industrial production