If today marked the onset of a recession, then this event would be at odds with the historical track with respect to
- oil prices (low)
- interest rates (low)
- job creation (high)
A constellation unseen before. But each recession is different. So: what does the data tell us, right now, about the eventual onset of an `at odds’ recession?
1. Indicator System
Instead of relying on a single-indicator design, I here propose a system of four different indicators with different timing/smoothing abilities. The aim is to exploit `typical patterns’ of the system, in order to retrieve – hopefully – more information from the data.
The first differentiation is a methodological one. I propose
- A classic trend-growth indicator (slightly lagging or coincident)
- A cycle-magnification indicator which assigns more weight to the business-cycle than to the secular trend (coincident or leading). The resulting indicator is neither a pure trend (as the previous one) nor a pure cycle design; it’s a hybrid derived from asset management purposes.
The second differentiation is data-based. Each of the above designs is applied to log-differenced data whereby the span of the differences are either a month (~24 days) or a quarter (~72 days).
- Quarterly differences (abbreviated as q-t-q in the figures below) are closer to growth-dynamics as observed in quarterly GDP
- Monthly differences (abbreviated as m-t-m in the figures below) are slightly faster and noisier as can be seen in the following figure
Combining the two indicator designs (trend/magnified-cycle) and the two data-sets (q-t-q/m-t-m), I thus obtain four different indicators for my system:
- Design 1: the q-t-q trend (tends to lag)
- Design 2: the q-t-q magnified-cycle (roughly coincident)
- Design 3: the m-t-m trend (roughly coincident)
- Design 4: the m-t-m magnified-cycle (tends to lead)
I here target the trend in the industrial production index, INDPRO (the target is the same for trend and for magnified-cycle).
- Specifically I target an ideal trend with cutoff two years: all components with duration shorter than two years (of INDPRO) are eliminated.
- The ideal trend is symmetric and cannot be computed towards the sample end (it’s not a real-time design).
- I estimate the ideal trend by using the new mixed-frequency extension of MDFA.
- Data: monthly (INDPRO, PAYEMS, UNRATE), weekly (ICSA) and daily (SP500)
3. Empirical Results
The main results can be seen here: indicator-system.pdf (most graphs below are pasted from the document).
- slides 3-12 are introducing the topic
- slides 13-16 compare (real-time) indicators, designs 1-4, and target signals i.e. ideal trends (which cannot be computed in real-time)
- Slide 17 compares all four designs
- Slides 18-21 assess timing abilities of the indicators: I trade the SP500 index based on indicator signals: design 1 (lagging) performs less well than design 4 (leading); designs 2 and 3 perform best (coincident)
- Slides 23-26 compare the sub-series of the multivariate designs
The following figures are pasted from the slides.
3.1 Comparison of Indicators
A signal is triggered whenever an indicator passes the zero line. Design 4 (brown line) is leading, followed by designs 2 (blue) and 3 (green), and design 1 (red) is last (lagging). As of now (data up 25.01.2016) three out of four designs trigger a downturn alarm in the INDPRO-series. Let’s plot the subseries of the multivariate design (the indicators are obtained as aggregate sums) in order to identify triggers.
Here I obtain the subseries of design 1 (lagging):
For design 1, the main (and only) trigger is INDPRO (orange). All other subseries are positive (though engaged on a declining path).
And now the subseries of design 4 (leading):
For the leading design 4, three out of five series drag the indicator down: INDPRO (orange) dominates, followed by ICSA (weekly initial claims, green) and SP500 (cyan). The employment series are at or near the zero-threshold.
4. Conclusion INDPRO
The recent dynamics of the system of indicators show a typical downturn-pattern:
- Increasingly pervasive (indicators as well as subseries)
- Consistent chronological ordering (leading to lagging)
The main trigger of the current downturn is INDPRO. Can the sector-dynamic spread over to the whole economy? Let’s briefly look at the GDP-indicators.
5. System of GDP-Indicators
The approach here is generic and allows for arbitrary target series: one can simply substitute GDP to INDPRO in the target specification and obtain the following figure (designs 1-4):
- Note that the set of explanatory variables is the same as for INDPRO above: GDP is not therein. Thus publication lags and revisions are alleviated.
- Month-to-month differences are not available for GDP. In the frequency-domain, however, one can easily generate a corresponding artificial target (even if the series is not observed). Therefore, one can obtain results for q-t-q designs 1,2 as well as for m-t-m designs 3,4 in the above figure.
In comparison to the INDPRO-indicators, the GDP-based indicators are dragged-down less severely: both q-t-q designs (1,2) are still anchored in the positive growth-zone; only the fastest design 4 entered negative territory affirmatively.
An overview of the subseries of design 1
suggests that all series, except INDPRO, are still positive; in contrast to section 3.2, INDPRO now stays at the zero line i.e. it is not pulling the indicator into negative territory, yet.
The above results suggest that the industrial production sector is in trouble. As claimed by the board of governors “Industrial production declined 0.4 percent in December, primarily as a result of cutbacks for utilities and mining“. The economic pressure from the contracting industry-sector is distributed heterogeneously over the states of the union (in conjunction with low oil prices, see 1). It is not clear, as of now, if the drawdown will spread over other sectors and/or over a majority of states. The GDP indicators do not suggest irreversible contamination, yet.