by Dave
I trained up a “recession detector” using as the training key the “USREC” (US recession declared) monthly timeseries from FRED, with the longest-dated macro timeseries I could find from FRED. These included: PPIACO, INDPRO, BAA, AAA, DJI, DJT.
In my training session, I trained the model on data from 1920-2004. Then I presented it with data it had never seen before: 2005-2018. Given that “new” data, it was properly able to predict the 2008 recession, but it also (curiously) saw Chris’s 2016 recession that never was.
What predicts a recession?
- Declining DJT/DJI ratio (weight 5..10, immediate)
- Declining INDPRO (weight 40..50, immediate)
- Declining PPIACO (weight 30, immediate)
- Rising AAA rates (weight 15..20, delay 12 months)
- Rising BAA rates (weight 20..30, delay 6 months)
- Rising BAA-AAA ratio (weight 15…40, immediate)
INDPRO tipped over in early 2015 and dropped for a full year. PPIACO did much the same thing.
Is this the holy grail? It might be. I watch it every month. You can watch the raw indicators too: INDPRO continues to look strong, so does PPIACO. Rates, on the other hand, have been rising; they are a ticking timebomb. Notice how they have a “delay” attached; and the pressure starts relatively low, and slowly builds until things pop.
Its really fun when the model matches up with how you think the world should work. The best part are the specific weight & delay numbers that pop out. Of course, these are just average correlations that happen across all recessions 1920-present, so…YMMV.
From what I can see, the time to panic is when you see INDPRO and PPIACO tip over, after seeing AAA and BAA rates rise for 6-12 months. Transports leading industrials lower is icing on the cake.
