Listen, there’s a lot of debate going on right now about late-cycle dynamics.
Chief among market participants’ concerns is the flattening curve, which some folks contend is trying to warn us that a recession is lurking in a dark alley a couple of blocks up the street, just waiting on us to walk by so it can reach out and grab us by the pussy (“when you’re a recession, they let you do it”).
And if all you’re looking at is the curve, well then I suppose there’s cause for concern. After all, the 5s30s touched 26bps this week following the CPI miss and a strong 30Y auction.
But that’s probably not all you should be looking at, and although predicting recessions is almost by definition an exercise in futility (if you could predict them accurately then you would presumably be able to avoid them), a look across a bunch of models suggests the probability is low, at least in the near-term. Here’s a smattering of charts:
(Clockwise from upper left: Deutsche, Goldman, BofAML)
Additionally, Wells Fargo really wants you to remember that “we’re late in the cycle” doesn’t count as “research”:
Investors realizing ‘We’re late in the Cycle’ is not Research. A year ago investors were told we’re late in the cycle. Two years ago investors heard the same proclamation as well as three. We think individuals are realizing that’s not research; it’s an observation.
And look, maybe the main problem here when it comes to drawing conclusions and making predictions about the trajectory of the economy is that we’re asking the wrong people or, more to the point, that we’re asking “people” at all. Maybe we should be asking this “guy”:
He looks friendly, what does he think?
Well BofAML is going to ask him. In a note dated Thursday, the bank makes what they call “a first attempt to combine macro analysis and machine learning in understanding markets.”
Call it “first contact” with Elon Musk’s “immortal robot dictator,” on the subject of macro.
Specifically, here’s what the bank did:
Instead of looking at past events in a time series format, we deconstruct history into individual economic configurations each month. Instead of arbitrarily comparing the present to past hiking cycles (one of strategists’ favorite exercises), we algorithmically quantify the closest historical parallels.
To achieve the former, we collect over 140 economic indicators, at a monthly frequency, encompassing various aspects of the economy going back to the beginning of 1993. We then apply PCA to reduce this dataset to eight factors (which roughly explains 50% of the variance in the original dataset). Intuitively, we now have a world where each month is characterized by eight factors.
As for the latter, we experiment with the k-nearest neighbor technique, a simple machine learning algorithm designed to group and classify data. Using this approach, we can mathematically compute the distance for a given month from every historical snapshot. The months with the shortest distance should be the best historical parallel.
Got that? Great.
Ok, so as BofAML goes on to explain, when you find “the top 20 historical parallels for each of the last three months, the periods that came up more often were months in ‘93-’94, ’96-97 and ’13-‘14.” Here’s the visual on that:
Apparently, previous hiking cycles aren’t in fact the best analogs when it comes to finding reference points for the current environment.
As the bank puts it after consulting with Skynet (which is mercifully not yet self-aware and thus still receptive to consultations as opposed to simply dictating orders to human slaves after becoming goal oriented), “in contrast to consensus, algos don’t think we are in a environment similar to the past two hiking cycles.”
BofAML goes on to say that the episodes identified by the algo were associated with “relatively strong economic growth” and thus we can conclude that “big data says the US economy is doing just fine.”
What about the curve? Well, the algos have something to “say” about that too, and here’s what it is:
Given the diverging views on the US yield curve among investors, we apply this classification tool and try to assess the most likely outcomes in the following month. For the purpose of this analysis, we look at the most likely directional move of 5y-30y curve in the months following the historical “neighbors”. With the information we have in April 2018 and its “neighbors” from the past, history suggests the curve is likely to continue to flatten with an average -3.4bp move.
There you go. That’s what the Terminators say.
One wonders whether the machines BofAML “talked” to for this exercise consulted with their market-making algo “friends” and/or with the bevy of algos that underpin the various systematic strategies that helped catalyze the February rout.
Take from this what you will and do note that BofAML readily admits the obvious which is that when it comes to this approach, “there is plenty of work to be done.”
Ignore the large laser cannon, he’s just here to help!