Models Vs Humans: Models Can’t Think, But People Think About The Wrong Things

Data from the BLS, chart by Mish, through October.

My recession model is based on 15+ weeks of continued claims.

I discuss my recession model at the end of this post but first let’s continue with a model of sentiment vs the economy in my previous post.


The Paradox of Bad Sentiment

Please consider The Paradox Between the Macroeconomy and Household Sentiment by the Brookings Institute.

Here’s what happened. The Brookings Institute has a model that looks at GDP, CEO sentiment, the stock market, and even airplane traffic.

On the basis of those factors, Brookings concluded that Consumer Sentiment should be higher than what it is.

My hoot of the day is that it’s obvious why consumer sentiment is sour. My second hoot is that despite the obvious, Brookings could not figure it out.

Bias gets in the way.


The Brookings Institute Wonders Why Consumer Sentiment is So Bad, I Can Help

I gave 16 reasons, admittedly some overlapping, in my post The Brookings Institute Wonders Why Consumer Sentiment is So Bad, I Can Help

Brookings trained its model on GDP, not employment. Full Time Employment is -1,000,600 from a year ago! Excluding government, year-over-year employment is negative for the last 9 consecutive months.

Brookings did not ponder a 0.7 percent rise in the unemployment rate despite when it’s the direction that matters, not the level.

Brookings believed bogus stats that say crime is down.

Evictions are at record highs in many states and might be everywhere were it not for eviction moratoriums.

I added a couple new points this morning. Here is one of them: Tens of millions of people want to buy a home but can’t afford one and a different set of tens of millions of people are trapped in their homes but won’t because of mortgage rates.

I also added this: Even if you have a home, what about flood insurance, fire insurance, and car insurance.

And as I noted originally, but failed to list in my bullet point synopsis (now added), a Bank of America survey shows over 40 percent of the nation is living paycheck to paycheck.

If the Mises Institute sees this they will be laughing out loud.

The Brookings errors here are twofold. First models can’t think. Second, Brookings failed to go beyond pat answers given by Paul Krugman (the election and Republican bias), Greg Ip at the Wall Street Journal, and others.

In this case, models can’t think but neither could humans. And of course humans create models.

Brookings had a bad model of sentiment that just happened to correlate nicely with the economy until it didn’t.


Atlanta Fed GDPNow

Recently, I have been amazed at the accuracy of GDPNow. Kudos to Pat Higgins at the Atlanta Fed.

But when GDPNow was first published, most of us mocked it. Misses were widespread, in random directions.

Random errors are a good thing actually.

I am not sure if Higgins tweaked GDPNow much over time, but the last 8 quarters of GDPNow have been outstanding.

I have discussed this before, and mentioned it on October 30, 2024 in Real GDP Increased 2.8 Percent in the Advance 2024 Q3 Estimate

The BEA reported 2.8 percent GDP with Real Final Sales of 3.0 percent. Higgins’ forecast was 2.8 percent GDP with Real Final Sales of 2.9 percent. Wow.

Models don’t think. And that’s a good thing (as long as the model is a good one).

I reflect back on one of the GDPNow model forecasts at the beginning of the Covid Pandemic. The numbers were so horrific Higgins thought the model was wrong.

The above is from memory, so I might not be stating this accurately, but the model was fine. What appeared to be massive outliers were in fact accurate, as we soon found out.

I have no edge over the final forecast of GDPNow. Unless you have a better model than GDPNow, then you don’t either.


Humans Can Think, That’s Good and Bad

My explanation of what went wrong with Brookings is an example of human intelligence (or lack thereof) vs a model.

However, I too have my fair share of misses that good models don’t have.

For example, a huge number of us forecast a recession in 2023. I was very confident of this. So were others. I recall a Bloomberg model that had the odds close to 100 percent within a year.

What happened?

I have written about this before, but not a single model factored in the right things.

Amusingly, many economists still do not know what happened. But I do know why all these models (and my seat of the pants non-model) went wrong.


The 2023 Non-Recession


Q: What happened to stave off recession in 2022Q4-2023Q1?
A: A huge tax cut coupled with big jumps in minimum wages in 27 states!

The IRS adjusted tax brackets based off inflation and that put a sudden and huge amount of extra income into people’s pockets every month.

In addition, 27 states raised minimum wages starting January. Disposable personal income jumped a whopping 2 percentage points in a single month!

GDPNow did not predict that outcome. Nor did any humans that I am aware of.

However, the GDPNow model reacted to the data. I initially didn’t, nor did any of the recession models.

So, shame on us. An unthinking model beat the pants off those of us who thought about the wrong things or ignored the right things.

Ignoring the right things (not even seeing them), is what happened to Brookings.


Understanding Bias

All of us are biased in some ways. Models may or may not be. One attribute of a good model is that it is unbiased.

Unbiased does not imply accuracy. So a model that’s totally and randomly wild is not a good thing although lack of bias is.

People accuse Nate Silver of being biased. So have I. He wants Harris to win, and that comes through. He has even admitted as such.

Importantly, bias, if present, is in Nate Silver the person, not his model. Silver’s model is not biased in any predictable way. It reacts in mostly predictable ways to polling data.

This was a subject of a recent debate with a few friends of mine. One of them suggested Selzer, a prominent pollster, should throw away a sample that had Harris winning Iowa by three percentage points.

But good models don’t do that. Imagine GDPNow throwing away its reaction to disposable personal income when it jumped an amazing 2 percentage points in a single month.

Of course, disposable income can be verified while the odds of Harris winning Iowa can’t.

My friend still believes Selzer should not have published those results. You get into trouble throwing away results just because you don’t believe them.

I have biases too. I tend to be pessimistic on the economy. This is why I did not immediately change my mind on a recession on the jump in income killing a recession. Guilty as charged.

But I did want to know where and why I went wrong, so I went back later and created the Change in Personal Taxes chart above. Coupled with huge minimum wages hikes a pending recession that many of us predicted did not happen.

Amusingly, those who said no recession all along and brag about it were mostly lucky. Only those who said “no recession due to tax cuts” truly got it right.

I am unaware of anyone who got it right, because without that tax cut and extra income we were headed for recession.

So if you are bragging about your “no recession” call you may wish to reconsider how smart or lucky you were.


Recession Model

I never had a recession model until now. I had been typically persuaded by prominent people calling for a recession and my typical reaction was to go along.

But not this last time. When Danielle DiMartino Booth repeatedly forecast recession stating in October of 2023 based off a McKelvey recession indicator, I repeatedly stated that I thought she jumped the gun.

In May of 2024, I went along with the call based with on a clear and persistent change in the McKelvey indicator.


What is the McKelvey Recession Indicator?

Take the current value of the 3-month unemployment rate average, subtract the 12-month low, and if the difference is 0.30 percentage point or more, then a recession has started.

Edward McKelvey, a senior economist at Goldman Sachs, created the indicator. The problem with the indicator is that it has many false positives.


Claudi Sahm Modification

Claudi Samn, a former Fed economist, revised the rule, claiming it as her own, without credit to McKelvey, then set the indicator to 0.50.

Because that still had false positives, she started her series in 1960.


McKelvey Improvements

Improvements to the model by Pascal Michaillat and Emmanuel Saez, economists at the University of California in Santa Cruz, eliminate all false negatives and all false positive dating to 1951.

I discussed McKelvey many times including a calculation of recession odds on September 10, 2024 in The McKelvey Recession Indicator Triggered, But What Are the Odds?

On October 8, I commented McKelvey-PMES Recession Indicator Weakens Slightly but Signal Still Firm

Since then, the PMES signal weakened a bit further still, but it is still above my preferred 0.40 trigger vs 0.30 McKelvey.


My Own Recession Indicator

Based off the McKelvey and Michaillat formulas, I came up with my own recession indicator.

Please consider my October 11 post Continued Plus Long-Term Unemployment Claims Suggest Recession Right Now

I created a new recession indicator based on claims. Hurricanes have nothing to do with this.

The trigger weakened slightly in October but is still deep in the range where 100 percent of the time, dating to 1951, that the economy has been in recession under current conditions.

My trigger is based off the lead chart which is current. The slight weakening is due to a particularly bad month my model now views as stale despite the fact the key number made a new high as shown.

I do not know how to accurately assess odds because these models don’t decide recessions. Humans decide. Specifically, the NBER decides elections based on subjective reasoning, after all the revisions come in.

In one instance, the NBER decided a recession after it was already over. The huge NBER lag is why people have developed models.

Meanwhile, negative revisions are mostly the norm.

Rather than let my bias get in the way, I am now basing recession calls off McKelvey-PMES and my own trigger. They say the same thing. Recession is here or likely.

Perhaps this is a Brookings-style or disposable income type of miss, but I think not. The data goes back to 1951 with no false positives or misses.

Call it 50-50 if you like. This can go on for a while.

Anyway, that’s my model and it isn’t biased. If it suddenly switches to no recession, I will report so.

If you disagree with my recession model or the McKelvey-PMES model, both of which have a perfect track record on recessions dating to 1951, can I ask: What’s the track record of your model?


More By This Author:

The Brookings Institute Wonders Why Consumer Sentiment is So Bad, I Can Help
No Matter Who Wins The Election, Expect The Cost Of Microchips To Soar
Excluding Government, Year-Over-Year Employment Is Negative 9 Straight Months

Disclaimer: The content on Mish's Global Economic Trend Analysis site is provided as general information only and should not be taken as investment advice. All site content, including ...

more
How did you like this article? Let us know so we can better customize your reading experience.

Comments

Leave a comment to automatically be entered into our contest to win a free Echo Show.
Or Sign in with