Models Can't Think, But Humans Overthink

Mish

I have some important observations on the strength and weaknesses of models.

Reflections on the Election and GDPNow

I discussed the fact that models cannot think this morning in Is Nate Silver Underestimating Democrat Senate Chances in Georgia?

It occured to me that I made similar statements previously about Pat Higgins' GDPNow model. 

Here is a snip of my Georgia discussion. Following that, let's look at the flip side of the coin.

Random Noise

Silver just flipped again on Iowa, and has been going back and forth for a week. Iowa is again in Trump's column.

I solidly disagree and may revisit Iowa soon. But now Silver has Ohio in the Biden camp.

When you change your mind every day, or even sometimes twice a day you are changing your mind on noise.

That's what we have: noise. 

Models Can't Think!

In reference to the Georgia Senate election, I also said Models cannot and do not do what I just did. That is take a look at a debate in which Perdue was clearly slaughtered, then make a judgement call on it.

Nor do they properly look at momentum.

I have the same criticism of Pat Higgins' GDPNow economic model. It cannot think about "why" or make value judgements on what is happening. 

When weather is a factor, GDPNow cannot make adjustments. When Covid hit, the numbers were so screwy the GDPNow model was programmed to discard the numbers as outliers that could not possibly be correct.

Models only do what they are programmed to do, typically look at data and historic trends.

There are clear errors being a slave to models, but there is a flip side.

Humans Overthink!

The flip side of the coin, and it's an important one is that humans overthink.

The lead image is from Psychologists Explain How To Stop Overthinking Everything

I have a huge collection of all the amusing things, people have said would matter in this election.

People believe all kinds of things matter that won't matter and in fact cannot possibly matter.

I will post the collection after the election, and some of them will have you laughing.

Beauty of Models 

The beauty of a model like Nate Silver's and Pat Higgins' GDPNow Models is they don't overthink.

Silver waits for the polls, GDPNow waits for the economic data.

I am very happy for Silver's forecasts and I find them helpful for my own thinking. The same applies to GDPNow.

It's easy to poke fun at models when they are wrong (and even when they aren't).

I think Nate Silver is brilliant. I feel the same way about Pat Higgins. 

But as I explained to a couple of friends recently, I add no value if all I do is parrot the calls of others. 

What's Happening and a Look Back at 2016

Mish

Comments (53)
No. 1-14
Eddie_T
Eddie_T

Overthink?.....Moi? :)

Rocky Raccoon
Rocky Raccoon

Donald Trump created a lot of economic damage in Iowa as crops rotted in his ill-advised trade war. Iowa has long been known as an independent state that goes back and forth all the time every four years. I just don't see how Iowa is rallying behind Trump as millions of dollars that was supposed to come into the state from China was lost.

silverdog148
silverdog148

Yes we overthink things no doubt, on Trump he has done some major damage to the standard GOP platform as well as the Dem platform, in fact it can be argued he inadvertently dealt a huge blow to the entire U.S political system that would under normal circumstances have taken a generation, so in that we should be thankful.
The playing field past this election is going to look very different. Washington D.C has been exposed inadvertently and they are now facing a real loss of confidence and even worse have lost any type of moral high ground, prepare for a very bumpy ride as the citizenry now demands Bread and Circuses now that the charade has been exposed and they want in on the action also.

Some observations:
The Age of Taxes is over: Any tax proposal is DOA at this point from both bases including the democrats base, it's simply not needed, the reserve currency can be leveraged via deficits, so any pain is unneeded, no new significant taxes at the national level until the collapse at least 3 generations if not more away. This knocks out a major party plank for the dems and they will face growing calls for real socialism in the next 10-20 years, which very few Dem politicians/Wall Street really want.

The Age of Austerity is over: The GOP plank of austerity/low deficits conversely is done even for their base, very few will support austerity, it's simply not needed, the reserve currency can be leveraged via deficits, so any pain is unneeded, no new significant austerity at the national level until the collapse at least 3 generations if not more away. This knocks out a major party plank for the GOP and they will face growing calls for real socialism in the next 10-20 years, which very few Republican politicians/Wall Street really want.

The Age of innocence is over:

Sechel
Sechel

Sounds like you want your model to react to momentum and volatility. That could be done but then it would no longer be parsimonious

anoop
anoop

any thoughts on supermodels?

Doug78
Doug78

The problem with models for presidential elections today is that the data is highly suspect. Ten of fifteen years ago a pollster had around a 50% response rate. Today it's around 5%. That skews the data too much to be normalized by mathematical methods so it's garbage in, garbage out. If you read studies in reviews dedicated to polling you see that they are well aware that it is the biggest problem and it's not getting better. I should mention that the polls seem to be tightening and the difference is narrowing. Soon the spread will be tight enough so that the pollsters can claim that they were in the range of statistical error giving themselves an out for being wrong and thereby saving their reputations. It's a common behaviour among the white-collar variation of the human race. I believe it's called "covering your ass".

njbr
njbr

TimeToTest
TimeToTest

Trump is going to win again. I was firmly in the Biden camp but I guess I over thunk it.

Trump will firmly win this thing on Election Day. I think emotion is playing into the “models” now. A model doesn’t think but bad data will make them worthless.

Biden has to much baggage. Very similar to Hillary. Are the democrats about to make the same mistake twice? Anyone in the primary against Biden could have beat Trump. Well except his running mate. Talk about the bottom of the barrel.

Tengen
Tengen

Overthinking in American politics is much more rampant than people seem to realize and it's a real problem.

The whole QAnon movement is an exercise in extreme overthink, where people feverishly try to decipher cryptic messages that don't mean anything, they're just gibberish. I suppose when people feel desperate or threatened they're more likely to perform mental gymnastics to reach a conclusion they like.

No matter how this election goes (I expect a narrow Biden win) a lot of people are going to look crazy, with wild predictions that didn't remotely reflect reality. Many people have been wrong in US elections before, but not THIS wrong.

Six000mileyear
Six000mileyear

The best model has been the stock market. Trump has a slight edge, but this week's sell-off has reduced his lead. Markets up on Monday and Tuesday favor Trump. Markets sideways to slightly down would be a coin toss. Heavy selling favors Biden.

AshH
AshH

Excellent post, Mish. Thanks.

Realist
Realist

These cases and deaths from the virus are not a model.

From Worldometer today:

Oct 26 M 69,844:529

Oct 27 T 75,980:1055

Oct 28 W 81,585:1030

Oct 29 T 91,829:1041

Oct 30 F 101,461:988 (so far)

AussiePete
AussiePete

I subscribe to Jim Rickards "Strategic Intelligence" investment newsletter and he is predicting a Trump win, as he did in 2016. He describes his model as "based on a combination of complexity theory, Bayes’ Rule, behavioural psychology, history, and proprietary applied mathematics arrayed in cognitive maps designed using subject matter expertise and natural language processing" - (whatever that means!) He predicts that Trump will win Florida, Georgia, Arizona, Nevada, North Carolina, Pennsylvania, New Hampshire, Iowa, and Michigan. He says that Trump also has a chance of winning Wisconsin, Colorado and even Minnesota.

He claims that current polls are flawed and these are some of the reasons...

• Polls oversample Democrats.
• Polls oversample middle-class and upper-income residential zip codes.
• Polls do not capture the ‘shy’ Trump voter.
• Polls are snapshots, not forecasts and don't take into account trends
• Pollsters rely too much on phone calls
• Turnout models are flawed, and don't take into account the "enthusiasm gap"
• National polls don’t matter and should be ignored as there are 50 state elections
• Many polls look at ‘all adults’ or ‘registered voters’ instead of "likely voters"
• Small polls should be ignored - he suggests a 1,300 minimum sample size
• Increased support for Trump among African American men and Hispanics is not being captured

After reverse engineering the flawed polls, he says that it is reasonable to add 2.5 percentage points to the claimed likely Trump vote

Rickards is also predicting a possible protracted legal and political struggle following the actual voting

He goes into much more detail but its paid-for content....

.


Global Economics

FEATURED
COMMUNITY