The Number That Predicts Nothing

At some point today, a cable news chyron will display a number. It will be the president’s approval rating, updated to the decimal point, presented as though it were a vital sign on a hospital monitor. A pundit will interpret it. A headline will frame it. A donor will adjust a contribution based on it. And none of them will pause to ask whether the number they are reacting to has any meaningful connection to the political outcome they care about. For most of the uses to which approval ratings are put, it does not.

The Correlation That Sounds Better Than It Is

The standard claim is that presidential approval predicts midterm seat losses. The data, at first glance, seems to support this. The American Presidency Project at UCSB has tracked the relationship since 1936 and reports a correlation of r=0.61 between presidential approval and House seat changes, with r=0.45 for the Senate.

Those numbers sound respectable until you do the arithmetic. A correlation of 0.61 means approval explains about 37 percent of the variance in House seat changes. The remaining 63 percent is driven by something else entirely. For the Senate, approval explains roughly 20 percent of the variance, leaving 80 percent unexplained. These are not forecasting tools. They are rough associations that leave most of the outcome unaccounted for.

The linear model derived from Gallup data makes the weakness concrete. At 40 percent approval, roughly where Trump sits now, the model predicts a loss of about 43 House seats, but with a standard deviation of 22. The 95 percent confidence interval stretches from losing 87 seats to gaining 1. A range that includes both a historic wipeout and a modest gain is not a prediction in any useful sense of the word.

The weakness runs deeper below the 50 percent line. When researchers tested the model separately for approval ratings below 50 percent, it failed its own statistical utility test. The relationship between approval and seat loss is statistically useful only above 50 percent, where the correlation is driven by a handful of popular presidents whose parties performed well in midterms. Below 50 percent, which is where every modern president spends most of their time and where approval ratings get cited most urgently, the historical relationship breaks down into noise.

The Counterexamples Are Not Edge Cases

If approval ratings were reliable predictors, the exceptions would be rare and explicable, but they are neither.

In 2022, Biden sat at roughly 42-44 percent approval heading into the midterms. The historical models predicted losses of around 30 House seats and 3 Senate seats. Democrats lost 9 House seats and gained a Senate seat. The model was off by a factor of three in the House and wrong on direction in the Senate. Every commentary piece written in mid-2022 citing Biden’s approval as a predictor of a “red wave” was proven wrong, and not by a small margin.

In 2016, Trump was elected president despite holding the highest unfavorability rating of any major party candidate in polling history, at 61 percent unfavorable. Favorability polls said he could not win. He won. His opponent also had historically high unfavorability, which illustrates a deeper problem: favorability is a relative measure being used as an absolute one.

George H.W. Bush reached 89-90 percent approval after the Gulf War in 1991. Eighteen months later he lost to Bill Clinton. Even Nate Silver has acknowledged that Bush’s Gulf War bounce distorted the timeline of when approval becomes predictive of reelection, which is a polite way of saying approval can move 40 or more points in a year and a half, making any snapshot a photograph of weather, not climate.

Why Approval Has Become a Constant

These failures reflect a structural change in what approval ratings measure, and a numerical illusion in how they are compared to election results.

Start with the comparison. Trump’s approval sits at roughly 35 percent of adults. Commentators compare this to his 2024 vote share of roughly 50 percent and conclude he has lost 15 points of support. But these numbers have different denominators. Approval polls sample all adults. Vote share is a percentage of the roughly 62 percent of adults who actually voted. Trump received about 50 percent of that 62 percent, meaning approximately 31 percent of all adults cast a ballot for him. His 35 percent approval rating is, by the measure that matches the poll’s sample frame, higher than his actual adult-population vote share. The narrative of eroding support depends on subtracting a percentage of voters from a percentage of adults, which is arithmetically meaningless.

This matters because the entire “he’s lost his base” framework rests on the comparison. When a commenter or pundit points to 35 percent approval and says Trump has shed the new voters who put him over the top in 2024, they are performing a subtraction across incompatible units. The approval number is not telling them what they think it is telling them. But this isn’t about Trump, there is a deeper change afoot.

The deeper structural issue is that approval has become inelastic. In the 1950s through the 1980s, it moved with events, with economic conditions, with crisis performance. Eisenhower, Kennedy, Johnson, Nixon, Carter, and Reagan all saw approval ranges of 30 to 40 or more points across their terms. The electorate was less sorted, more voters were persuadable on the merits, and approval captured something resembling an evaluation of performance.

That elasticity is gone. Since the mid-1990s, and accelerating through the Trump era, approval ratings have locked into a narrow band determined almost entirely by partisan composition. Trump’s first term was the most stable approval rating in Gallup’s tracking history; he barely moved from the low-to-mid 40s across four years regardless of events. Obama was similar, ranging only about 10 points across eight years. Biden barely moved. Roughly 40 percent of adults will approve and roughly 50 percent will disapprove of any president, regardless of performance, because approval has become a proxy for party identification. The remaining 10 percent in the middle is responsive to conditions, but their movement is swamped in the topline by the partisan anchors on both sides. The daily-updating approval tracker that Silver and others maintain is mostly tracking a constant with noise, and treating every half-point fluctuation as a story.

The Legislative Connection Is Even Weaker

If approval had limited value for predicting elections, it might still matter for governing. The common assumption is that a popular president can push legislation through Congress and an unpopular one cannot. The academic research does not support this.

Canes-Wrone and de Marchi, in the most rigorous study of the question, found that approval affects legislative outcomes only on high-salience, low-complexity issues, essentially easy votes where members of Congress worry about being on the wrong side of a popular president. On complex legislation, technical bills, or anything where the member’s constituency has a distinct interest, presidential approval has no independent effect on vote outcomes.

This makes intuitive sense if you think about how a member of Congress actually decides. They vote based on what their donors want, what their district wants, what leadership is pushing, what the bill does to their constituents, and whether a primary challenger could use the vote against them. Presidential approval enters this calculation only as a vague background signal about whether association with the president is an asset or a liability, and even then it is filtered through the member’s own read of their own district, not the national number.

The current Iran situation provides a clean test. The War Powers resolution vote went 53-47 along party lines in the Senate. Trump’s approval is at roughly 35 percent. If approval drove legislative behavior, you would expect some Republicans in competitive states to break with Trump on a politically toxic war vote. Only Rand Paul, whose brand is non-interventionism, defected. Every other Republican voted with the president despite his historically low approval, because their calculation had nothing to do with a national polling number and everything to do with their own primary electorate and donor base.

What Actually Predicts Things

If approval ratings are weak predictors, what works better? Recent research by Algara et al. offers an answer that is both more useful and more humbling. Using 80 years of high-frequency data, they found that congressional election outcomes are predicted by party brand differentials (measured by the generic ballot) while presidential elections are predicted by presidential approval combined with candidate polling differentials heading into election day. Crucially, these three variables, approval, generic ballot, and candidate differential, are weakly correlated with each other, meaning they measure distinct things that cannot be substituted for one another.

For midterm prediction specifically, the generic ballot measured close to the election is the single best predictor of House seat changes. Consumer confidence, especially the Michigan Consumer Sentiment Index, predicts presidential reelection outcomes better than approval in several models. And actual voting behavior, such as special election results and primary turnout, tells you more about the current political environment than any opinion survey, because it measures what people do rather than what they say.

The most useful forecasting models combine multiple inputs. An NBER model using real GDP per capita growth, excess inflation, and approval ratings accurately predicted 15 of 17 presidential election outcomes, but approval alone missed several. The approval rating contributes to the prediction; it simply cannot carry it alone, and the other variables in the model are doing most of the work.

What Approval Ratings Actually Are

Strip away the forecasting pretensions and ask what approval ratings functionally do in the political ecosystem.

They give cable news a number to put on screen. They give pundits a hook for commentary. They give political donors a signal to calibrate contributions. They give the White House communications shop something to manage. They give the opposition a rhetorical weapon or, when the numbers are high, something to explain away. And they give the polling industry a recurring revenue stream from media clients who will always pay for a new number regardless of whether that number means anything.

What they do not reliably do is predict elections, forecast legislative outcomes, or measure the thing they claim to measure, which is how the public evaluates the president’s performance. In an era of rigid partisan sorting, they measure party identification with a thin veneer of performance evaluation on top, and the veneer gets thinner with every election cycle.

The next time a headline announces that the president’s approval has hit a “new low” or “ticked up two points,” consider what that information actually enables you to do. Can you predict the midterms from it? The research says no, at least not from approval alone, and not at these levels. Can you predict legislative outcomes? The research says no, except on easy votes. Can you even be confident the number will hold for more than a few weeks? Bush’s 40-point drop from 1991 to 1992 says no.

What you can do is nod along with the commentary it generates, which will tell you what you already knew: that the president’s opponents disapprove and his supporters approve, in proportions that have barely moved since inauguration day. That is what the number describes. Everything else built on top of it is conjecture dressed in decimal points.

Historical approval data from the American Presidency Project (UCSB), Gallup, and Wikipedia’s compilation of Gallup tracking. Forecasting models from Algara et al. 2025 and Gordon 2024 (NBER). Legislative analysis from Canes-Wrone and de Marchi 2002. Midterm prediction analysis from Niskanen Center interview with Algara, March 2026.