A reasonable assumption about elections is that when petrol gets dearer and the weekly shop costs more, the governing party suffers. It is the oldest idea in election forecasting, and this site's own model leans on a version of it: the fundamentals term that carries a race nobody has polled is built partly from presidential approval and an index of economic conditions.

So it seemed worth checking. We took twenty economic series — crude oil, retail petrol and diesel, headline and core inflation, unemployment, jobless claims, payrolls, real disposable income, the S&P 500, the VIX, the dollar, mortgage rates, breakeven inflation, consumer sentiment — and tested them against 2,130 individual polls published between December 2024 and August 2026. Two dependent variables: how Americans rate the President, and which party they say they will vote for in November.

The answer splits cleanly, and not where most people would expect.

What moves approval

Economic conditions move presidential approval, at sizes that are modest, stable and consistent with fifty years of published research.

A ten-cent rise in the price of a gallon of petrol costs roughly a quarter of an approval point. A percentage point of extra inflation costs about two. Half a point on the unemployment rate costs around one and a half. Higher consumer sentiment and a rising stock market help, by smaller amounts.

Every one of those signs is the direction theory predicts, and the petrol estimate survives dropping any single month from the sample. It also lines up with the literature: Krosnick and colleagues, working on data from 1976 to 2007, found about six tenths of a point per ten cents. Ours is roughly 40% of that, which is what the recent work on polarisation would lead you to expect — the same shock moves a more entrenched electorate less.

What does not move the congressional vote

The same conditions, tested against the generic congressional ballot, produce effects four to eight times smaller, and mostly indistinguishable from zero.

This is not a quirk of our window. Kramer's 1971 paper, the founding study in this literature, found real income growth mattered for congressional voting while inflation and unemployment did not. A 2022 study in PNAS found that voters who switched their congressional vote that year did not do so in response to inflation; abortion drove the switching. We reproduce both nulls on 2025–26 data.

The mechanism the PNAS authors identify is the part worth dwelling on. Attributions of blame for inflation were, in their phrase, "starkly partisan or completely nonpartisan." People experienced the same prices and assigned responsibility according to what they already believed. Economic perception is downstream of political identity rather than an independent input to it.

That explains the split we measure. Approval is a question about the President, and it moves. The ballot is a question about which side you are on, and it does not.

The part that should make you cautious about all of it

Two further findings temper everything above.

The relationship is contemporaneous, not predictive. When we fit the change in conditions over the period before the change in vote intention — which is what forecasting would require — almost nothing survives. Petrol and approval move together within the same few weeks; neither leads the other. The economy is useful for understanding where approval sits, and close to useless for guessing where it goes next.

Almost nothing moves at all. Across twenty months, the total standard deviation of the generic ballot is 3.2 points. Which polling firm asked the question explains more of the variation in a single poll than every economic factor combined. In an electorate this settled, there is very little movement left for anything to explain, and a real effect and a negligible one look much the same.

We also tested whether the economy reaches the congressional vote indirectly, by moving approval which then moves the ballot. That is the classic reading of midterms as a referendum on the President. It appeared to work, and then did not survive its own robustness check: the approval-to-ballot link is strong when both are measured in the same survey and vanishes when they are measured a month apart on different samples. Approval and vote intention are not cause and effect. They are two ways of asking the same person which side they are on.

What this means for reading the forecast

Three practical things.

When the national environment on this site moves, the cause is almost always new polling rather than new economic data. The "what moved the forecast" panel on the front page is computed by re-running the model with recent polls held out, and it attributes movement to specific surveys for exactly this reason.

When you see a story arguing that a change in petrol prices or an inflation print will reshape the midterms, treat the claim about approval as plausible and the claim about seats as unsupported by this data.

And treat all of it as provisional. We observe one administration over twenty months, which means every slowly-moving economic series is nearly collinear with the calendar itself. We can distinguish "the economy did nothing" from "the economy did something large", and we cannot reliably distinguish it from "the economy did something small". Extending the analysis across earlier cycles is the fix, and it is the obvious next step.

The code, the data pull and every regression are in the public repository under research/.