Nineteen for nineteen. That is the claim. Every midterm election year since 1950, the S&P 500 delivered a positive return in the twelve months following the vote. Average gain, roughly 14%. A widely circulated analysis bolted a headline onto it — 95% probability — and the financial press repeated the number without asking what it was measuring.
I pulled the arithmetic apart last week. It does not close. Not because anyone fabricated returns, but because the sample definition moves under your feet and nobody published the definition.
Nineteen observations. One headline. Zero confidence intervals. That is the entire evidentiary base for a claim currently shaping positioning ahead of a November 3 vote.
I don't trade equities. I trade crypto. But the failure mode here is the exact failure mode I have spent nine years auditing in on-chain data, and it is about to be recycled for the fourth Bitcoin halving narrative. So it is worth doing the forensic work in public.
What the study actually says
The artifact is a 2026 midterm preview. The date anchor is unambiguous: November 3 is a Tuesday in 2026, and the only cycle years that match are 2018 (Nov 6), 2022 (Nov 8), 2026 (Nov 3). 2010 and 2014 don't fit. So this is event-driven, forward-looking content with a hard expiry — everything in it goes stale on election night.
The core claims: a 95% positive rate since 1938. Nineteen of nineteen since 1950. A 14% average twelve-month return after the vote. A 5% average return in year two of the presidential term. An average drawdown near 9%, described as half the historical norm.
Now run the arithmetic. 1938 through 2022 contains 22 midterm elections. A 95% rate implies 21 of 22. But the same article notes that 1938 was down 5% and 1946 was flat. If flat does not count as up, then at most 20 of 22 rose — 90.9%. That contradicts the headline.
There is a reconciliation: start the sample in 1942 instead of 1938. You get 21 elections, 1946 as the sole non-positive, and 20/21 = 95.2%. The number survives — but only if the stated start date is wrong. That is the first signal. The headline says 1938; the math says 1942. Not fraud. Window selection, undisclosed.
The second tension is more interesting. A 5% average in year two and a 14% average after the election cannot both be right unless the windows overlap in a specific way. The election happens in November of year two. The following twelve months are two months of year two plus ten months of year three. Weight it: (2/12 × 5%) + (10/12 × 15%) ≈ 13.3%. That lands on 14%.
The study's own figures confess the answer. The "midterm effect" is the year-three effect wearing a political costume.
The identification problem
Midterm elections are not a random variable. They are a deterministic function of the calendar — always year two of a presidential term. The twelve months that follow are 83% year three. So a midterm dummy and a year-in-cycle dummy for year three correlate above 0.9 across any sample built from 1938 onward.
When two regressors move together, the regression cannot separate their coefficients. The claim that midterms cause the rally is not merely unsupported. It is unidentifiable from this data. You cannot run the counterfactual because there are no non-midterm year-two observations. None exist. The design has no control group.
This is not a subtle econometric footnote. It is the first thing you check. When an explanatory variable is a calendar identity rather than a random draw, the coefficient is a restatement, not a cause.
Apply the same test to Bitcoin. Every halving is followed by a year that is definitionally the year after the halving. "Post-halving years are bullish" and "the twelve-to-eighteen months following a supply-shock event are bullish" are the same sentence in different hats. Four observations. Correlated dummies. No control group. Same disease, smaller sample.
The base-rate check nobody bothers to run
A base rate is the price of admission for any frequency claim. US equities in a randomly selected rolling twelve-month window since 1938 rise roughly 65–70% of the time. Take 65% as the conservative figure.
Probability of nineteen consecutive wins at a 65% base rate: 0.65^19 ≈ 0.03%. At 75%: 0.4%. At 80%: 1.4%.
Two explanations exist. Either the true conditional win rate exceeds 90%, or the windows were not randomly selected.
I know which one I would bet on. A 19-for-19 record is not evidence of edge. It is evidence that the sample was drawn from the strong part of the cycle and reported as though it were drawn from the whole. The perfect score is the warning label, not the selling point.
This is where retail readers get flipped. They see 19/19 and read certainty. The statistically literate see 19/19 and read a hand-picked window. Same number, opposite priors, and only one of them is doing inference.
The window is undefined, and the conclusion is window-dependent
The article never states its measurement rule. Election day to election day? Calendar year? December to December? Price return or total return?
Total return adds roughly two percentage points a year from dividends. On a 15% figure that is 15% versus 17% — not trivial when you are comparing against a threshold.
Then the fragile data point. The article cites 1986 as the weakest post-election year at +1%. October 1987 — Black Monday, a one-day drawdown near 20% — sits at the tail of a 1986-to-1987 election window. Shift the start by one month and 1986 becomes catastrophic. Shift it the other way and it becomes strongly positive. A single observation that flips sign under a one-month translation means the "weakest year" narrative is an artifact of an unreported choice.
I have run this audit before. In 2017 I spent six months manually scraping Ethereum block data across 45 ICO projects, reconciling whitepaper distribution schedules against actual token movements. Three projects carried inflation discrepancies around 40%. The teams were not lying about the numbers. They were lying about the denominator — which wallets counted, which vesting cliffs counted, which supply figure got called circulating. Change the denominator and any schedule becomes whatever you need it to be.
The midterm study has the same structure. No denominator. No link to primary data. No reproducible calculation.
Frequency is not probability
The title says 95% probability. The body delivers a sample frequency.
These are different objects. A frequency describes what happened across a set of realized observations. A probability is a claim about future states conditional on current information. Converting one into the other requires an assumption — usually stationarity — that nobody stated and nobody tested.
The 95% figure arrives without a confidence interval, without a p-value against the base rate, and without any adjustment for valuation or rate regime. A binomial interval on 19/19 against a 65% base rate is wide enough to make the headline meaningless. Nineteen observations cannot carry a probability claim. They can barely carry a frequency claim.
Drawdown misused as a quality metric
One data point claims the average post-midterm drawdown is near 9%, roughly half the historical average, and treats the smaller number as evidence of a stronger regime.
That is backwards. Realized drawdown measures volatility that has already occurred, not return that should be expected next. Low-volatility regimes revert. If anything, a compressed maximum drawdown raises forward risk, because the volatility was suppressed rather than eliminated. Presenting a smaller drawdown as a bullish quality signal is a category error — it substitutes a risk statistic for a return statistic and trusts that nobody checks.
Also unstated: is 9% the average of annual maximum drawdowns, or the average of discrete drawdown events? The gap can run several percentage points. If 9% is genuinely half of history, the implied historical average is 18%, and I want that number sourced.
Risk stress-test before you act on any of this
If I were sizing a position on the calendar, I would build three checks first.
One: recompute the win rate under three window definitions and translate each by ±3 and ±6 months. If the sign flips, the signal is dead.
Two: compare the conditional win rate against the unconditional rolling twelve-month base rate. The excess, not the raw number, is the only quantity with information content.
Three: stratify by starting valuation and by real rate regime. A 2026 setup with an expensive index and a restrictive policy backdrop is not comparable to 1986. Without that stratification, the prior is borrowed from conditions that no longer hold.
None of the three checks appear in the source material. That is the gap between a statistic and a strategy.
Where this becomes crypto's problem
I flagged this piece in my own research stack as non-crypto content. The article contains zero policy substance — no legislative agenda, no fiscal stance, no monetary discussion. Labeling it macro-policy is generous. It is market statistics wearing a political skin. I am importing the methodology, not the claim.
And that methodology is about to be pointed, for the fourth time, at the halving.
The 2024 halving cut issuance from 6.25 to 3.125 BTC per block. The consensus frame has been: supply shock, then twelve to eighteen months later, a parabolic move. Four data points since 2012. Four.
The 2024 cycle is also the first in which the marginal buyer changed identity. Spot ETFs moved institutional flow into a market where the supply shock is mechanically trivial against secondary volume. At $100k BTC, 450 BTC of daily issuance is roughly $45M against a multi-billion daily spot tape. The halving removed something like 0.4% of daily float supply. That is a rounding error wearing a regime-change costume.
Yields die where liquidity dries up. That is the variable to watch — not a supply schedule drafted in 2008.
Now test the crypto version of the frequency claim. "Bitcoin has never had a negative year following a halving." True, on the current window. Also: three observations, all inside the largest monetary expansion in modern financial history. The base rate for any twelve-month BTC window since 2012 being positive is roughly 70%. Three consecutive positives at a 70% base rate has a probability of 34% — unremarkable. Nobody would publish 0.7^3 as a natural law.
The halving is a calendar identity. "Post-halving year" and "year of maximum retail attention surplus" are collinear. The coefficient is unidentifiable from four points. Follow the chain, not the hype — and the chain here shows four observations, not a mechanism.
The contrarian angle, twice over
The tempting conclusion is that the calendar has no effect at all. That is also wrong, and it is the mistake the skeptics keep making.
The year-three pattern has a plausible weak mechanism: policy uncertainty resolves after midterms, the fiscal and monetary calendar clears, and the equity risk premium compresses slightly. That is a real channel. But it runs through uncertainty resolution, not through the vote itself. The error in the midterm study is not claiming an effect exists. It is attributing the effect to a political event when the causal pathway — if there is one — is calendar-clearing and risk-premium compression.
The second blind spot is survivorship. The S&P 500 is a survivorship-biased index by construction; failures get delisted and removed. Any long-horizon positive-return statistic inherits that upward drift automatically. Part of 19/19 is a fact about index methodology, not about elections.
And the most checkable unverified claim in the whole piece: "independent of which party wins." Testable. Not tested. I would want split-sample returns for unified versus divided government before accepting it. No group data is provided. Data doesn't lie; windows do.
The takeaway
In a chop tape, positioning should be driven by measurable carry and observable liquidity depth, not by calendar folklore. That is the whole discipline. The market is sideways because the marginal buyer has no urgency; narratives fill the vacuum where flow is absent.
What I am watching this week, and what you should be too: perp funding dispersion across venues, stablecoin net issuance, and the spread between ETF creations and realized volatility. Those are quantities that can be measured, revised, and falsified. A nineteen-year-old win streak cannot be falsified — which is exactly why it keeps getting sold.
The interesting question is not whether the midterm pattern survives the next sample. It is whether a market that has already learned to price the halving can still be moved by it. If Bitcoin's cycle premium compresses toward its base rate, the trade is to sell the calendar narrative, not buy it.