Strategies·November 17, 2026·8 min read

A Simple Polling Model for Election Contracts

How to aggregate polls, estimate house effects, turn a lead and its standard error into a win probability, and decide when an election contract is worth betting against.

election pollingforecastingcalibrationpolymarket
Ballot box — illustrative image
Photo: Wikimedia Commons · CC BY-SASource

A polling model for election contracts is not a prediction of the outcome. It is a way to turn a pile of polls into a probability with an explicit error bar, then compare that number with the price. The useful output is not 54.3%. It is a distribution: 54% plus or minus 4, with a known tendency to be wrong in the same direction across many states.

That distinction is the difference between a model and a story. A model that says the lead is four points, the standard error is three, so the fair probability is about 91%, can be checked against results. A model that says the momentum is with the challenger cannot.

Build the poll average

Start with a list, not a single poll. Record dates, sample size, population, mode, sponsor and margin. Decide the weights before you look at the result.

Weight by recency, sample size and pollster quality. Recency can be a 14-day half-life. Sample size can be weighted by the square root of n, so a 5,000-person poll does not dominate a dozen 800-person polls. Pollster quality can be a shrinkage factor based on past error.

Average margins, not percentages. If one poll says 48-44 and another says 46-45, the average margin is 2.5 points. Allocate undecideds explicitly, either by normalizing the two candidates or proportionally.

A worked example: +3 with n = 800, +5 with n = 1,200 and +1 with n = 600. Weight by the square root of n: 28.3, 34.6 and 24.5. The weighted average is 3.4 points.

House effects

A house effect is a pollster's average deviation from the consensus in comparable races, after controlling for timing and geography. It is not the same as a bad poll.

Estimate it over many races. If a pollster has 40 comparable races and averages 1.5 points more Democratic, subtract 1.5 from its margin. Shrink small samples toward zero.

House effects change with likely-voter screens and weighting. Use a rolling window. Do not confuse a house effect with herding: herding is suspicious similarity, a house effect is a persistent deviation.

Sampling error arithmetic

The margin of error for a proportion is 1.96 times the square root of p(1 - p) divided by n. At p = 0.5 and n = 1,000, that is 1.96 times 0.0158 = 3.1 percentage points. That is the number quoted in most poll stories.

The number that matters for a race is the margin of error on the lead, which is larger. For two candidates with shares p1 and p2, the standard error of the difference is the square root of (p1 + p2 - (p1 - p2)^2) divided by n. At 50-46 with n = 1,000, that is the square root of (0.96 - 0.0016) divided by 1,000, or 3.1 points. The 95% margin is 6.1 points. A four-point lead is not statistically distinguishable from a tie.

That assumes a simple random sample. Real polls have design effects, so multiply the standard error by 1.2 to 1.5. A state poll with n = 600 has a lead margin of about 8 points at 95%.

A poll average reduces sampling error by the square root of the number of independent polls. Five polls with a 3-point standard error produce 1.3 points. But the polls share house effects and timing. The average reduces the random component, not the correlated one.

Correlated state errors

National polls measure the national vote, not the Electoral College. State polls share a common national swing, so their errors are correlated.

A simple structure is: state margin = national margin + state lean + state-specific error.

If the national error has a standard deviation of 2 points and the state-specific error is 3 points, the total is 3.6 points. Averaging across states does not drive this to zero, because the national component is common.

Jennings and Wlezien found that polling errors are correlated within an election and that accuracy has not improved much over time. A large average can be unbiased and still wrong in the same direction across every state.

From margin to probability

Assume the error on the margin is approximately normal. The probability of winning is the normal cumulative distribution of the lead divided by the standard error. A four-point lead with a three-point standard error gives a z-score of 1.33 and a probability of about 91%. A four-point lead with a five-point standard error gives a z-score of 0.8 and a probability of about 79%.

That is the entire conversion. The price of the contract should be compared with this probability, not with the poll lead. A market at 85 cents on a 91% estimate has a 6-cent gross edge. At 85 cents in a politics market, 1,000 shares pay a fee of 1,000 times 0.04 times 0.85 times 0.15 = $5.10, or 0.51 cents per share. The break-even probability is 85.51%. The edge after fees is 5.49 cents per share, or 6.5% of the stake.

Sizing is where most models fail. The Kelly fraction is (0.91 - 0.85) divided by 0.15, or 0.40, which says bet 40% of bankroll. That is absurd for an estimate with a 3-point standard error. Use quarter Kelly, or 10%, and cap the event at 2%. If the standard error is 5 points, the probability is 79% and the trade disappears at 85 cents.

When to bet against the market

Only when your error bar is tighter than the gap and you can explain why the market is wrong. A different point estimate is not enough. If your model says 91% and the market says 85%, the gap is 6 points; if your standard error is 5 points, the gap is inside the noise.

Legitimate reasons include a stale market that has not absorbed a new poll, a thin market dominated by one trader, and a convention bounce. A feeling about momentum and a single outlier poll are not reasons.

Avoid the final 48 hours unless you have a turnout model. Polls stop mattering and news dominates. Check the resolution rule too: win the election and win the popular vote are different contracts.

How to execute

  1. Build the average from every poll in the window, with house effects removed.
  2. Set the error bar from the lead, the design effect and the correlated state error.
  3. Convert the margin to a probability with the normal approximation.
  4. Compare that probability with the executable ask, then subtract the taker fee and half the spread.
  5. Size at quarter Kelly and cap the event at 2% of bankroll.
  6. Write down the thesis and the evidence that would change it before you enter.
  7. Recheck the resolution source and deadline. If the contract settles on a different question, the model does not apply.

Model maintenance

Log every poll and every output. Score with Brier and log loss, not just win or loss. A model that wins 70% of the time when it says 70% is calibrated; one that wins 90% is underconfident; one that wins 50% is overconfident.

Plot calibration by bucket after the cycle, not during it. If your 70% bucket wins 60%, widen the error bars.

Freeze the model during an election. Changing weights after a bad poll is how a model becomes a narrative. Version it, keep a changelog, and decompose the largest errors into house effect, swing and turnout.

Risks and limits

The biggest risk is correlated error. A model can be right about the average and wrong about every state. The second is late swing: polls taken two weeks before the vote may not reflect the final decision. The third is turnout: likely-voter screens are models themselves, and they can be wrong in the same direction.

The fourth is herding. When pollsters converge on a comfortable narrative, the average looks precise and is not. The fifth is resolution: certification, recounts and legal challenges can make the contract payoff different from the vote count. Read the rule before you trust the probability.

A short checklist

  1. How many polls, and what is the average margin?
  2. Are the polls weighted for recency, sample size and house effect?
  3. What is the standard error on the lead, including the design effect?
  4. What is the correlated state error, and does it change the conclusion?
  5. What probability does the margin imply, and what is the fee-adjusted break-even?
  6. Is the market gap larger than my error bar?
  7. Is the resolution rule the one my model is forecasting?

A polling model is useful because it makes the error explicit. The number is not the edge. The edge is the difference between a probability with a known error bar and a price that has not priced that error.

Sources

  1. 1.Resolution— Polymarket Documentation
  2. 2.Standard Definitions and Methods— American Association for Public Opinion Research
  3. 3.Methods Research— Pew Research Center
  4. 4.Election Polling Errors across Time and Space— Nature Human Behaviour
  5. 5.Dynamic Bayesian Forecasting of Presidential Elections in the States— Journal of the American Statistical Association

More from the journal

All articles