Prediction markets have not systematically overpriced left- or right-leaning candidates, according to a National Bureau of Economic Research working paper by Eric Zitzewitz that Andreessen Horowitz highlighted on October 9, 2026 in its weekly "This Week in Charts" newsletter. The headline partisan coefficient across all markets is +0.03 percentage points, with a standard error of 2.66.

In Short

A researcher checked whether betting markets on elections lean towards one political side, and found they do not, at least not by any amount that can be told apart from chance. This matters because Google now accepts ads from regulated prediction markets, shows their odds in Google Finance, and has spent the year adding US states where those ads are banned. The finding does not settle whether these markets are accurate, only that their errors do not consistently favour one party.

What the a16z newsletter published

The October 9 edition of "This Week in Charts," written by Moses Sternstein and distributed through the a16z Substack, ran under the headline "Are Prediction Markets Politically Biased?" It covered four subjects: prediction markets, the declining usefulness of the Rule of 40 as a software valuation benchmark, the narrowing of analyst earnings upgrades towards energy and hardware, and a contraction in global call-centre employment. The newsletter opened by noting that a16z is leading an investment in TypeSafe AI, the company behind a product called Jev.

The prediction market section rests on two charts. The first tracks the Kalshi market "Brazil Presidential Election Winner." The second, titled "No Apparent Political Bias In Prediction Markets," reproduces estimated return coefficients from Zitzewitz's paper, which the chart cites as "NBER Working Paper No. 35846 (Oct. 2026), Tables 3-5." The NBER website lists the paper with an October 2026 issue date.

a16z attaches a standard disclaimer to the newsletter. According to the firm, the content "should not be relied upon as legal, business, investment, or tax advice," and "a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information" drawn from third parties.

The Brazil case

The newsletter used the Brazilian presidential race as a lead-in. Flávio Bolsonaro won the first round, held on October 4, 2026. The author was candid about the limits of the commentary. "Charts knows next to nothing about Brazilian politics, so we really couldn't say," according to the newsletter.

The Kalshi chart, sourced to data dated October 5, 2026, shows two lines. Luiz Inácio Lula da Silva's contract sits between roughly 35% and 60% from February 2025 through early 2026. Bolsonaro's contract reads zero until around December 2025, when it appears at close to 19%. It climbs to roughly 50% in the spring of 2026, falls back to around 25% by mid-summer while Lula's line peaks near 67%, and then rises steadily from August. The two lines cross at a vertical marker labelled "Sep 10." By the chart's final point, Bolsonaro stands at 85% and Lula at 17%.

Those two final figures add up to 102%. That is not an error in the chart. Each contract on Kalshi trades separately, so the prices need not sum exactly to 100, and the gap is a rough measure of the friction built into the market. It also matters that an 85% price on October 5 is not a resolved outcome. The market was still pricing a contest that had not ended.

The newsletter drew a clear conclusion from the timing. "Whatever the surprise, prediction markets were early and right on this one by ~3 weeks," it stated. a16z added that MSCI's Brazil ETF "jumped ~15% after the news of Bolsonaro's first-round win," though the ETF chart referred to in the text does not appear in the captured version of the newsletter, so that figure cannot be checked against the graphic. According to the newsletter, the ETF "also started climbing back in August (when Bolsonaro's Kalshi odds began to turn), but not nearly as steeply or as uniformly."

One market getting one election right says little about systematic bias. The newsletter acknowledged as much, calling the Brazil section "a somewhat longwinded segue" into the research.

How the Zitzewitz paper tests for bias

The method is simple in principle. If traders consistently overrate a certain kind of candidate, then betting against that candidate would earn money over time, and betting on them would lose it. A persistent return in either direction is the fingerprint of bias. The newsletter put it this way: "if, e.g., prediction markets consistently underrate female candidates, then taking the 'over' on female candidates would be a winning strategy (reflected by a positive coefficient)."

A negative coefficient therefore means the market overprices that attribute: buyers pay too much and lose. A positive coefficient means the market underprices it.

According to the newsletter, the paper "aggregated over 100 years' worth of prediction markets (albeit with most of the sample coming recently)." The abstract on the NBER website is more specific, describing prediction market returns from 1880 to 2025. That span covers the informal election betting markets of the late nineteenth and early twentieth centuries as well as the modern commercial platforms.

The political coefficients

The a16z chart, in percentage points, shows four political rows for right-leaning candidates:

  • All markets: +0.03, 95% confidence interval -5.18 to +5.24, standard error 2.66.
  • All, commercial markets only: -0.43, interval -5.55 to +4.69, standard error 2.61.
  • US elections: +0.84, interval -7.76 to +9.44, standard error 4.39.
  • Non-US elections: -1.42, interval -3.97 to +1.13, standard error 1.30.

Every one of the four intervals contains zero. The US elections row has the widest band of any on the chart, spanning more than 17 percentage points, which says as much about the noise in the data as about any underlying tendency. The commercial-only row is a useful check, since commercial platforms are the ones now buying advertising, and it moves the estimate by less than half a point.

The non-US row is the one exception the newsletter flags. "The only exception is for non-US elections, where prediction markets appear to reflect some bias in favor of right-leaning candidates (i.e. the coefficient is negative because markets overrate their chances), albeit over a much smaller sample," according to a16z. The coefficient of -1.42 is the only political figure larger in absolute terms than its own standard error of 1.30. Even so, its confidence interval runs from -3.97 to +1.13. "Even there, however, the 95% confidence interval indicates that the bias is not statistically significant," according to the newsletter.

The demographic coefficients

The second panel covers candidate characteristics:

  • Female versus male: -0.95, interval -8.48 to +6.58, standard error 3.84.
  • Nonwhite versus white: -0.85, interval -8.73 to +7.03, standard error 4.02.
  • Black versus other nonwhite: -0.69, interval -8.86 to +7.48, standard error 4.17.
  • Age, per additional 10 years: +1.17, interval -1.57 to +3.91, standard error 1.40.

The signs point in a consistent direction. Negative values for women and nonwhite candidates mean markets priced them slightly higher than outcomes justified; the positive age coefficient means older candidates were priced slightly lower. a16z described this as "a small (statistically insignificant) bias in favor of women and non-white candidates, and against older candidates." None of the four coefficients exceeds its standard error, and the three identity comparisons carry intervals roughly 15 to 16 points wide.

"When it comes to partisanship, the coefficient is a bare 0.03," the newsletter stated, concluding that "when it comes to elections, prediction markets are pretty good at leaving their biases at the door."

What the chart leaves out

The a16z summary does not mention two findings that the NBER abstract describes. According to the abstract, most political markets in the sample display a favourite-longshot bias, the well-documented pattern in which long shots are overpriced and favourites underpriced. The abstract also reports a very small overpricing of left-leaning outcomes in markets on near-term polling averages, which the author says is detectable only because that sample is so large.

Neither finding contradicts the headline. Favourite-longshot bias is a pricing error unrelated to party, and the polling-average effect runs opposite to the non-US election result and sits in a different kind of contract. But both show that "no partisan bias" is a narrower claim than "no bias." A market can be politically even-handed and still misprice the odds of an underdog.

The study also measures something different from accuracy. Earlier this year, PPC Land reported on a Vanderbilt study of the 2024 US election that found accuracy rates of 93% for PredictIt, 78% for Kalshi and 67% for Polymarket. A platform can be unbiased between parties in the Zitzewitz sense while still being wrong often, as long as the errors do not consistently favour one side.

Why this research lands in advertising

Prediction markets have moved from the fringe of the ad business to a defined, if contested, category in the space of a year. Google opened its ads system to prediction markets on January 21, 2026, after posting the policy on January 5. Only two kinds of advertiser qualify: designated contract markets authorized by the Commodity Futures Trading Commission, and brokerages registered with the National Futures Association that offer access to those contracts. The same policy continues to exclude binary options with fixed payouts and unregulated online gambling.

The geography has narrowed since. Nevada was excluded from the start. On June 2, Google added Ohio to the list of excluded states, effective the day the update was posted, after a federal court in the Southern District of Ohio denied Kalshi a preliminary injunction against state regulators. Michigan and New York were later added, and on October 7, two days before the a16z newsletter, Google posted a change log entry prohibiting prediction market ads in Connecticut with immediate effect, according to its Advertising Policies Help Center.

Google has also put prediction market prices in front of users outside the ad auction. On November 6, 2025, Google added Kalshi and Polymarket probabilities to Google Finance, letting users ask natural-language questions about events such as GDP growth. By April 2026, that integration had reached more than 100 countries, and Polymarket contracts had briefly appeared in Google News, which a Google spokesperson called an error.

The commercial stakes behind the platforms are large. Kalshi's monthly trading volume passed $1 billion by mid-2025, and the company closed a $1 billion financing round at an $11 billion valuation, figures PPC Land detailed in its coverage of the Ohio ban. The same coverage noted that state-regulated New York sportsbooks face a 51% tax rate that CFTC-overseen exchanges do not.

Against that backdrop, the Zitzewitz paper matters in a specific way. State regulators have mostly attacked prediction markets on gambling grounds, focusing on sports contracts. The political critique is different: that platforms whose traders are not representative of voters will push election odds towards their users' preferences, and that those odds then circulate in news coverage, finance products and search results. The paper's answer, on the evidence presented, is that the historical record does not support that worry for election contracts. It does not address sports contracts, which analyst Dustin Gouker estimated at 85% to 90% of Kalshi's contracts.

The US midterm elections on November 3, 2026 will be the first national contest since Google opened the category to advertising. Political odds from Kalshi and Polymarket will be quoted widely in the weeks before. Whether markets that may be unbiased on average perform well in a single cycle is a separate question, and one the paper does not claim to answer.

Rule of 40 and the software multiple

The newsletter's second section addresses a metric familiar to anyone who follows listed ad tech and marketing software companies. The Rule of 40 adds revenue growth to profit margin; a combined score of 40 or more was long treated as a sign of a healthy software business, whatever its stage.

According to a16z, that relationship has broken down. A chart sourced to CapIQ data as of October 6, 2026, through the second quarter of 2026, plots median software "Rule of X" against median enterprise value to annualized revenue. The sample ranges from 71 to 146 companies over time, and the score is defined as year-over-year revenue growth plus net income margin including stock-based compensation. The text puts the median Rule of X rise at "from ~17% to ~23% over the past ~3 years." On the chart, the revenue multiple falls from around 6x through most of 2024 and 2025 to roughly 3.4x in 2026, before ticking up to about 3.6x at the last data point. For comparison, the multiple peaked at about 13x in 2021.

A second chart shows the eight-quarter rolling Pearson correlation between the two series. It rose to about +0.94 in 2023, held positive through mid-2025 at around +0.48, then fell through zero and reached roughly -0.6 around the end of 2025. The final reading is near -0.38.

a16z's explanation is that the composition of the score has shifted. "Public (and private) techcos responded to post-ZIRP regime change by trading growth for profitability," according to the newsletter. Margins are lifting the Rule of X, while investors reward growth more heavily.

A four-panel scatter chart, again from CapIQ data as of October 6, splits 140 companies (excluding MercadoLibre and neoclouds) into Rule of X buckets: 56 companies at 0 to under 20, 55 at 20 to under 30, 17 at 30 to under 40, and 12 at 40 and above. The fit between growth and enterprise value to next-twelve-month sales is strongest in the 20-to-30 bucket, with a quadratic R-squared of 0.43, and weakest in the 40-plus bucket, at 0.02. "For Rule of 20 all the way to Rule of 40, most companies are clustering in the same range between 5-10x EV/NTM sales," according to a16z.

The chart colour-codes an AdTech category. Most of its points sit below 5x across the buckets. MNTN, the only labelled ad tech company, appears in the Rule of 40-plus group at about 1x next-twelve-month sales, with projected growth of roughly 23%, while Palantir in the same bucket trades at about 44x on growth near 65%. MNTN, a CTV performance advertising platform, reported second-quarter 2026 revenue of $82.5 million, up 21%, with adjusted EBITDA of $21.5 million and net income of $6.7 million. On a16z's reading, that profile is exactly the one the market now discounts: a high combined score without the top-line growth that earns a premium.

Earnings upgrades narrow to hardware and energy

The third section shifts to the wider equity market. A Morgan Stanley chart dated October 5, 2026 tracks S&P 500 earnings revisions breadth, the share of upgrades minus downgrades. It turned positive in mid-2025, climbed to a peak near 28% in the summer of 2026, and has since slipped to roughly 22%. The newsletter describes the trend as "analyst optimism is training its fire."

A FactSet table, as of October 5, breaks revisions out by sector. For 2026 estimates, the S&P 500 is up 7.0% year to date but only 2.3% in the third quarter. Energy rose 19.3% in the quarter, tech hardware 7.6% and semiconductors 9.8%, while software and services moved just 0.7%. For 2027, semiconductors lead with a 27.3% upward revision, against 0.4% for software and services.

Communication services, the sector that houses Alphabet and Meta, shows the largest year-to-date upgrade for 2026 at 20.7%. In the third quarter, however, its estimates were cut by 1.8%, and its 2027 revision stands at 0.6%. Consumer discretionary followed a similar path: up 17.9% year to date, down 1.7% in the quarter.

There is a discrepancy between the text and the table. The newsletter states that "for the year, earnings revisions have been positive across the board (outside of utilities)." The FactSet chart shows three negative year-to-date figures for 2026: health care at -0.9%, materials at -1.3% and utilities at -0.2%.

Two further charts support the "atoms" thesis. Institute for Supply Management survey data, via Morgan Stanley, show the three-month average share of industries reporting a higher order backlog at close to 60%, which the newsletter calls the highest level "since pandemania." A pair of Carlyle charts, based on portfolio company data and Bloomberg as of October 5, show the firm's Global AI Hardware Shipments Index, set at 100 in August 2021, approaching 1,000 by mid-2026, while annual growth in its IT Services Spending Index has plateaued at around 29%. The newsletter text refers to "~30%" growth in "the broader category of tech-spend," while the chart itself is labelled as IT services spending. That fits the pattern visible in company guidance: Alphabet, for example, set 2026 capital expenditure guidance of $180 billion to $190 billion, roughly double its 2025 level.

Call centres

The final section cites Revelio data showing global call-centre employment "has gone sharply negative, after more than a decade of fairly consistent growth." According to the newsletter, higher-income countries turned negative in 2022, lower-middle-income countries stayed positive for about two more years, and "as of Q4'25, call center employment growth has gone negative everywhere, even in the poorest countries." Outside China, call-centre employment has lagged broader white-collar service work in every major non-US market, according to a16z.

The three Revelio charts that accompany this section did not render in the captured copy of the newsletter, so the figures above rest on the prose alone. The newsletter itself hedges the causal claim. "Could there be other reasons for below-trend call center employment, besides AI? Probably," it stated, noting that growth "began slowing when interest rates went up, which is slightly before GPT was released into the wild."

Reading the evidence

The four sections share a theme the newsletter does not spell out: headline numbers that hide what is underneath. A partisan coefficient of 0.03 hides a favourite-longshot bias. A rising Rule of X hides slowing growth. Positive revisions breadth hides a narrowing set of winners.

For the prediction market discussion in particular, the data presented is limited by its own uncertainty. Confidence intervals of 10 to 17 points on most rows mean the study can rule out large, persistent partisan tilts, but not small ones. The non-US result, the only row whose point estimate exceeds its standard error, comes from what a16z calls "a much smaller sample." Is that a real effect that more data would confirm, or noise? The chart cannot say.

The finding is nonetheless one of the more substantial pieces of evidence yet published on a question that has followed prediction markets into mainstream products. It arrives as election contracts are about to be tested in a live US cycle, and as the advertising rules around these platforms continue to shift state by state.

Timeline

Summary

Who: Andreessen Horowitz (a16z), through Moses Sternstein's "This Week in Charts" newsletter, and economist Eric Zitzewitz, author of NBER Working Paper No. 35846. Kalshi, Google, MNTN and data providers CapIQ, FactSet, Morgan Stanley, Carlyle, the ISM and Revelio also feature.

What: The newsletter presented research finding no statistically significant partisan or demographic bias in election prediction markets, with a headline partisan coefficient of +0.03 percentage points. It also argued that the Rule of 40 has lost its link to software valuations, that earnings upgrades have narrowed to energy and hardware, and that call-centre employment is contracting globally.

When: The newsletter was published on October 9, 2026. The NBER paper carries an October 2026 issue date, and most chart data is dated October 5 or 6, 2026.

Where: The research covers US and non-US elections in data from 1880 to 2025. The Brazil example concerns Kalshi's market on the 2026 Brazilian presidential election. The advertising context centres on Google's US prediction market policy, which excludes several states.

Why: Prediction markets have become an advertising category and a data source inside Google products, and their political odds circulate widely ahead of the November 3, 2026 US midterms. Evidence on whether those odds tilt towards one party bears directly on how much weight they deserve, even though the paper does not measure accuracy and the newsletter's own figures leave room for small effects.