{"path":"research/folded-routes-and-shared-premises/markets-and-experts-report-2026-10-23.md","content":"# Are prediction-market prices and political-science forecasts independent reasons? A research helper's report, as returned (2026-10-23)\n\n*A research helper (Claude Sonnet 5.5) was asked, read-only, for evidence on how far prediction\nmarkets and expert or political-science forecasts share information and err together, who trades,\nreal elections where both pointed one way, and any published correlation of their errors. Asked for\nthe first independence case he read (his reading:\n[../how-independent-reasons-add/his-readings.md](../how-independent-reasons-add/his-readings.md)).\nKept verbatim; nothing below was rewritten. Read by the note that uses it:\n[../folded-routes-and-shared-premises.md](../folded-routes-and-shared-premises.md).*\n\n---\n\nREPORT: Are prediction-market prices and political-science forecasts independent reasons for \"the incumbent will win\"?\n\nRead-status tags: [full text] = read in full; [part] = read key sections or substantial excerpts; [abstract only]; [secondary] = reported by someone else. I wrote no lasting files. I briefly saved one PDF to the scratchpad and deleted it. A PDF fetched by the web tool landed in the tool-results folder automatically.\n\nBOTTOM LINE\nThe two reasons are partly dependent. They are not independent and they are not one story. They share a lot of public input (polls, economy, approval). In the big recent misses (UK 2015, Brexit 2016, US 2016) both leaned the same wrong way. But the day-to-day error series of a market and an expert panel are only weakly to moderately correlated, and markets do often react to things models and polls miss. Two caveats up front:\n- I found NO published estimate of the market-vs-expert-panel error correlation across many elections. The closest is five US elections (item 4).\n- I found NO study that measures what share of traders are political scientists.\n\n---\n1. HOW FAR DO MARKETS AND POLLS/MODELS/EXPERTS SHARE INFORMATION?\n\na) Markets mostly follow polls when polls exist.\n- Erikson & Wlezien 2012, \"Markets vs. polls as election predictors: An historical assessment\", Electoral Studies. [part]\n  - Using US presidential markets 1936–2008, they find \"when we have both market prices and polls, prices add nothing to election prediction beyond polls.\" For recent markets, \"candidate prices largely follow the polls.\"\n  - The correlation between poll results and market prices is +.86 for 1936–2008 and +.91 for 1952–2008.\n  - Even in 1948, when the polls were wrong, \"the market followed the polls.\"\n  - Markets were far better in the pre-poll era (1880–1932).\n  - https://statmodeling.stat.columbia.edu/wp-content/uploads/2024/08/Erikson-and-Wlezien-Electoral-Studies-2012-1.pdf\n- Erikson & Wlezien 2008, Public Opinion Quarterly, on Iowa Electronic Markets (IEM) 1988–2004. [part]\n  - With poll leads properly discounted, poll-based forecasts beat vote-share market prices.\n  - Winner-take-all prices were dominated by poll-based win projections in all four elections tested.\n  - \"Market prices contain little information of value for forecasting beyond the information already available in the polls.\"\n  - https://academic.oup.com/poq/article-pdf/72/2/190/5414848/nfn010.pdf\n- Duquette et al. 2014, Atlantic Economic Journal. [abstract only]\n  - In 2012, RealClearPolitics poll averages Granger-cause Intrade prices, but not the reverse.\n  - https://link.springer.com/article/10.1007/s11293-014-9430-6\n- Goodell, McGroarty & Urquhart, International Review of Financial Analysis. [abstract only, via a secondary blog summary]\n  - The Iowa and Intrade 2012 markets and Nate Silver's forecasts were strongly cointegrated.\n  - Silver and Iowa Granger-caused Intrade.\n  - https://nationalaffairs.com/blog/detail/findings-a-daily-roundup/popularity-contest\n\nb) Counter-evidence: markets also carry non-poll information.\n- Rothschild 2009, Public Opinion Quarterly. [part]\n  - In 2008 (74 state-level races), debiased Intrade prices were more accurate than debiased polls and FiveThirtyEight, especially early in the cycle and in uncertain races.\n  - He argues the poll information is public and so in traders' information set, and that prices add dispersed or unpublished information.\n  - https://academic.oup.com/poq/article/73/5/895/1868587\n- Khan & Lieli, 2008 Democratic primaries. [abstract only]\n  - Two-way Granger causality: prices anticipate part of the \"pollster's surprise\", and the surprise is also news to the market.\n  - https://economics.ucr.edu/repec/ucr/wpaper/201610.pdf\n- Tufts paper on PredictIt 2016 (30 markets). [abstract only; looks like a student thesis, so low weight]\n  - Found no significant influence of polls on prices.\n  - https://dl.tufts.edu/downloads/4m90f7221?filename=w6634g725.pdf\n  - Evidence on market dependence on polls is therefore not unanimous.\n\nc) Markets are partly a different kind of forecaster than models.\n- Seager & Sethi, \"Models, Markets, and the Forecasting of Elections\" (2020 US battleground states). [part]\n  - The model and the market \"make different kinds of errors in different states\". The model was confidently wrong in some cases and the market excessively uncertain in others, so a simple average beat either.\n  - They also write that traders \"are free to use any inputs, including model predictions.\"\n  - https://ar5iv.labs.arxiv.org/html/2102.04936\n- Sethi et al., \"Political Prediction and the Wisdom of Crowds\" (2024 US election, 3 statistical models vs Polymarket). [part]\n  - Models serve \"as information aggregators that traders incorporate into belief updates.\"\n  - All three models lost money against the market on the headline presidential contract.\n  - https://dl.acm.org/doi/full/10.1145/3715928.3737483\n- Atanasov et al. 2017, Management Science (261 events, 2,400+ forecasters randomly assigned). [part]\n  - Prices beat the simple mean of prediction polls.\n  - Team prediction polls, statistically aggregated, beat markets by about 12–13% on Brier score.\n  - This concerns market vs crowd-poll accuracy, not dependence.\n  - https://pubsonline.informs.org/doi/10.1287/mnsc.2015.2374\n\nd) Who trades. (Not representative, and I found no direct evidence on political scientists.)\n- Forsythe et al. 1992, American Economic Review. [part]\n  - The original Iowa market (1988) was restricted to the University of Iowa community.\n  - Traders were 71% male and 90%+ white. Students were the largest group.\n  - Traders were well educated, middle or upper income, and more politically active than average (80% registered to vote vs 67% in a national poll).\n  - https://gwern.net/doc/statistics/prediction/election/1992-forsythe.pdf\n- Berg & Rietz, IEM review. [part]\n  - Traders are \"overwhelmingly male, well-educated, high income, and young (the average age is close to 30)\".\n  - https://biz.uiowa.edu/faculty/trietz/papers/IEM%20Results.pdf\n- Crane & Strumpf, \"Diversity, Efficiency, and Incentives in Political Prediction Markets\". [part]\n  - Caveat: I only found it on a personal-looking host (borowiecki.dev), and it is an unreviewed preprint.\n  - Kalshi 2024 traders were 86% male and half were under 30.\n  - PredictIt 2016: 92% of traders with a recorded sex were men, and about 42% were under 30.\n  - It quotes Rothschild & Sethi 2016 that Intrade traders were \"predominantly young and male\".\n  - Price-setting heavy traders showed no link between position direction and home-area politics. Casual one-trade users did.\n  - https://borowiecki.dev/pdf/2610.28277\n- PredictIt's own demographic claim. [secondary; company-reported]\n  - Affluent millennials, many in \"finance, law, politics and technical fields\".\n  - https://www.fia.org/marketvoice/articles/political-prediction-markets-turn-pro\n- Rothschild & Sethi 2016, Journal of Prediction Markets. [abstract only]\n  - Intrade 2012: about 6,300 trader accounts. Most directional traders bet consistently in one direction, and there is suggestive evidence of manipulation by one large trader.\n  - https://doi.org/10.5750/jpm.v10i1.1179\n- Bottom line: academics were in the original Iowa pool and PredictIt has academic sponsors. I found no study counting political scientists as traders. Whether they \"bet\" is unmeasured.\n\n---\n2. REAL CASES WITH KNOWN OUTCOMES\n\nCase A: Brexit, 23 June 2016. Both favoured Remain. Leave won with about 52%. [Hanretty op-ed, full text: https://goodauthority.org/news/heres-why-pollsters-and-pundits-got-brexit-wrong/]\n\n| Source | Date | Stated forecast |\n|---|---|---|\n| Political Studies Association (PSA) expert survey (496 academics, 13 pollsters, 33 journalists, 54 others) | fieldwork 24 May–2 Jun | 87% said Remain more likely; 5% Leave; 8% equal. Mean probability of Leave 38% (academics 38%, pollsters 33%, journalists 32%). Mean expected vote Remain 55 / Leave 45. |\n| Betting and prediction markets | same period | \"just 26%\" or lower for Leave |\n| Fisher & Shorrocks final combined forecast, data of 22 Jun (expert row = the late-May PSA survey, not updated) | 22 Jun | Remain win probability: experts 62.0%, betting markets 76.7%, prediction markets 73.4%, polls 55.6%, poll-based models 68.5%, citizen forecasts 55.2%. Combined 66.5%. |\n| Market-implied Leave probability (Hanretty) | morning of vote | 23% |\n| Market-implied Leave probability (Hanretty) | 90 min after polls closed | 11.3% |\n| Betfair Remain | around 23 Jun | 84–86% on the day [secondary]; at one stage over 90–93% |\n\n- The PSA report itself warns that consensus across expert groups \"may just be a sign of groupthink\", and that experts and bettors \"have access to the same opinion polls\".\n  - https://eprints.soton.ac.uk/396305/1/PSA_20EU2016_20Report.pdf [part]\n- Fisher & Shorrocks 2017: \"Betting and prediction market participants and volunteer forecasters were the most overconfident\" for Remain. They suggest bettors distrusted polls after 2015 and expected a late swing to the status quo, as in Scotland 2014 and Quebec 1995.\n  - https://doi.org/10.1080/17457289.2017.1380656 [abstract only]\n  - Final forecast table: https://electionsetc.com/2016/06/23/final-combined-eu-referendum-forecast/ [part]\n  - PSA numbers: https://www.psa.ac.uk/news/leading-academics-journalists-and-pollsters-predict-eu-referendum-results/ [part]\n- Reading: same direction, but the experts were noticeably less confident than the markets, and so less wrong.\n\nCase B: UK general election, 7 May 2015. Both favoured a hung parliament, and the Conservatives won a majority. [Lewis-Beck et al., part]\n- PSA expert survey, Feb 2015: Conservatives 33% vs Labour 32%; seats 278 vs 282.\n  - Only 6% of respondents said one party would win a majority.\n  - Source: https://psa.ac.uk/sites/default/files/PSA%20EU2016%20Report.pdf [part]\n- All 12 academic forecasting teams foresaw a hung parliament.\n  - Betting markets gave a 6% implied probability of a Conservative majority in the final week.\n  - https://ora.ox.ac.uk/objects/uuid:5379074d-2e4d-4629-aeff-69976cbba74b/files/rht24wk45c [part]\n- Result: Conservatives 330 seats, 37.7% of the vote, ahead of Labour on 232 seats and 31.2%. Final polls understated the Conservative lead by about 7 points.\n  - https://doi.org/10.1111/rssa.12329 [abstract/part]\n- The authors say the main failure was the polls: \"the problem was also present\" in other data types, and \"conventional wisdom\" was \"strongly informed by the opinion polls.\"\n\nCase C: US presidential election, 8 Nov 2016. Mixed.\n- Markets: PredictWise (a market aggregator) about 89% Clinton on election night. PredictIt traders about 80% (Rothschild's own fundamentals model had a narrow Trump electoral-college win). Source: PBS [part]: https://www.pbs.org/newshour/show/prediction-market-prophets-bet-wrong-president\n- Political-science models forecast the two-party POPULAR vote, which is a different quantity. Clinton won it with 51.1%. They were mostly accurate; seven of 11 were within one point (Tien & Lewis-Beck, LSE blog [part]).\n  - Stated probability that Clinton wins the popular vote: Lewis-Beck/Tien 83%, Campbell 75% and 69%, Abramowitz 66%.\n  - Norpoth was the odd one out, giving Trump 87% in March 2016.\n  - https://researchonline.lse.ac.uk/id/eprint/69308/1/blogs.lse.ac.uk-In%20forecasting%20the%202016%20election%20result%20modelers%20had%20a%20good%20year%20Pollsters.pdf\n  - https://centerforpolitics.org/crystalball/how-accurate-were-the-political-science-forecasts-of-the-2016-presidential-election/ [part]\n- The panel of 15 political scientists in PollyVote (Graefe) had 75 of 77 individual forecasts in the last 100 days overestimating Clinton's vote share. Its combined mean absolute error was 2.1 points (vs a 1.5-point average in 2004–12).\n  - IEM overestimated Clinton by 4.8 points on average.\n  - Graefe's account: \"Experts – including self-selected experts in prediction markets – apparently thought that the polls would underestimate Clinton.\"\n  - https://mpra.ub.uni-muenchen.de/83282/1/MPRA_paper_83282.pdf [part; I read the expert, prediction-market and bracketing sections]\n- Reading: popular-vote models and electoral-college markets did not err on the same question. But the expert PANEL and the market did err in the same direction.\n\nCase D: US presidential election, 2024. Disagreement among methods, market and fundamentals aligned.\n- Markets on 4 Nov: Polymarket 56–58% Trump; Kalshi 54–56% Trump; PredictIt about 55% Harris. [secondary, Axios and Newsweek] https://www.axios.com/2024/11/04/polling-models-trump-harris-win-2024-election\n- Poll-based models: Silver Bulletin 50.4% Trump; FiveThirtyEight 53% Trump; Economist 50/50 (same Axios page).\n- Crane & Strumpf: through the fall, markets priced Trump 5–10 points above leading poll models. [part]\n- PS Special Issue: popular-vote predictions \"almost evenly split\"; among models giving electoral-college forecasts, 5 predicted a Trump win and 3 a Harris win. https://www.cambridge.org/core/journals/ps-political-science-and-politics/article/introduction-to-forecasting-the-2024-us-elections/3B0CD7678126F613C38AA011D3C034E2 [part]\n  - The state-level model of Enns et al. gave Trump about 73%.\n  - PollyVote gave Harris 56% in the electoral college, with trial-heat poll methods favouring Harris and \"alternative measures\" less so.\n- Trump won; I did not source that separately. Here the market and the fundamentals models pointed the same way, away from the polls. That suggests markets are not simply copying polls.\n\nCase E: Iowa caucuses 2016. Markets and polls both favoured the wrong winner. [secondary]\n- Trump had roughly a 70% PredictIt price; Cruz won.\n- https://www.igamingnews.com/article/predictit-aims-to-find-out-if-political-bettors-could-be-more-accurate-than-pollsters-216448/\n\n---\n3. POLITICAL-SCIENCE MODEL INPUTS AND SHARED INPUTS\n- Standard political-science models use presidential approval, economic growth or indicators, and incumbent-party time in office. Many also add trial-heat polls.\n  - Mongrain & Stegmaier 2024: the Tien/Lewis-Beck model \"relies solely on presidential approval and economic growth\", while \"many models also rely on polling information.\"\n  - Source: PS introduction (URL in Case D) [part].\n- 2016 inputs, from the Sabato's Crystal Ball model descriptions [part]:\n  - Erikson & Wlezien: leading economic indicators plus the incumbent's trial-heat poll share.\n  - Campbell: trial-heat polls plus the economy.\n  - Holbrook: national conditions plus trial-heat polls.\n  - Norpoth: primaries plus the electoral cycle.\n  - https://centerforpolitics.org/crystalball/the-political-science-election-forecasts-of-the-2016-presidential-and-congressional-elections/\n- Markets and polls (item 1a) show prices following public polls. Rothschild: poll data \"is public and hence should be in the information set of Intrade investors.\" Economic and approval data are public too.\n- Graefe et al., \"Combining forecasts: an application to elections\" [part]: low gains from combining models with poll projections. \"A possible reason … might be that many models already include information from polls.\" Combining poll projections with the IEM also gave small gains, \"conform[ing] to results by Erikson and Wlezien (2012b), who showed that prediction market forecasts mostly follow the polls.\"\n  - https://faculty.wharton.upenn.edu/wp-content/uploads/2014/03/16.-JSA.-Combining-forecasts-An-application-to-elections.pdf\n- So the shared channel is real: polls sit inside both the models and the market. I found no study directly measuring whether market traders read political-science models.\n\n---\n4. PUBLISHED CORRELATION OF MARKET VS EXPERT ERRORS\n- No multi-election estimate found. The closest evidence is below.\n- Graefe, Cuzán, Jones & Armstrong 2009 (PollyVote), Table 3. [part] https://politologist.com/files/publications/I.B.19-CuzanandGraefeJonesArmstrong-2009-Combining-Forecasts-for-U.S.-Presidential-Elections-The-Polly-Vote-Oct-30-2009.pdf\n  - Correlations of the daily errors of component forecasts over the last 93 days before election day.\n  - The \"expert\" component is a periodic survey of experts on American politics, only available for 2004 and 2008.\n  - IEM errors vs expert errors: .01 (2008), .36 (2004).\n  - IEM vs polls: .60 (2008), .06 (2004), .19 (2000), .35 (1996), -0.5 (1992, as extracted from the PDF text; the sign or size may be a typo).\n  - IEM vs models: .39, .43, .61, .49, .79.\n  - Experts vs models: .22 (2008), .68 (2004). Experts vs polls: .05 and -.01.\n  - The authors say \"there is no pattern of correlations between components across years\" and speculate that experts are \"immune\" to disturbances because they are surveyed only monthly or quarterly.\n  - Limits: five elections at most, and the correlations are within-election daily series, which are autocorrelated.\n- Graefe et al. (IJF/JSA paper above). [part]\n  - Of all two-method combinations, IEM plus experts gave the largest error reduction (29%). Models plus poll projections gave the smallest (17%).\n  - That implies market and expert errors are less correlated than model and poll errors.\n- 2016 same-sign error pattern. [part, Graefe 2016 paper above]\n  - Average errors over the last 100 days: prediction markets +4.8, index models +2.6, experts +2.1, poll aggregators +1.5, citizen forecasts +1.1, econometric models -1.5 (positive means overpredicted Clinton).\n  - Five of six components erred in the same direction, so \"little bracketing occurred\".\n- Chernov, Elenev & Song (UCLA Anderson summary). [secondary]\n  - They say polling averages and prediction markets are error-prone but \"uncorrelated\", and so useful together. This is a summary of a working paper I did not read.\n  - https://anderson-review.ucla.edu/prediction-markets-polls-economic-indicators-better-election-forecasting/\n\n---\nTHE THREE FINDINGS THAT BEST ANSWER \"INDEPENDENT, PARTLY DEPENDENT, OR ONE STORY?\"\n\n1. PARTLY DEPENDENT through shared public inputs: polls feed the market, and polls and fundamentals feed the models.\n   - Source: Erikson & Wlezien 2012, Electoral Studies. [part] Prices \"largely follow the polls\", with a price–poll correlation of +.86 to +.91, and prices add nothing once the final-week polls are in the regression.\n   - Supported by Duquette et al. 2014 [abstract only]. Contradicted in part by Rothschild 2009 [part] and Khan & Lieli [abstract only], who find markets carry some non-poll information.\n   - Link: https://statmodeling.stat.columbia.edu/wp-content/uploads/2024/08/Erikson-and-Wlezien-Electoral-Studies-2012-1.pdf\n\n2. NOT ONE STORY in the ordinary course: the error series of a market and an expert panel were only weakly to moderately correlated, and they combine better than most pairs.\n   - Source: Graefe et al. 2009 Table 3 [part]: IEM-vs-expert error correlations .01 (2008) and .36 (2004).\n   - Source: Graefe et al., JSA, combining paper [part]: the IEM+experts pair gave the largest error reduction of any pair of methods (29%).\n   - Links: https://politologist.com/files/publications/I.B.19-CuzanandGraefeJonesArmstrong-2009-Combining-Forecasts-for-U.S.-Presidential-Elections-The-Polly-Vote-Oct-30-2009.pdf and https://faculty.wharton.upenn.edu/wp-content/uploads/2014/03/16.-JSA.-Combining-forecasts-An-application-to-elections.pdf\n\n3. BUT THE ERRORS CLUSTER IN THE YEARS THAT MATTER: when a shared narrative or shared polling error exists, both fail the same way, and the markets are often more confident than the experts.\n   - UK 2015: the expert survey gave a 6% chance of a majority and markets gave about 6%; the Conservatives won a majority. [part: https://ora.ox.ac.uk/objects/uuid:5379074d-2e4d-4629-aeff-69976cbba74b/files/rht24wk45c]\n   - Brexit: experts put Leave at 38% in late May / early June (62% Remain in the final combined forecast), betting and prediction markets put Remain at about 73–77% on 22 June and 84–93% on the day; Leave won. [part: https://eprints.soton.ac.uk/396305/1/PSA_20EU2016_20Report.pdf and https://electionsetc.com/2016/06/23/final-combined-eu-referendum-forecast/]\n   - US 2016: the political-scientist panel (75 of 77 forecasts) and the Iowa market (+4.8 points) both overpredicted Clinton. [part: https://mpra.ub.uni-muenchen.de/83282/1/MPRA_paper_83282.pdf]\n   - Implication for your test case: treat the two reasons as correlated, with a correlation that rises in exactly the cases where the shared polling picture is wrong, so do not simply add them.\n\nPractical caveat for the test case: the strongest dependence runs through polls and through the common elite view of the race, not through political scientists placing the bets. I found no evidence on that last channel either way.\n"}