Hypothesis
When the normalized Bayesian poll-aggregate win probability for a 2028 GOP nomination candidate diverges from the Polymarket price by >5pp, the poll-based estimate is closer to true probability than the market price, yielding positive-EV trades net of fees.
Data used
- Same poll dataset as idea
c1c4cd77-df23-48ed-993f-36a85666af78(75 GOP primary polls, 2024–2026) - Polymarket CLOB prices-history for 7 Republican 2028 nominees: ~704 hourly points per candidate, 2026-04-20 – 2026-05-20
- Market IDs (YES token, first 16 chars):
- JD Vance:
4008127555885222... - Marco Rubio:
1356545876122014... - Ron DeSantis:
1212797565073611...
Sample market prices vs poll aggregate (2026-05-20 snapshot): | Candidate | Market p | Poll share | Norm win% | Δ | |---|---|---|---|---| | JD Vance | 36.1% | 42.3% | 54.6% | +18.5pp | | Marco Rubio | 23.8% | 13.5% | 17.4% | −6.3pp | | Ron DeSantis | 4.5% | 8.5% | 10.9% | +6.5pp | | Donald Trump Jr. | 2.6% | 13.3% | 17.1% | +14.5pp |
Method
Trade rule: go LONG when $\hat{p}{\text{poll,norm}} - p{\text{mkt}} > \theta$; go SHORT (buy NO) when $p_{\text{mkt}} - \hat{p}_{\text{poll,norm}} > \theta$.
EV per dollar (LONG): $\text{EV} = \hat{p}(1 - p_{\text{entry}}) - (1-\hat{p})p_{\text{entry}} - f$, where $p_{\text{entry}} = p_{\text{mkt}} + \delta_{\text{spread}}$, $f=0.02$ taker, $\delta_{\text{spread}}=0.005$ half-spread.
Time-series consistency (t-test on 30 daily observations): - JD Vance: mean $\Delta = -4.9\text{pp}$, $t=-33.6$, $p\approx 0$ (market persistently below polls) - Marco Rubio: mean $\Delta = +10.7\text{pp}$, $t=+67.4$, $p\approx 0$ (market persistently above polls)
Result
| θ | Signals | Avg EV net | Notes |
|---|---|---|---|
| 2% | 5 | +9.8% | Vance LONG, Rubio SHORT, DeSantis LONG, Trump Jr. LONG, Ramaswamy LONG |
| 5% | 4 | +9.8% | Same minus Ramaswamy |
| 10% | 2 | +14.0% | Vance LONG (+16.6%), Trump Jr. LONG (+11.5%) |
| 15% | 1 | +16.6% | Vance LONG only |
Persistence of Vance gap: The market has lagged polls by 4–7pp every single day of the 30-day window. The AtlasIntel poll (2026-05-07, Rubio 45%, Vance 30%) caused a visible Rubio price jump (+3pp in one day) and Vance decline (−2pp) — markets respond to individual polls, not the aggregate, suggesting price dislocation vs. aggregate truth.
Why INCONCLUSIVE (not PASS)
- Poll share ≠ win probability: We normalize poll shares assuming linear mapping; real nomination probability depends on delegate math and convention dynamics.
- No convergence observation: Only 30 days of market history. The divergence is persistent but has not resolved — we cannot confirm the gap eventually closes.
- Rubio market behavior: The market repriced Rubio on a single outlier poll (AtlasIntel May 7). If outlier polls systematically move markets, the gap may not narrow toward aggregate but could widen.
- Trump Jr. anomaly: 13% poll share vs 2.6% market may reflect rational market discounting of name-recognition polls where respondents list Trump Jr. by association, not genuine support.
- Long-dated market risk: The 2028 primary is ~2 years away; holding costs and opportunity cost are not modeled.
Reproduction
source ~/.pmvenv/bin/activate
python3 /mnt/projects/tnt_85c10df4451042ca/prj_c7cb91b70b2f42ac/d6_bayesian_poll_agg.py
# ev_analysis.json written to /tmp/pm_data/ev_analysis.json
Failure mode / next step
- Killer test: Get 6+ months of market history. If Vance consistently underprices and polls remain stable, EV is robust.
- Rubio SHORT is the cleaner bet (market at 2x poll share; less explainable by non-poll factors than Vance discount).
- Add general-election electability model: markets may be discounting Vance for VP-succession narrative; Rubio priced as strongest general-election candidate.
- Check CLOB order book depth: if bid-ask is 3–5pp on these markets, the half-spread assumption of 0.5pp is too optimistic.