# Prediction markets — methodology **The dataset:** `prediction-market-mentions.csv` — 2,623 settled Polymarket "What will ⟨Trump⟩ say at ⟨event⟩?" markets across 224 speeches, 2024-06 to 2026-08. Every row is a market that closed and resolved, and that had a real traded price before it closed. --- ## Where the numbers come from **The venue.** Polymarket, harvested through its public Gamma and CLOB APIs under the tags `mention-markets` (100343) and `trump-speech-mentions` (105481). **The universe, and how it narrows.** Reported honestly, because the sample is a choice and every step of it throws something away: | slice | n | |---|---| | all settled word-markets, all speakers | 19,715 across 1,066 events | | Trump only | 8,558 across 445 events | | real single speeches (not "this week", not a date range) | 3,803 across 314 | | **…that also have a real pre-close price — this dataset** | **2,623 across 224** | The last cut is the one that matters: a market with no traded price cannot tell you whether the price was any good, which is the entire question. **The price** is `price_1h`, the CLOB midpoint one hour before the market closed. One hour, not the close, because the last minutes of these markets often trade on the speech already happening. `price_24h`, `price_close` and `price_first` are included so you can check whether that choice matters. It does not: the calibration result holds on all four. **The outcome** is exact-term matching against the speech transcript. Pluralisation counts, compounds count, other inflections do not. `threshold` is how many times the word had to be said — most markets are 1, some are "3+ times". Resolution is mechanical, which is why this corpus is usable at all: nobody is adjudicating intent. --- ## The three claims in the video, and how to rebuild them **1 · The market is calibrated — the price is right to about two cents.** Bucket by `price_1h`, then compare each bucket's mean price to the share of its markets where `resolved_yes = 1`. ```python import pandas as pd d = pd.read_csv("prediction-market-mentions.csv") edges = [0, .05, .10, .20, .35, .50, .65, .80, .90, .95, 1.001] d["bucket"] = pd.cut(d.price_1h, edges, right=False) g = d.groupby("bucket").agg(n=("price_1h","size"), price=("price_1h","mean"), actual=("resolved_yes","mean")) g["gap"] = g.actual - g.price print(g, "\nmean |gap| =", g.gap.abs().mean()) # 0.0204 ``` **The binning changes the headline, so it is stated rather than hidden.** These edges give a mean absolute gap of **0.0204**. Ten even bins give **0.0234**; fifteen give 0.0231. All of them support "about two cents"; none of them is the only defensible choice. **The honest caveat: the coin flip is the worst band.** The 319 markets priced between 45¢ and 55¢ happened just **43.3%** of the time — off by 6.4¢, three times the average. That is not a footnote, it is the setup: it is exactly where the fee is largest. **2 · The fee is a parabola that peaks at 50¢.** Both venues charge on `price × (1 − price)`. Kalshi publishes `0.07 × contracts × P × (1−P)`; Polymarket's "Mentions" category peaks at 1.56% of notional, which is the same curve with a coefficient of 0.0624. The `fee_cents_per_contract` column is that second one evaluated at each row's own price. It peaks at **1.56¢** at exactly 50¢ and falls to nearly nothing at both ends. **Where the trading is.** By market count this corpus is U-shaped — the 0-10¢ and 90-100¢ bins hold the most markets. **By volume it is centre-weighted: 60.5% of all trading sits between 25¢ and 75¢**, where the fee is still at least 75% of its peak. A market priced at 3¢ is settled, and nobody trades a settled market. "Where everybody trades" is a claim about money, so it is measured in money. **3 · A calibrated price plus a fee is a negative expectation, and it compounds.** If the price is the true probability, both sides of the bet are worth the same and expected profit is exactly zero: `p(1−p) − (1−p)p = 0`. Then subtract the fee. At 50¢ with fee `f` you stake `0.5 + f` to win 1, so a bet returns `+1/(0.5+f) − 1` with probability 0.5 and `−1` otherwise. At the Mentions rate (`f = 0.0156`) the per-bet mean is **−0.0303** with sd **0.9697**, so ``` P(ahead after N bets) ≈ Φ( −0.0312 · √N ) ``` | bets | you are ahead | |---|---| | 100 | 37.8% | | 1,000 | 16.2% | | 10,000 | **0.090%** — about 1 in 1,100 | Cross-checked against a 40,000-path simulation at the Kalshi rate: theory 13.42% vs simulated 13.23% at N = 1,000. --- ## What we looked for and did NOT find Published because a methodology that only lists its wins is an advertisement. - **No edge survives costs.** Selling every market indiscriminately returns +4.25% ROI (+0.0208 per contract). At a 2¢ spread that is +0.0008 — nothing. - **The favourite–longshot bias is not in this data.** Fading longshots (p < 0.10) returns +2.41%, *worse* than selling indiscriminately. The classic story does not hold here. - **A "+21% selling the joke" result was leverage, not skill.** Profit per contract is flat at ~2¢ across the price range while ROI% swings from 2% to 19%, purely because a cheap NO position divides by a small number. Always read profit per contract next to any ROI%. - **Residual YES-overpricing is not significant.** −2.4 points, but markets inside one speech are correlated: clustered p = 0.060, bootstrap CI [−0.049, +0.001]. - **A real model lost to the price, 30 times out of 30.** A latent-exposure Poisson model fitted by maximum likelihood scored Brier 0.2338 against the market's 0.1589. Even had it won, it needed to win by more than the fee. --- ## The claim this is careful about "Prediction markets always take your money" is literally false, and the video says so. **Makers pay no fee and collect a rebate funded by taker fees**, and a genuinely better-informed trader can win. The narrower claim is the one the arithmetic supports: **the taker loses by construction.** If you are tapping buy in an app, you are the taker, every time. --- ## Reuse Free to use, including commercially. A credit to **reelgorithm.py** is appreciated but not required. Prices and outcomes are Polymarket's; the derived columns and the analysis are ours. Questions or a correction: reply to any newsletter email.