Illustrative mathematical surface · Research note
What can prediction market prices really tell us? A framework for separating collective information from market noise. This illustrative research note sets out an approach, its assumptions, and the questions left open.
Research
repositories
Get the notebooks, modeling code, and supporting datasets behind the videos as they become available.
- Full research archive
- New published projects
- Reproducible experiments
Membership opens when the first repository files are ready. No payments are being collected.
Recurring monthly subscription · Cancel anytimeA price is a starting point
A prediction market compresses many opinions into one number. A contract trading at 0.62 is often read as a 62% probability. That interpretation is useful, but incomplete: execution costs, liquidity, position limits, and resolution rules can all influence the price.
The question is not whether a market is always right. It is what the quoted price tells us, and what uncertainty we should preserve around it.
A minimal model
Begin with a binary event and a latent probability p. We observe a sequence of quoted prices, rather than p itself. A simple observation model is:
Here, q is the quoted price and ε collects observation noise and market frictions. This is a teaching model; real prices are bounded between zero and one, and errors need not be independent or normally distributed.
Methodology
- Define the event and read its resolution criteria before collecting prices.
- Record timestamps, bid–ask spreads, and available depth alongside the midpoint.
- Build a synthetic baseline with a fixed underlying probability.
- Perturb liquidity and observation noise separately.
- Compare model estimates against the baseline, keeping evaluation data separate.
import numpy as np
rng = np.random.default_rng(42)
latent_p = 0.62
quotes = np.clip(latent_p + rng.normal(0, 0.04, 200), 0, 1)
print(np.mean(quotes))
The fixed seed makes this experiment reproducible. These quotes are generated data, not a record of an actual market.
What to look for
A useful model should explain when its estimates become less certain. A narrow spread is not a guarantee of accuracy; repeated observations are not necessarily independent information. Compare calibration across many resolved events rather than judging a model on one outcome.
A sharper estimate is valuable only when the evidence supports it.
Limitations
This note proposes a framework, not a validated trading strategy. It does not report measured returns, live prices, or empirical forecasting performance. Event dependence, selection bias, and changing participation remain open modeling questions.
Further reading
The Polymarket documentation explains market mechanics. For probabilistic forecasting, explore calibration and proper scoring rules before optimizing a trading rule.
