ProbCast aggregates odds across prediction-market venues and scores every move so you can tell a real repricing from thin-market noise. Everything below is split into the production methodology — the scores and labels shown across the site — and clearly-labelled experimental research still accumulating the history needed for validation.
Production methodology
How much the YES probability has moved recently, weighting recent windows more heavily (1h ×3, 6h ×1.6, 24h ×1, 7d ×0.25). A saturating curve keeps very large moves near — but never above — 100.
Reading it: High = a big, recent repricing. Low = roughly flat.
How much real market depth backs the price: log-scaled 24h volume (40%) + order-book liquidity (30%) + tight spread (20%) + price recency (10%).
Reading it: High = deep, actively-traded market. Low = thin, easy to push around.
How much weight to put on a move: 0.7 × Liquidity + 0.3 × (100 − Movement), minus penalties for thin liquidity, wide spreads, stale prices, missing data, a very young market, and imminent expiry. When a fresh order-flow reading exists, a move dominated by a single account is penalised and flagged CONCENTRATED_FLOW, while broad whale-free participation earns a small bonus. A large move on thin support (including a fragile order book) is penalised further and flagged LOW_TRUST_MOVE.
Reading it: High (80–100) = liquidity-supported. Medium (60–79) = solid but watch. Watch (40–59) = mixed conditions. Low (0–39) = treat as noise.
The composite used to rank movers: Movement × 0.45 + Liquidity × 0.35 + Trust × 0.20. It rewards large moves that are also backed by real liquidity and trust, so a thin-market spike never tops a genuine repricing.
Reading it: High = a big move you can actually lean on. This is the default sort on the Movers page.
Every move gets one headline label, derived from the scores above. It's the fastest read on whether a probability move is worth trusting.
A real move supported by liquidity and trust — ProbCast reads this as a genuine repricing worth paying attention to.
A large probability move that isn't backed by trustworthy conditions: thin liquidity, a wide spread, a stale price, or trading dominated by a single account. Easy to over-read.
Moving, but on mixed conditions — or barely moving at all. Not yet a confirmed repricing; keep an eye on it.
Heuristic flags surfaced alongside a move so you can weigh it in context.
We don't claim the market is wrong — we flag reasons the number may deserve less trust. Each failure mode below maps to a specific, observable ProbCast detector, so you can see why a price is being discounted rather than taking the discount on faith.
Liquidity Score · spread · book depth
last-trade age · stale-but-real source badge
Meaningful Move Score · reversal rate
tape whale concentration · participant breadth
Resolution-risk score
market-pairs registry · cross-venue agreement card
localized irreducibility / predictability (experimental)
probability durability · forecast half-life (experimental)
GDELT / event features (experimental)
circuit breaker · source freshness
young-market penalty
expiring-soon penalty
Experimental research
Research signals — shown on market pages only inside the “Advanced analysis” section. They do not affect the production scores above, and the historical sample is not yet large enough to validate them; treat them as context, not conclusions.
The Trust Score asks whether a number is structurally well-supported. The market regime asks the complementary, time-varying question: is this probability trustworthy right now? It is an observable Markov chain over six named, explainable states — stable consensus, trend continuation, news repricing, thin-liquidity spike, choppy/mean-reverting, and illiquid. We use named states, not a hidden model, so the read is something you can see and check.
Two layers, kept honest and separate: (1) the current stateis a real classification of the market's recent path (move size, short-vs-day reversal, spread, liquidity); (2) where it tends to go next is an interpretable transition model, Bayesian-blended with the transitions actually observed in tracked history (measured over a ~1-hour horizon, not the raw refresh cadence). The observed sample size is shown, so you can see how grounded the estimate is; with little history it is model-dominated and labelled as such.
It describes near-term price dynamics — not a prediction of the final outcome. Whether a high-reversal state actually precedes reversals or improves forecast accuracy is an open question we test under the same leakage-free, multiple-testing-corrected backtest as everything else, as resolved history accumulates.
The scores are transparent, rules-based summaries of liquidity and price movement — not predictions and not a recommendation to take any position. ProbCast provides prediction-market data and analytics for informational purposes only. Not financial, trading, betting, investment, legal, or tax advice.