What Moves Prediction-Market Prices

Prediction-market quotes can change when participant orders or available interest change. Information releases, beliefs, positioning and market rules can help explain a move, but the price path alone does not identify its cause. Start with the exact payout and distinguish a displayed price from an executable quote.

This guide treats potential drivers as research hypotheses. It makes no trading-volume claim or promise that one category of news, flow or timing supplies a profitable signal.

News: record the information clock

An information release can motivate order changes. Preserve the publication time, receipt time and exact content relevant to the contract’s determination rule. A revised report downloaded later cannot automatically be used at an earlier decision time.

Measure the observed reaction rather than assuming prediction markets universally move before other financial indicators or within seconds. Compare synchronized records with known coverage; a last-trade timestamp can differ from a current quote.

A news-response hypothesis should define the permitted input and later evaluation window in advance. An apparent lead can come from clock errors or missing observations. Earlier receipt also does not establish an executable opportunity or a profitable order.

Polls: signal, but not the way people assume

A poll or forecast measures a different object from a contract quote. Define how the sample, question and outcome map to the selected market. Do not assume an observed quote movement was caused by one release merely because their timestamps are close.

A survey estimate and a tradable price have different participants, sampling and measurement conventions. Preserve those differences and the forecast version without inventing a universal demographic description of traders.

An external-feature model can be investigated using known-at-decision-time inputs and held-out outcomes. A residual from a fitted price-reaction model is a measurement under assumptions, not an automatic trade signal.

Orders and liquidity: inspect the actual book

A large request relative to resting quantity can change available prices. Record whether the observed event was a trade, addition, cancellation or quote update. The public price path does not reveal a participant’s motive.

Depth and spread help describe the available book, but cannot by themselves establish informed flow, manipulation or a change in true probability. Preserve quantity and timing and examine the selected instrument rather than applying a universal thin-versus-liquid conclusion.

A calibration result from one dataset cannot be transferred to every venue, political contract or horizon. Do not turn an unverified study coefficient or confidence interval into a fixed adjustment to a current quote.

Repeated observations can show whether a move persists in the sampled window. That remains limited evidence about a path, not a reliable classifier separating informed trades from distorted quotes.

Event timing: preserve the payout and horizon

A later observation may have more available information, but the last pre-resolution quote is not guaranteed to be the most accurate. Distinguish trading closure, event outcome, determination and settlement. Scalar and exceptional payouts need their own model.

To investigate horizon-dependent calibration, specify a sample, quote convention, prediction time and final outcomes. A price of 0.70 does not universally imply a true probability of 0.83. Preserve independent-event grouping and uncertainty instead of hardcoding a purported bias.

Sentiment: real but slippery

If using a sentiment feature, define its source, timestamp, construction and relationship to the hypothesis. A prevailing narrative is not directly observed probability. Do not prescribe wider risk bands merely because an unlabeled sentiment measure changes.

Turn candidate drivers into measurable questions

Choose an explicit question and preserve the inputs and assumptions that could test it. The list below describes possible investigations, not a completed profitable strategy or recommended allocation:

• News: compare known publication/receipt times with synchronized quote observations.
• Forecasts: preserve the sampled question, version and mapping to the market outcome.
• Order flow: distinguish trades, cancellations and quote changes under the actual protocol.
• Horizon: compare consistently sampled prices with actual outcome rules.
• Sentiment: define the feature and retain its construction and timing.

Banger’s public beta supports paper research. Its actual on_market_tick(market, tick, ctx) callback can log delivered normalized observations, and current CLI account setup is required before a local run of your existing strategy.py. This is different from a promised external-news connector or a real execution feature:

pip install bangertrades
banger run strategy.py --paper

Start with a bounded no-order observation when inspecting the feed. Preserve missing and stale data, exact identifiers and completed shutdown. When adding simulated orders, review the packaged non-crossing assumed fills and fee model; paper data cannot prove real queue position, returns or a trader’s information advantage.

One caution on calibration

A price-implied probability is a convention under a specified payout and cost model. Calibration needs observed outcomes and a justified sampling procedure. Quote movements can motivate useful questions without establishing that markets are universally accurate, that a particular trade is informed or that the proposed signal is profitable.