Methodology

Final p = logistic(logit(base) + Σ evidence + Σ conditions), clamped to 0.1%–99.9%. You may override p directly. This is an illustrative heuristic you control — not a trained or calibrated forecasting model.

Each item contributes weight × reliability × (½ if rumor) × 0.5^(age / half-life). Age is measured at the as-of date. Weights, reliabilities and half-lives are your assumptions.

Items sharing a correlation group are summed, then capped at the magnitude of the strongest single member — so five reports of one fact count once. Leave the group blank for genuinely independent signals.

Asks are executable prices, not exact probabilities; YES + NO asks usually exceed 1. All-in cost c = ask + fees + slippage. YES pays 1 if the event happens; NO pays 1 if it doesn't (use 1 − p). EV = N(p − c), win N(1 − c), loss −N·c. Outlay may not exceed capital × max allocation.

Monte Carlo samples p from Beta(mean·k, (1−mean)·k), then resolves the event once per trial: every contract wins or loses together. Widening uncertainty at the same mean does not change the single-event win chance. More trials reduce simulation noise only; they do not validate the model.

At an as-of date, evidence with a later available-at time is removed before computing — not just hidden in the UI. Outcomes and market snapshots appear only once their timestamps pass. Synthetic episode scores demonstrate mechanics, not real out-of-sample skill.

Brier score and log loss (clipped at 1%/99%) use resolved forecasts only. A benchmark is compared only when a frozen bid/ask midpoint exists or you record one. Small samples are flagged; no Sharpe ratios or return track records are shown.

All events, prices and episodes are SYNTHETIC / DEMO or typed by you. Source URLs are stored, never fetched. There is no live feed, exchange connection or real money. Everything is saved only in this browser's localStorage; clearing site data erases it.