Methodology

Interpret the post

For each post the AI extracts the asset, direction (bullish or bearish), whether intent is explicit or inferred, any stated horizon, and the exact text supporting it, with a confidence score. Each asset in a post gets its own prediction record.

Only context available at publication is used: the post and the thread before it. Later replies and price data never reach the model. Post text is treated as data, never as instructions.

Excluded: historical commentary, pure questions, two-sided statements, sarcasm that cannot be resolved, and unresolved conditional calls. Extractions below 60% confidence go to manual review rather than being forced into a score.

Two time axes: window and horizon

Horizon is how long after each call we check the price. Every accepted call is measured at all of them, so the same call can be right at 4 hours and wrong at 30 days:

• 1 hour · One hour after the post
• 4 hours · Four hours after the post
• 7 days · One week after the post
• 24 hours · One day after the post
• 30 days · One month after the post
• Stated · The timeframe the caller wrote; 7 days when none was stated

Window is which calls count: those whose outcome resolved in the period, so a short window holds calls that have actually matured. Within a window, recent calls weigh more (half-life = half the window; all time 180 days).

Timeframes written in the post (“today”, “this week”, “this month”) set the call's stated horizon; vague ones go to review. The share of calls with a stated timeframe is shown on each profile.

Measure every call against price

The entry price is the first observation at or after publication, never before it. Exit is the observation at the end of the horizon, within 60 minutes (10 minutes on 5-minute candles for 1h and 4h). Missing or stale prices leave the call unresolved; nothing is estimated. We also store the best and worst point along the way and the asset's 30-day volatility at entry.

Combine into a Caller Score

Each view (window × horizon × asset) scores every caller on five components:

• Hit rate (20%) · Share of directional calls that were right (neutral moves excluded)
• Follow-the-calls drawdown (15%) · Worst peak-to-trough of taking every call at equal stake (lower is better)
• Mean alpha (25%) · Average return minus the same bet on BTC over the same period
• Consistency (15%) · Share of the window's weeks (days for short windows) with a scored call
• Risk-adjusted return (25%) · Mean alpha divided by the volatility of the assets called

Shrinkage keeps a caller with three lucky calls near the middle until the record grows; with hundreds of calls the true value shows. Callers with fewer than 15 scored calls in a view are shown with a “low sample” badge and are not ranked. The score is relative: it says where a caller stands among the accounts we track for that window, horizon and asset, not how good the calls are in absolute terms. It is never shown without its sample size, and every component is on the profile.

Returns are shown per call: “+2% avg 7d return” means each call moved 2% in the called direction over the following 7 days on average. “Following every call” adds up an equal stake on every call resolved in the window. We do not annualise returns: calls overlap, and compounding them back to back would overstate what anyone could have earned.

Repeats, corrections and integrity

Same-account, same-asset, same-direction calls within a fixed 24-hour episode count once, anchored to the first eligible post. Later posts are kept as evidence and reversals are flagged.

Claiming a profile verifies ownership only. It cannot hide, remove or change predictions. Corrections go through a logged review and every affected snapshot is recomputed under the published policy.

Known edits are tracked with timestamps. Provider-required deletions are applied, and any resulting coverage limits are disclosed on the profile.