How verdicts work
MarkMyCall scores public predictions against public data. These are the rules, in the order a call meets them. Last updated .
What counts as a call
A call is the author's own prediction about something real-world data can check, with a target and a time frame stated in the post. “BTC hits $150k before the end of the year” is a call. These are not:
- no time frame (“ETH is going to 10k”);
- questions (“BTC 100k by Friday?”), jokes and sarcasm;
- someone else's forecast reported or quoted (“Tom Lee says…”);
- conditionals on events that may never happen (“if the Fed cuts, SOL 500”);
- looking back at an old call (“I told you…”).
Only deadlines the author stated count toward a score. A deadline we had to infer is shown but never scored.
How a post is read
A language model reads the post and proposes each call: the asset, the line, the direction, the deadline and the author's exact words for the time frame. It never decides a verdict. Before a reading reaches anyone's record, rule-based checks must all agree with it:
- the author's words for the deadline appear in the post, and re-reading the whole time expression by rule gives the same day (the rule's date is the one stored);
- the target is a number the author wrote (with the unit they wrote), belongs to the asset named, and sits in the same clause as the deadline; the direction and "reaches" versus "closes at" must match the author's words; any "not" or "if" in the sentence, or a second time qualifier on the deadline, sends it to a person;
- the price when the post was made is looked up on an exchange, and the line is not already met, not within 1% of that price, and not off by a unit (“78” meaning $78,000);
- the asset is a large, liquid one with an unambiguous ticker;
- the model is at least 85% sure this is the author's own sincere call.
Anything that fails a check waits for a person. We measured this on 170 real posts from X, hand-labelled, 75 of them held back until the end (as of 25 October 2026, with Gemini 3.5 Flash reading): of 81 calls it published on its own, all 81 were read correctly, and it published nothing from 52 posts that only looked like calls.
How a call is settled
Price calls are settled on exchange candles. Coinbase decides and Bitstamp has to agree; if they disagree, a person decides. The window runs from the minute of the post to the end of the deadline in UTC, built from whole days in the middle and hours and minutes at the edges, so a price printed before the post can never count.
- “Hits / reaches / touches $X by…” is right the first minute price trades through the line, even before the deadline, and wrong if the deadline passes without it.
- “Will be above $X on…” is settled on the price when the deadline ends.
- “Won't reach $X” is the same rule turned around: it is wrong the moment price gets there.
A verdict decided within 1% of the line without a second exchange to confirm it goes to a person. So does any window with missing data, and any dispute. Verdicts are never edited: a correction is a new entry under the receipt, and the newest counts.
How accounts are scored
- Hit rate: right calls divided by settled calls. Void calls count for nothing.
- Rank score: the lower end of the 95% Wilson interval of the hit rate. It is how the leaderboard is sorted, so a lucky 3 for 3 cannot outrank 70 for 100. Accounts need 10 settled calls to rank.
- Confidence: a stated percentage wins (“70% chance”). Otherwise it is read from wording: “mark my words” about 95%, “will” 85%, “I think” 70%, “could” 60%. When the author hedges, the hedge wins.
- Brier score: the average squared gap between how sure the author sounded and what happened. 0 is perfect; always saying 50% scores 0.25.
- Says vs hits: average confidence minus hit rate. Sounding 95% sure and landing 30% of calls is +65 points.
Corrections and right of reply
Every receipt links the post, the rule it was held to, the prices and the source. Every receipt also has a form to say it was misread or settled wrong; a person reviews each one. If you are the author, that form is your right of reply.
What this is not
A record of past calls, not advice about future ones. Nothing here is financial advice.