Source-backed
Each logged pick is tied to a public source, with the relevant moment retained so readers can check what was said.
Methodology
How FightCardIQ counts analyst track records and AI predictions.
FightCardIQ logs public analyst calls, verifies their sources, and calculates results after bouts settle. Each section explains which calls count, how results are shown, what uncertainty remains, and how the AI forecasts are produced.
See the dated version history.
Included calls
A call starts with a documented public prediction, not a retrospective interpretation.
Result-derived
Metrics are calculated only after a qualifying bout result is available. The site does not let analysts edit a settled record.
Separate samples
Winner, method, round, and underdog calls have different eligibility rules, so each carries its own denominator and none of them can be added together. Method and round are counted inside the correct winner calls, not alongside them; underdog calls are a different sample again.
How each metric is counted
A correct call is a hit; an eligible incorrect call is a miss. We never replace the denominator with a bare percentage.
Winner called
Every selected winner in a resolved bout with a winner is eligible. A correct selection is a hit; a wrong selection is a miss.
Method called
Method is scored separately, only when the analyst selected the correct winner and stated a method that can be checked against the result. A wrong winner does not enter the method denominator.
Round called
Round is also separate, and it is a claim about a finish: naming a round says the fight ends in that round. It is scored when the analyst selected the correct winner and the result can settle that claim — a finish that records a round, or any decision, which refutes a finishing-round call outright. A fight that went the distance is a round miss, never a round hit, whichever round the scorecards covered. It does not require a correct method call.
Perfect calls
A perfect call is a hit where everything the analyst committed to was right: the winner, the method, and the round when they named one. Naming no round does not disqualify a call and does not count as a round hit. Naming one on a fight that went to a decision does cost the call, because the decision refutes it. Perfect calls are reported as a count against hits, never as a percentage, because their denominators are small.
Correct sportsbook underdog calls
This counts selected winners that were strict underdogs in the qualified frozen pre-event sportsbook consensus. A winning underdog selection is a hit. A losing selected underdog remains an attempted call and is a miss.
How to read a scorecard
The analyst card has four rows: Winner, Method, Round, and Underdog; this example shows the first three.
Method and round are independent details of a correct winner call, so their eligible samples sit within the winner sample. The illustrative rows below show each hit count against its own denominator. What makes a round call eligible differs from method — see Round called above.
Illustrative record · 99 resolved calls
Winner67 hits / 99 resolved calls
Hit: 67 · Miss: 32
Method34 hits / 53 eligible calls
Hit: 34 · Miss: 19 · Outside this sample: 46
Round5 hits / 11 eligible calls
Hit: 5 · Miss: 6 · Outside this sample: 88
Every row uses the same resolved-call width. Only its Hit and Miss segments are drawn; any remaining space is intentionally blank. That blank space keeps a method, round, or underdog result in proportion to the full resolved record instead of stretching a smaller sample across the whole chart. The Underdog row counts only picks with a known pre-fight price.
Where a step shows a sentence instead of a number, its denominator is zero, and the sentence says which of two different things happened. “Calls the winner, not the round” means the analyst has never named a round — a style, not a failure. “Named rounds — none on a correct call yet” means they do name rounds, but none has yet landed on a correct call whose result could settle it. Neither is shown as “0 of 0” or as a zero bar, because a zero would state something false in both cases. A genuinely earned zero — a 0 of 1 on someone who named a method on their one qualifying call and got it wrong — keeps its real numbers and is never hidden.
What these samples can and cannot show
Published so readers can judge a difference between two analysts for themselves.
On 19 August 2026, across the 26 analysts with a shown record, resolved-call counts ran from 5 to 99, with a median of 27.5. These examples round to the same 68% observed accuracy, while their 95% Wilson intervals narrow as the sample grows.
19 of 28 calls correct (67.9%); 95% Wilson interval
67 of 99 calls correct (67.7%); 95% Wilson interval
The practical consequence: at these sample sizes most differences between two analysts are not distinguishable, and a higher percentage on a smaller sample is not evidence of a better analyst. This is why every figure on the site is published with its denominator, why records below five resolved calls show no ratio at all, and why a larger sample is worth more than a higher percentage when comparing two records.
For the analysts as a group rather than one at a time, the dated analyst accuracy study reports how often their logged picks were right and whether the analyst majority beat simply picking the favourite, with the same denominators and exclusions described here.
Frozen pre-event sportsbook consensus
Underdog status is a historical market context, not a live price or a recommendation.
For a bout to qualify, the consensus must be captured before its event starts, frozen at that boundary, and be less than two hours old. The selected fighter must have a strictly lower consensus chance of winning than their opponent. Tied markets do not create an underdog.
We use the frozen pre-event consensus so later market movement cannot change a historical underdog result. The metric does not use a price at the moment an analyst made a pick.
Directory underdog row
A directory display: qualifying underdog picks and whether they hit or missed.
The analyst directory always includes an Underdog row. A pick qualifies once it has a pre-fight price: the frozen pre-event sportsbook consensus described above when one was captured, or, when it was not, a price reconstructed after the fact from Polymarket’s public trading history for that fight. The row shows only the qualifying picks that backed the underdog, split into hits and misses; it states when there are no qualifying underdog calls.
This is a wider net than Correct sportsbook underdog calls, which uses the sportsbook consensus only. The row uses this wider net; it remains separate from the sportsbook-only underdog call count.
Exclusions and thin samples
Not every logged row belongs in every metric.
Excluded from Track Record
- Upcoming calls and rows without a selected winner.
- Draws, no-contests, cancellations, and results without a winner.
- Method or round details that are not eligible for their own metric.
- Underdog candidates without a qualified, strictly unequal frozen consensus.
Thin samples stay visible
Small samples are not converted into an overconfident verdict. Analyst pages show real counts, and detailed result splits appear only once the relevant metric has enough eligible calls to be useful.
Winner accuracy
The main comparison measure uses only correct winner calls and resolved winner calls.
Winner accuracy is the number of correct winner calls divided by all resolved winner calls. The rankings board orders analysts by exact hit-rate fractions; each displayed percentage keeps its record denominator beside it. Analysts with the same exact hit rate share a rank. Within a tie, the larger sample appears first; equal-sized samples use a consistent order.
A ranking position requires at least 20 resolved calls. Records with 5 to 19 calls remain listed with their Winner accuracy but without a rank or uncertainty range; records below five show counts without a percentage. Each ranked record also shows a 95% Wilson interval to make sample uncertainty visible. The interval describes uncertainty in the observed hit rate; it does not measure skill or performance against betting markets.
The analyst directory ranks each selected metric within its selected period. Winner minimums are Last event: 6; Last 5 events: 10; Last 20 events: 20; All time: 40. Method, Round, and Underdog minimums are Last event: 3; Last 5 events: 5; Last 20 events: 10; All time: 20. Each metric qualifies on its own denominator, rather than on Winner call volume. Below-minimum records retain their percentage and exact sample without a rank; a period can have no qualified analysts for a metric. These minimums do not guarantee reliability. The separate Winner accuracy board above retains its 20-call minimum.
Method, Round, and Underdog stay as separate measures on analyst scorecards. Method and Round are scored only within the eligible correct-winner calls; Underdog uses calls with a qualifying known pre-fight price.
AI predictions
ChatGPT (OpenAI), Claude (Anthropic) and Gemini (Google) each independently forecast every way a fight can end. The numbers are theirs; FightCard sets the rules and does the arithmetic.
What the models are given
The event name and start time, the two fighters’ names, the weight class, the number of scheduled rounds, whether it is a title fight, and the counting rules below. FightCard does not supply records, statistics, base rates or odds. Each model researches the public web itself, using only information from before the event starts, and chooses what to look at — which can include outside analysis or public odds. We do not control or check that research, and we do not publish the pages a model read.
When forecasts are made
Each model is asked again as the fight gets closer, at four checkpoints: T-14 (14.5 to 7.5 days before the event starts), T-7 (7.5 to 3.5 days), T-3 (3.5 to 1.5 days) and T-1 (1.5 days or less). “T-14” means roughly 14 days to go. Nothing is made earlier than 14.5 days out, and a forecast made at or after the event start is not accepted. A published forecast is never edited, and a model is not asked again for a checkpoint it has already answered; the page shows each model’s latest forecast made before the event start. If one model fails to deliver, the others are published and the missing one is added to the same checkpoint later, while there is still time. Forecasts made before the outcome grid existed keep their older six-number layout.
The outcome grid
- Each model gives a chance for every way the fight can end: each fighter winning by KO/TKO or by submission in each scheduled round, each fighter winning by decision, and one shared cell for a draw or no contest.
- That is 15 cells for a three-round fight and 23 for a five-round fight, and they add up to 100%. A fight scheduled for any other number of rounds gets no grid.
- KO/TKO includes referee, doctor and corner stoppages, and disqualifications. A decision includes a fight stopped early and scored on the judges’ cards. A stoppage between rounds counts as a finish in the round that just ended. A draw and a no contest share one outcome.
- Every cell is stored to a tenth of a percent, so the total is exactly 100.0%. The page rounds each to a whole percent and shows “<1%” for the very small ones.
- A fighter’s win chance is the total of all that fighter’s cells. The draw/no-contest cell is part of the 100%, so a favourite can sit below 50%.
- Small slips in a model’s numbers are corrected by code and the correction is recorded. A clearly broken forecast is left out for that one model only.
Confidence
Confidence restates the favourite’s win chance on a 10-step scale: the win chance in tens, rounded, so 55% is 6 of 10 and 95% or more is 10 of 10. 1 to 3 reads Low, 4 to 6 Medium and 7 to 10 High. A favourite usually has at least half the chance, so most picks show 5 or more. When the model has the fighters exactly even, there is no pick and no confidence level. It is not a separate judgement by the model.
Written explanations
Models also write short reasons for their three most likely outcomes and short answers to some reader questions, and are told to leave out percentages, odds, betting language and links. Wherever the page shows that text it is labelled “AI model’s own reasoning — not checked by FightCard”: it is the model’s opinion, and FightCard has not verified the facts in it. Everything else on the AI tab — the percentages, the total for any outcomes a reader selects, the three most likely outcomes, the “what if” numbers — is calculated by FightCard’s code from the model’s grid, not written by the model.
Scoring
Each model’s winner record is shown on its tab as “N of M winner picks correct” with a hit rate. For each settled fight with a winner it counts the model’s last forecast made before the event start. Draws and no contests are left out, and so are fights where a model had the fighters exactly even (no pick). The record always shows its denominator (M), and appears only once AI forecasts have been settled across at least two events. Once a model’s outcome-grid forecasts have been scored across at least two settled events, its tab also says how well it calls method and round: for three-round and five-round fights separately, the average chance it gave the actual result and a grid score (lower is better). Where the fights have one, it is compared with a simple baseline — the betting market’s winner split, spread like past UFC fights of the same length — and the answer says how many fights that comparison covers. It also shows how often the model picks the winner at each confidence level and on fights like the one shown, and whether the chance it gave the actual result rises as fight day gets closer. Every figure says how many fights it rests on.
Limits
Models research public information before the fight, such as records, styles, recent results and coverage. They cannot see inside training camps, injuries nobody has reported, or what happens on the night. A small chance is not no chance: outcomes a model rates at a few percent still happen, and big favourites lose sometimes. A hit rate over a handful of fights says little — see what these samples can and cannot show. These forecasts are not betting advice: FightCard never turns a model’s chances into betting odds and does not use tipster wording.
How versions work
A published figure should never change silently.
Rule changes
A changed metric, eligibility rule or ranking floor gets a dated version. Earlier published figures stay as they were.
Data corrections
Verified pick, result or source errors are corrected and marked on the analyst’s page. Affected figures are restated in full; corrections do not create a new version.
Version history
Version 8, 18 October 2026
Kept the analyst-directory Winner minimums at 6/10/20/40 eligible calls; lowered the Method, Round, and Underdog minimums from 6/10/20/40 to 3/5/10/20 eligible calls respectively (Last event, Last 5 events, Last 20 events, All time). Below-minimum records remain visible without rank. The separate Winner accuracy board’s 20-call rule, metric definitions, and call records are unchanged.
Version 7, 11 October 2026
Raised the analyst-directory per-metric minimums from 1/3/5/5 to 6/10/20/40 eligible calls (Last event, Last 5 events, Last 20 events, All time). Below-minimum records remain visible without rank. The separate Winner accuracy board’s 20-call rule, metric definitions, and call records are unchanged.
Version 6, 10 October 2026
Corrected the public ranking floor: a position now requires at least 20 resolved calls, rather than five. Records with 5 to 19 calls are now listed with Winner accuracy and their denominator, but no rank or uncertainty range; records below five show counts only. No call records changed.
Version 5, 9 October 2026
Winner accuracy became the default comparison, and the earlier public comparison display was retired. Call records and metric definitions were unchanged.
Version 4, 31 August 2026
Activated the existing underdog Score component and began publishing each priced pick’s market price and earned credit. The formula, weights and 20-call floor were unchanged; no Score fell and no figure was restated. Analyst positions changed.
Version 3, 20 August 2026
Published FightCardIQ Score, its formula, weights, floor and worked example; retired the statement that no overall ranking existed. No metric definition changed or figure moved; the Score used existing figures. The underdog term was defined but inactive.
Version 2, 18 August 2026
Added the analyst-directory scorecard funnel, explained both zero-denominator cases and published the sample-size arithmetic. Metric definitions and published figures were unchanged.
Version 1
Initial definitions for Winner, Method, Round, Perfect and Underdog calls.