AI Boxing Predictions — 83.3% Accuracy Exposed
The Numbers
Thirty out of thirty-six. Our AI boxing predictions are running at 83.3 per cent accuracy across recent fight cards — the highest of any sport we cover. Here is every pick, every miss, and exactly how the model works.
The headline pick was Tyson Fury over Arslanbek Makhmudov by unanimous decision, which landed exactly as called — Fury controlled the fight from the opening bell and won 120-108 on the cards. Conor Benn dominated Regis Prograis to a 98-92 unanimous decision, another comfortable win for our model. The Justis Huni pick over Frazer Clarke was the most impressive call of the week: Huni won by majority decision (95-95, 96-94, 96-94) in a fight where many analysts favoured Clarke's size and power. Our model read the data differently, identified Huni's superior hand speed and ring IQ, and backed the upset.
The wins were not limited to headline fights. Felix Cash stopped Liam O'Hare as predicted. Breyon Gorham finished Eduardo Costa Do Nascimento. Arjan Iseni outpointed Christian Figueroa 60-54. Francis Gorman beat Ryan Labourn. Mikie Tallon beat Leonardo Blanc. Norman Neely stopped Gabriel Garcia Perez. On the April 10 undercard in Newark, Hebert Conceicao Sousa beat Johan Gonzalez by unanimous decision and Francisco Veron beat Raul Garcia by unanimous decision — both correctly called. Yan Marcos won a split decision over Dwyke Flemmings Jr. on the Atlantic City card, and Chantel Navarro won her Las Vegas debut by unanimous decision over Perla Lomeli. The model called all of them.
The Six We Got Wrong
Six misses out of thirty-six graded fights. Each one teaches something. Tenshin Nasukawa's retirement TKO of Juan Francisco Estrada in round nine was our biggest miss by profile. The model picked Estrada, the experienced former champion, but Nasukawa's speed and pressure were too much. The model underestimated how effectively a fighter moving up from kickboxing could translate his skills at the top level of professional boxing.
Jeamie Tshikeva upsetting Richard Riakporhe was a genuine surprise. The model had Riakporhe, the more established cruiserweight, but Tshikeva fought the best night of his career and earned the stoppage. Simon Zachenhuber was picked to beat Pawel August but lost a close decision 56-58 — a fight that could have gone either way. Sultan Almohammed was picked over Hector Avila Lozano but lost. Daiyaan Butt was picked over Heidan Martinez Morillo but lost 90-100 on the cards. Jursly Vargas was picked to lose to Isaac Dowuona but actually won, which was a case of the model getting the underdog wrong rather than the favourite.
The pattern across the misses is revealing. Five of the six incorrect picks came from lower-profile fights on undercards where the fighters had limited professional records and thin online coverage. When our model searches for fighter data, there is simply less to find for a six-fight professional with no Wikipedia page and minimal news coverage. The Nasukawa-Estrada miss was the exception — that was a high-profile fight where the model had plenty of data but drew the wrong conclusion from it. This is an honest limitation: the model performs better on bigger fights because more information is available to search and synthesise.
Why Boxing Is Our Best Sport
Boxing outcomes are more predictable than MMA, and the data explains why. In mixed martial arts, a wrestler can take down a superior striker and win on the ground. A jiu-jitsu specialist can submit a more experienced opponent with a single mistake. Boxing removes those dimensions entirely. The fight takes place standing up, with a limited set of offensive weapons, and the better technical boxer usually wins — especially over the longer scheduled distances of ten or twelve rounds where skill gaps compound.
Judges' decisions in boxing follow more predictable patterns than in MMA. The ten-point must system means that each round is scored individually, and the cleaner, more active fighter almost always gets the nod. Upsets happen, but they tend to happen via knockouts rather than decisions, and knockout probability can be estimated from a fighter's power statistics and their opponent's chin. Boxing also has seventeen weight classes compared to MMA's eight, which means fighters are more precisely matched by size. You rarely see the kind of physical mismatch that produces freak results in heavier MMA divisions.
The result is that boxing rewards the kind of analysis our model does well: searching for fighter records, win-loss-knockout ratios, recent form, age trajectories, and stylistic matchups. A fighter who has won his last eight bouts by decision is unlikely to suddenly become a knockout artist. A fighter with a 40 per cent knockout ratio facing an opponent who has never been stopped has predictable dynamics. These patterns are exactly what a search-grounded AI model is built to identify.
How We Predict Boxing Fights
Our boxing prediction system uses Gemini AI with Google Search grounding. Before every fight, the model searches the web for both fighters' records, recent results, training camp news, weigh-in reports, and available odds. It synthesises that information into a prediction that includes the predicted winner, method of victory, confidence level, and detailed reasoning. There is no structured boxing API that provides the depth of data we need at a price we can afford, so the model relies on its ability to search, read, and interpret boxing news and statistics from across the web.
We do not apply a hard confidence ceiling for boxing the way we do for UFC (capped at 82 per cent) or NHL (capped at 80 per cent). The 83.3 per cent accuracy rate suggests the model is already well-calibrated for boxing — it is not systematically overconfident. That said, we watch the numbers carefully. If accuracy drops below 75 per cent over a sustained sample, we will introduce tighter calibration. For now, the model's natural conservatism on lower-profile fights and appropriate confidence on marquee matchups is producing strong results.
Results grading for boxing is more manual than for other sports. The ESPN scoreboard API does not cover boxing — it returns a 404 error. We use Gemini with search grounding to look up fight results from BoxRec, Tapology, BoxingScene, and ESPN's boxing results page. When results cannot be found automatically, we verify and grade them manually. Several fights this week were graded by hand after the auto-grader could not find published results for smaller undercard bouts. This is a bottleneck we are actively working to improve.
Limitations and What We're Building
Honesty about limitations is more valuable than pretending they do not exist. Our boxing prediction system has three clear weaknesses right now. First, there is no structured fighter database feeding the model. We investigated the Odds API for boxing and it works, but the free tier gives only 500 requests per month across all sports, and we burn through that quickly. A dedicated boxing data API would improve predictions on lower-profile fights where the model currently struggles to find reliable information.
Second, undercard coverage is inconsistent. Major fight cards often have eight to twelve bouts, and the lower prelim fights feature fighters with minimal online presence. Results for these fights sometimes take 48 hours to appear on any public website, which means our auto-grader cannot find them immediately. We manually graded several orphan fights this week, including bouts from the Accra Legacy Rise card and the Atlantic City ProBox TV undercard. We are exploring boxing-data.com and other specialised APIs to close this gap.
Third, the model occasionally gets the date wrong when indexing fights. Boxing cards that start at 10 PM Eastern and run past midnight create timezone ambiguity — is a fight that happens at 12:30 AM on Sunday part of the Saturday card or the Sunday card? We built deduplication logic that checks for the same two fighters within a one-day window to prevent double-counting, but the underlying indexing could be cleaner. These are infrastructure problems rather than prediction problems, and we are solving them incrementally. The boxing prediction pages show the current state of every upcoming fight, with predictions published before the first bell and results graded as soon as they are available. The same-game parlay tool lets you combine boxing picks with other sports, and the accuracy page tracks every prediction we have ever made.