One of the first questions people ask when they see SaltyTrack is: “How does it know?” You open the extension, two fighters you’ve never heard of are about to throw down, and SaltyTrack is already telling you who’s probably going to win — sometimes with 99.8% confidence. How?
This post is the high-level answer. We’re going to walk through what goes into our prediction model, what comes out, how accurate it actually is, and where it falls short. We’re not going to reveal the exact recipe — that’s our competitive edge — but we believe in transparency about how well it works and, just as importantly, when it doesn’t.
Every SaltyBet fight is AI vs AI. The characters are MUGEN fighters with their behavior coded into them. They don’t learn, they don’t adapt, they don’t have good days or bad days. A fighter’s AI is the same every single time it steps into the ring.
That’s what makes prediction possible. If a fighter has won 87% of their last 200 matches, that’s not a streak — that’s a pattern. And patterns are what machine learning is built for.
We trained a prediction model on historical match data — specifically a subset of the most recent matches. We experimented with training on the full 570,000+ match database, but found that recent data is more predictive than old data. Fighters get rebalanced, the roster evolves, and what happened three years ago matters less than what happened last month. The model looks at a range of factors about each fighter and the matchup between them, then outputs a prediction with a confidence level, that’s what you see in the extension overlay. We re-train the model on a regular basis to ensure prediction accuracy is optimal.
We’re not going to list every input — that’s our secret sauce — but at a high level, the model evaluates:
Fighter track records. How has each fighter performed historically? Win rate, match volume, and consistency all factor in. A fighter with an 89% win rate across 200+ matches is a very different proposition than one with an 89% win rate across 9 matches.
Matchup context. Not every fight happens in a vacuum. Have these two fighters met before? How did it go? Head-to-head history is one of the strongest signals we have, when it’s available.
Tier dynamics. Different tiers behave differently: S-tier is relatively predictable, A-tier is chaos. The model understands these patterns and adjusts accordingly.
Recency. Is a fighter on a hot streak or in a slump? Recent performance can indicate whether something has changed — maybe the roster was updated, maybe the matchmaking is hitting different opponents.
The model weighs all of these factors (and others) together to produce a single prediction: which fighter is more likely to win, and how confident we are in that call.
Not all predictions are created equal. Some fights are near-certainties, others are genuine coin flips. We break our predictions into four confidence levels so you know how much weight to give each one:
Lock — We’re nearly certain! The data overwhelmingly favors one fighter. These are the fights where a dominant S-tier veteran is matched against someone with a losing record, or where head-to-head history is extremely one-sided.
High — Strong signal: the data clearly points to one side, but there’s enough uncertainty that upsets happen. These are good bets, but not guaranteed.
Medium — We think we know, but the margin is slim. Both fighters have reasonable cases. This is where smart bet sizing matters — don’t go all-in on a Medium.
Low — The model is finding some signal, but not much. Both fighters look roughly even, or we don’t have enough data on one or both of them. These are the fights where you either bet small or skip entirely.
We track every prediction we make and score it against the actual outcome. No cherry-picking, no asterisks. Here’s where we stand across the last 90 days (34,777 scored matches):
| Confidence | Accuracy | Matches |
|---|---|---|
| Lock | 99.8% | 998 |
| High | 86.8% | 9,695 |
| Medium | 70.1% | 15,138 |
| Low | 56.0% | 8,946 |
| Overall | 72.0% | 34,777 |
Let’s talk about what these numbers mean in practice.
Lock at 99.8% — Two wrong out of 998. When the model says Lock, it’s essentially a sure thing. These don’t come up often (roughly 3% of matches), but when they do, the data is overwhelming.
High at 86.8% — Nearly 9 out of 10. High confidence predictions cover about 28% of matches and are correct the vast majority of the time. This is where consistent, disciplined betting pays off.
Medium at 70.1% — Seven out of ten. Medium is our largest bucket (44% of matches) and the one where bet sizing matters most. You’ll win more than you lose, but the margin for error is real.
Low at 56.0% — Six points above a coin flip. The model is still finding signal even when it’s least confident. Over thousands of bets, 56% vs 50% is meaningful — but match-to-match, it’s noisy.
Overall at 72.0% — Beating the crowd’s 67.9% by 4.1 points. That gap compounds. Over 1,000 bets, a 4-point edge is the difference between growing your stack and treading water.
Known fighters with deep history. When both fighters have 100+ matches in the database, the model has a lot to work with. Win rates stabilize, matchup patterns emerge, and the predictions get sharp. This is why S-tier and X-tier tend to be the most accurate — those fighters have been around.
Lopsided matchups. When a dominant fighter faces a weak one, the model catches it quickly. These generate Lock and High confidence predictions, and the accuracy shows.
Head-to-head repeats. If two fighters have met before multiple times and one always wins, that’s gold. The model weighs head-to-head history heavily because in a game where fighters never change, past results are genuinely predictive.
The prediction model is the core of what SaltyTrack does, but it’s not the only way we use AI and data to make SaltyBet more interesting.
Quick Insights on the homepage. The SaltyTrack homepage features a set of AI-powered data questions you can explore — things like “Who has the highest win rate right now?”, “What’s the most active fighter this week?”, “Which recent match had the closest odds?”, and more. These rotate through different angles on the data so there’s always something new to dig into. Think of it as a curated window into 570,000+ matches of SaltyBet history, updated continuously.
AI-generated matchup commentary, our unique “AI Insights”. When you’re looking at a matchup in the extension overlay, you’ll see a one-sentence AI-generated insight about the fight. These aren’t generic descriptions. They’re contextual takes that factor in each fighter’s history, the matchup dynamics, and what makes that particular fight interesting. Sometimes it’s a stat drop. Sometimes it’s a hype call. Sometimes it’s deadpan commentary about a fight that’s not going to be close. The personality varies, and that’s intentional — it keeps the experience fun and unpredictable, just like SaltyBet itself.
Both of these features are built on the same foundation as the prediction model: hundreds of thousands of matches of historical data, processed and surfaced in ways that make the SaltyBet experience richer. Predictions tell you who’s going to win, our AI Insights tell you why it’s interesting.
We believe in being honest about this. The model isn’t perfect, and some scenarios are genuinely hard.
New or rare fighters. If a fighter has fewer than a handful of matches in our database, we don’t have much to go on. The model can still make educated guesses based on tier placement and opponent history, but confidence drops. This is the biggest source of Low confidence predictions.
A-tier chaos. A-tier has the highest upset rate at 34.3% — one in three favorites loses. A-tier fighters are still being sorted. Some are on the way up, some on the way down, and the model can’t always tell which. If you’re going to get burned, it’s probably happening in A-tier.
When both fighters are unknown. Occasionally two fighters show up with almost no match history between them. The model does its best, but these are essentially educated guesses. You’ll see Low confidence on these, and rightfully so.
Character updates. SaltyBet’s admins occasionally update the roster — rebalancing characters, adding new ones, tweaking AI. When a character changes, our historical data for them becomes less reliable until new matches accumulate. We don’t always know when this happens, which can cause temporarily inaccurate predictions for affected fighters.
The crowd — meaning the collective betting behavior of everyone on SaltyBet — gets it right 67.9% of the time. That’s decent. Thousands of people pooling their knowledge and pattern recognition arrive at a roughly 68% hit rate.
SaltyTrack hits 72.0%. That’s a 4.1-point edge. Take us to the bank.
Four points might not sound like much, but in the world of prediction, it’s significant. Over 1,000 bets, that’s 41 additional correct calls. Over 10,000 bets — roughly a month of SaltyBet — that’s 410 extra wins. Compounded over time with consistent bet sizing, that edge builds.
The crowd also has a structural disadvantage: you don’t see which side the crowd is backing until after your bet is locked in. SaltyTrack gives you data-driven insight BEFORE you bet, not after.
We get asked about this. “What’s the algorithm? What features do you use? Can I see the model?”
We keep the specifics private for two reasons:
1. It’s our competitive edge. SaltyTrack’s value comes from having a better prediction model than what’s publicly available. If we published the exact approach, anyone could replicate it and we’d lose what makes SaltyTrack unique.
2. We’d rather show you results than methods. We track every prediction, score it against reality, and publish the numbers. You can see our accuracy on the SaltyTrack Accuracy page updated in real time. We think that transparency about outcomes matters more than transparency about inputs.
What we will say: the model is trained using machine learning on historical match data. It’s not a simple rules engine and it’s not just looking at win rates. It evaluates multiple dimensions of each fighter and matchup to produce its predictions. And it gets better as we collect more data, and is routinely refined.
SaltyTrack’s prediction model turns 570,000+ matches of historical data into actionable predictions for every fight on SaltyBet. It’s right 72% of the time overall, it’s nearly perfect when it says “Lock,” and it beats the crowd by 4 points consistently.
Is it perfect? No. A-tier is chaotic, new fighters are hard, and sometimes the MUGEN gods just have other plans. But over thousands of bets, data beats guessing every time.
Want to see these predictions working live? Install the SaltyTrack Chrome extension — it’s free, installs in seconds, and gives you AI-powered predictions, win rates, and head-to-head records right in the SaltyBet overlay. The data is already there. You just need to use it.
SaltyBet uses virtual currency only. No real money is wagered or exchanged. SaltyTrack is not affiliated with SaltyBet.