Today, almost anyone can feed basketball, soccer, baseball or any other sport's statistics to an AI agent and produce a metric, a ranking, a prediction and a report explaining it all.
The language can sound professional. The interface can look analytical. Neither establishes whether the underlying reasoning is useful.
For me, the important question is what happens next. Where does that metric go? What decision does it support? How do you find out whether it contributed anything?
Those questions connect the work we do at Statbase.
We build player and team metrics to organize different parts of the evidence: performance, form, workload and competitive context. Those metrics then feed StatPredict, where they contribute to assessments of teams, matchups and possible outcomes.
The prediction is therefore connected to the analytical system behind it. It can be traced back through the inputs and assumptions that produced it, and subsequently compared with what happened on the field.
That is a considerably harder test than asking an AI agent to explain why its answer sounds reasonable.
If your moat consists of a synthetic metric assembled from public data, familiar formulas and a new acronym, I have bad news: someone else can generate another one tomorrow.
We build composite metrics too. Their names are not our argument for their value. Their value depends on what they help us understand, calculate and test.
A metric can be useful as a description of recorded performance. Using it to recommend a lineup or forecast a game introduces further requirements.
Take nine selected hitters. They can be arranged into 362,880 batting orders, all with the same sum of individual WAR values. That sum alone cannot choose the order expected to produce more runs. You need an objective and a model that can distinguish the alternatives.
The same principle applies to forecasting. A collection of player scores does not become a win probability merely because you run a simulation.
Simulation is a calculation method. It does not supply the baseball 'reality-check' or validate the assumptions.
StatPredict calculates its forecasts from the metrics and inputs behind our analysis. Simulation explores possible outcomes under that specified model. The usefulness of the result depends on the quality of the evidence, the relationships we model and how well the framework holds up when tested.
We have results against which to examine that work.
In our internal audit of the available May-August 2026 French D1 archive, StatPredict correctly identified 60 of 80 on-field winners: 75%.
A July 27 forecast, frozen before the closing month’s 22 games, selected 18 eventual winners: 81.8%. That result is recorded in our regular-season review.
In that closing cohort, a standings-based benchmark using the records available at forecast time selected 14 winners. StatPredict selected four more.
The closing games are part of the broader archive. These are defined historical samples, with recurring teams and doubleheaders, and they need to be interpreted accordingly. They provide evidence about this use of our framework; they do not certify every metric or establish a permanent predictive advantage.
But they give us something concrete to defend. Saved forecasts. Specified inputs. Actual results. Comparisons that can expose where the additional modelling helped, and where it did not.
That connection between our metrics, the model and subsequent outcomes matters far more to me than having a distinctive number to put on a player card.
It also leads directly to a question we have been hearing from teams.
Someone told us their team knew our work, and remembered it as the platform that said they were going to lose.
As a former player, I understand the reaction. You put work into preparing for a game, and then encounter a number that appears to discount your chances before the first pitch.
StatPredict was never intended to tell a team to surrender to a percentage.
Its purpose is to help people understand the situation well enough to prepare for it.
The favourite shown at the top of a report is the most immediately visible result. The analysis underneath asks why that advantage exists and what could change it.
How dependent is the assessment on the assumed starter? What happens when the matchup reaches deeper into the pitching staff? Which roster changes materially affect the model? Where does the available evidence suggest a vulnerability worth investigating?
A coach can use those questions to prepare for several plausible games before the actual one begins.
That is where the metrics become practically useful. They help identify which parts of a matchup deserve attention. The scenarios help examine the consequences of different assumptions and decisions.
If all someone takes from the report is “you said we would lose,” we need to make that useful work easier to see.
Baseball should make us particularly receptive to this way of thinking. A .300 hitter can be excellent while recording just three hits per ten at-bats. We know from playing that good preparation and sound decisions do not guarantee the next result.
A 30% win probability describes a meaningful possibility of winning. Across comparable cases, a well-calibrated set of 30% forecasts should see that outcome roughly three times in ten.
An underdog winning does not automatically disprove a forecast. A favourite winning does not automatically validate one. We have to examine performance across outcomes, alongside the quality of the probabilities and the assumptions behind them.
And those assumptions can be challenged before the game.
A different starter, a changed lineup or newly available information may alter the conditions on which the forecast was based. That is one reason to make the reasoning accessible. Someone working with the team may know something our records do not.
If “breaking the math” means finding a better plan than the one reflected in our assumptions, that is a conversation I want to have.
Those conversations also reveal what the system needs next. Sometimes the limiting factor is the model. Sometimes it is the evidence.
The ecosystem was never a buzzword
Our Czech Extraliga regular-season analysis covered 140 games and 10,936 plate appearances, yet the feed contained no usable pitch-velocity or exit-velocity values.
The volume of recorded events could not supply measurements that were absent.
That boundary matters when interpreting a player assessment. It also explains why we are building several connected parts of Statbase. That’s basically why we built the Diamond Vision.
StatIndex makes player records, history and assessments usable. It gives people a place to examine the evidence behind a comparison and follow an athlete over time.
When the question requires evidence beyond the recorded statistics, Diamond Vision, available to selected early access members, brings our work into analysis of game footage. The observations and measurements we claim from that footage carry their own validation requirements.
StatPredict puts selected parts of the analytical framework to work in forecasts and scenarios. Subsequent results help us investigate the framework’s strengths, limitations and missing information.
That is the connection we mean when we talk about an ecosystem: evidence informing metrics, metrics supporting analysis, analysis supporting decisions, and outcomes giving us something to learn from.
The word was never intended to make a collection of features sound bigger.
We use AI throughout this work. We welcome tools that help us build, inspect and communicate more effectively. Faster development makes it possible to attempt more. It also makes it easier to publish claims before understanding what they require.
Vibe-code the prototype if you want. You still have to answer for what it tells an athlete or a coach.
We are funding Statbase ourselves and putting years of work, relevant experience and our professional reputations behind it. That is what being all-in means to us: maintaining responsibility after the demonstration, after publication and after the result.
Our experience helps us ask questions. Our methods have to make the answers testable. Our results have to remain available when they become uncomfortable.
We want younger athletes to have tools we would have valued as players. We want coaches to find useful preparation beneath a headline probability. And we want teams to challenge our work with better information, sharper questions and performances that force us to learn.
Who we personally want to win should have no influence on the probability we publish.
Math has no favourites. Our models have assumptions. Teams have choices.
The forecast is ours to defend. The game is yours to play.
