I want to tell you why this exists, because I think it matters.
Credence Sports didn't start as a business idea. It started with baseball and a long-running frustration that the gap between what I knew about a game and what I could actually do with that knowledge was too wide.
It Started with Baseball
I've been a baseball fan long enough to know that the game rewards the patient observer. Baseball is the sport that invented sabermetrics. It's the sport that spawned Moneyball, that gave us WAR and FIP and xwOBA, that has more publicly available data than any other sport on the planet. Pitch-level Statcast data, spin rates, exit velocity, expected batting average on every ball in play; it's all out there, free, updated in real time.
I used to spend time on Fangraphs the way other people browse sports now. Looking at a starting pitcher's xFIP compared to his ERA to see if the market was pricing in luck or skill. Checking a lineup's platoon splits against a specific arm. Trying to understand whether a team's recent offensive output was signal or noise given their underlying contact quality metrics.
And then I'd think: I know something is here. The line doesn't reflect what I actually know about this game. But I had no rigorous way to turn that knowledge into a decision. I was still operating on intuition dressed up with data. Which is better than pure intuition, but not by as much as it should be.
The Vegas Problem
The thing about going to Vegas and betting on baseball, or betting anywhere really, is that it forces a confrontation with your own reasoning that casual fandom doesn't. When you have money on a game, you can't hide behind "well, I thought they'd win." You have to answer a harder question: how confident are you, and is that confidence justified by the evidence?
I found that I could justify my picks with data. I could point to Fangraphs and say here's why I like this pitcher tonight, here's why this offense has been running hot. But I couldn't answer the follow-up question: is the line already pricing in everything you just said? And if so, by how much?
Edge isn't knowing that a pitcher is good. Edge is knowing that the pitcher is better than the market currently believes, by a specific quantifiable amount, with an honest accounting of how uncertain you are about that estimate.
That's a completely different problem. And it's one that no amount of Fangraphs or Baseball Savant browsing can solve on its own.
Why Bayesian Statistics
I want to be honest about something. Most people who build sports models reach for frequentists tools first: regression, classification models, cross-validation accuracy metrics. I went a different direction, and I want to explain why, because it's not arbitrary.
Frequentist statistics ask: given that the null hypothesis is true, how surprising is this data? It gives you p-values and confidence intervals that have a specific, often misunderstood technical meaning. It treats probability as a property of repeated experiments, not of individual events.
Bayesian statistics asks a different question: given this data, how should I update my beliefs? It treats probability as a measure of rational confidence, exactly what you need when you're trying to decide how much to trust a model's output on a single baseball game that will happen once and never be repeated under identical conditions.
For sports specifically, Bayesian inference just makes more sense to me. Here's why.
Every morning our model processes information about a game: pitching metrics, lineup quality, ballpark factors, bullpen availability, weather conditions. It combines these signals to produce a probability estimate. But that estimate isn't a fact. It's a belief, and beliefs come with uncertainty. A Bayesian model forces you to be explicit about that uncertainty rather than collapsing it into a point estimate that implies false precision.
When you see a pick on Credence Sports. you don't just see "58.3% probability." You see an 80% credible interval, the range from 48% to 61% within which the true probability is likely to fall. That range is the honest answer. The width of that range tells you how confident the model actually is. A narrow interval means the signals align and the estimate is tight. A wide interval means there's genuine uncertainty and you should size your bet accordingly.
No frequentist model produces that output naturally. A Bayesian model produces nothing else.
There's also the small-sample problem. Baseball seasons are long but individual matchup histories are short. A pitcher with a 1.21 ERA after three starts is almost certainly not a true 1.21 ERA pitcher. A Bayesian model with sensible priors handles this automatically. It pulls extreme early-season observations toward realistic baselines, weighting them appropriately given how much evidence has actually accumulated. You don't have to tune this manually. It emerges from the math.
And when data is missing - when a lineup hasn't been confirmed, when a bullpen usage figure is unavailable - a Bayesian model can incorporate that uncertainty into its output rather than silently ignoring it. Our signal completeness score on every pick tells you exactly how much information was available when the model generated its estimate. If the completeness is low, the interval is wider and the conviction score reflects it. The uncertainty is visible, not hidden.
What We're Actually Building
Credence Sports runs on a gradient-boosted ensemble model, built on XGBoost for win probability and NGBoost for run-based markets, that has been systematically enriched with Bayesian components at every stage where they add genuine value over pure machine learning.
The distinction matters and is worth being precise about. We're not running a full Bayesian generative model end to end. What we've built is more honest than that claim would be: a well-tested gradient boosting foundation where Bayesian reasoning handles the problems it's actually best suited to solve, rather than being applied everywhere for the sake of brand consistency.
Here's concretely what that means.
Empirical Bayes stabilization sits at the input layer. Before any game-level feature reaches the model, player and team statistics are smoothed using Empirical Bayes priors, estimates that pool information across similar player and contexts to. produce more reliable measurements from limited data. A pitcher with a 1.21 EAR is not a true 1.21 ERA pitcher. Empirical Bayes priors pull that estimate toward a realistic baseline, weighted by how much evidence has actually accumulated. This applies to park factors, lineup quality, starter performance, and bullpen metrics; everywhere small samples could otherwise produce overconfident features.
Sequential posterior updates run end of day. After each night's games resolve, the model updates its beliefs about player and team quality using the new evidence. These aren't dramatic swings; they're principled small adjustments that accumulate over a season into a meaningfully more accurate picture of what each team is actually capable of. The posteriors from yesterday's games feed directly into today's features.
Distributional outputs from the sub-models mean the system tracks uncertainty rather than just point estimates. The run differential model outputs a mean and a standard deviation. The total runs model outputs parameters of a probability distribution. These parameters propagate through the prediction pipeline so the final output carries honest uncertainty information rather than false precision, which is what enables the credible interval you see on every pick.
Market blending in the probability layer combines the model's independent probability estimate with market-implied probability using a calibrated alpha weight. This is Bayesian in the technical sense - it's a principled update of a prior belief given additional evidence - and it's what distinguishes our output from a model that either ignores the market entirely or takes it at face value.
What we haven't shipped yet, and won't claim to have shipped, is a full hierarchical Bayesian generative architecture. We built one, tested it rigorously, and it didn't clear its own evaluation gates on 2026 data. Rather than deploy a model that doesn't meet our standards and quietly hope users don't notice, we closed that path and returned to the foundation. The Bayesian components that are live are live because they earned it. The ones that haven't yet aren't.
A pick reaches the dashboard only when it clears five simultaneous criteria: edge above the minimum threshold, the full credible interval above the market line, signal completeness above 0.80, conviction score above the gate threshold, and CLV meta-model confirmation where sufficient historical data exists. When a pick doesn't qualify you can see it in the EV tracker anyway - every market, every edge calculation, every stake recommendation. We don't gate the data. We only gate what we're willing to call a qualified pick.
What Makes This Different
The honest answer is transparency, in a specific and meaningful sense.
There are platforms that give you picks. There are platforms that give you percentages. There are platforms that show you sharp money movement and consensus data and beat writer opinions. What almost none of them give you is the full output of the underlying model: the uncertainty range, the signal breakdown, the gate criteria, the reasons a specific pick was or wasn't selected.
We give you all of it. Not because transparency is a marketing claim, but because a model that hides its uncertainty is a model you can't actually evaluate. And a model you can't evaluate is a model you're taking on faith. We're not asking you to take anything on faith. We're showing you the math.
Action Network tells you what other people think. Credence Sports tells you what the evidence supports, how confident the evidence warrants being, and exactly what would have to be true for the model to be wrong.
That's a different product. It requires different infrastructure, different modeling choices, different design decisions. It also requires a different kind of intellectual honesty - the willingness to say, in public, that a model didn't pass its own evaluation criteria and is therefore paused. That a pick with positive edge but a wide credible interval gets a LOW conviction score rather than being dressed up to look more certain than it is.
Why "Credence"
The name was deliberate.
In Bayesian epistemology, a credence is a degree of belief, and probability assigned to a proposition based on available evidence. When the model processes tonight's pitching matchup, it's computing credences. When it updates on lineup confirmation, it's updating credences. The mathematical framework underlying every pick on this platform is literally a system for computing and revising credences.
But credence also means something in ordinary language: credibility, trustworthiness, the quality of deserving to be believed. As in "lend credence" to an argument.
The name works in both registers at once. A sharp bettor who knows Bayesian statistics reads "Credence Sports" and immediately understands what kind of platform this is. A casual bettor reads "Credence Sports" and understands the promise - a platform that has earned the right to be believed because it shows its work.
We're trying to be both things simultaneously. That's the whole project.
Who This Is For
If you want a simple lock of the day with a compelling narrative about why a team is due, this probably isn't the right platform for you.
If you're the kind of person who spent time on Fangraphs trying to understand whether a pitcher's ERA was telling the truth, who looked at a betting line and wondered whether it was actually accounting for the underlying performance metrics, who found the Moneyball story interesting not because it was a movie but because it was proof of concept for evidence-based decision-making in a domain dominated by intuition - this is built for you.
It's also built for people who are curious about Bayesian statistics and want to see what it actually looks like when you apply it seriously to a real domain with real stakes. Every pick comes with the full model output. The credible interval isn't decoration - it's the most important number on the screen.
Where We Are
Credence Sports is in beta. The head to head model is live and in active evaluation. We're tracking CLV, monitoring for signal drift, and running the full evaluation framework against every pick the model generates.
We started with baseball because it's the sport with the richest publicly available data and the deepest analytical tradition. The same framework extends naturally to football and basketball, and that expansion is on the roadmap for the fall.
For now: if you're a beta tester, thank you for being here early. The model gets better with more data, the platform gets better with more feedback, and the methodology improves with more scrutiny. All of that is welcome.
If you're reading this and you're not a beta tester yet, reach out at charlie@credencesports.com. We're selective about who gets early access, but not because we're trying to be exclusive. We're selective because early feedback shapes what this becomes, and we want it to come from people who care about the same things we do.
Credence Sports is a product of Penumbra Partners.
We operate in the penumbra - the space between certainty and ignorance where careful reasoning has an edge.