No sportsbooks. No syndicates. No narrative. Just probability — derived from hierarchical Bayesian inference, weighted by recency, validated out-of-sample on the held-out 2025-26 season (— games), and disclosed honestly about what it cannot see.
POSEIDON's forecasts come from TIDAL Core, a hierarchical Bayesian Poisson regression model. We treat scoring as a probabilistic process whose rate parameters are inferred from training data, weighted by recency. Every prediction is a probability distribution — not a point estimate.
TIDAL Core models each game as two independent Poisson processes — one for each team's scoring rate — whose parameters depend on team strength, opponent strength, home advantage, rest days, and pace.
For a game between team A and team B, we estimate the expected scoring rate for each side as a function of latent team-strength parameters that the model infers from historical scores. Bayesian inference means we don't just produce a single estimate of a team's strength — we produce a full posterior distribution describing our uncertainty.
This means stronger teams produce higher expected scores, weaker opponents reduce them, home advantage adds a documented bonus, and recent rest reduces fatigue penalties. From these two scoring rates, the full distribution of possible game outcomes follows mathematically — not just "who wins" but "what's the chance of winning by 7 vs 12 vs 3 points."
POSEIDON's NBA model is trained on — games across 10 seasons (2016-17 → 2025-26), with recency weighting (a —-day half-life) so recent games count more than older ones.
Weight halves roughly every six months (a —-day half-life), so a game about a year old counts about a quarter as much and older seasons fade further still. We don't throw old data away — older seasons help anchor team-strength priors — but the model knows that the league has evolved (pace is up, three-point rates are up, defenses are different) and weights accordingly.
Excluded: preseason games (rotations are non-representative), games with missing or corrupted score data (rare but they exist), and any single-game outliers we could verify were affected by extraordinary off-court factors.
Every accuracy claim POSEIDON makes comes from walk-forward backtesting — training on data
through date T, predicting day T+1, recording the result, then advancing.
This is the closest you can get to honest forward-looking accuracy estimation.
Across the walled-off 2025-26 hold-out season — — games from 2025-10-21 to 2026-06-04, a season never used to tune the model, with every game predicted using only games played before it — POSEIDON's moneyline picks hit at —. When the model is more confident (>60% probability), accuracy rises to —; at very high confidence (>70%), to —.
Our Brier score — the standard measure for how well-calibrated probabilistic predictions
are — is —, meaningfully better than the uninformed
baseline of —. Lower is better; a perfect forecaster would score 0.
Honest forecasting requires honest disclosure of what we don't capture. Here's the list.
Player-level injuries and rest decisions. Our model treats teams as composite entities. A late scratch of a star player can shift a real-world probability by 10+ points; POSEIDON won't see that until you tell it (via the TIDAL Pulse overlay) or until the market moves and we re-anchor.
Travel, altitude, time zones. An East Coast team flying to Denver for an 8pm tipoff is observably different. POSEIDON does not model this directly — it shows up only insofar as it correlates with rest-day patterns the model does see.
Officiating and game flow. Foul rates, technical fouls, ejections, and replay review variance are not in the model. These add real-world noise we don't try to predict.
Playoffs differ from regular season. Playoff intensity, slower pace, and tighter defenses mean
the regular-season-trained model's totals predictions tend to overshoot in playoff games. We grade
totals at C for this reason.
Black-swan events. Coaching changes, mid-season trades, locker-room dynamics — all invisible to a stat-based model until enough games have happened to update the team-strength estimate.
Five thinkers whose work, properly applied, forms the actual mathematical backbone of any serious sports-betting system. Each has a real, computed contribution to Poseidon — no metaphors, no vocabulary borrowing.
Scoring events in basketball — possessions ending in points — are well-approximated as a Poisson process. Each team has a latent offensive rate and defensive rate; points scored are draws from a Poisson distribution whose mean depends on both teams plus context.
The log(λ) link makes the model multiplicative: a strong offense vs. weak defense compounds, exactly as in reality.
Gaussian priors on each team's offensive/defensive coefficients with mean zero. Pulls estimates toward league average when evidence is weak — that's what stops early-season noise from mattering.
Equivalent to L2-penalized Poisson regression with penalty λ = 1/σ², tuned via walk-forward CV.
Refit every 7 days walking forward through history. For each game, predictions use only data available before the game tipped off. No leakage. This is the only honest way to estimate live performance.
Anyone showing you in-sample accuracy on the same games the model was fit to is selling you a fantasy.
Once we have the Poisson rates, we draw 5,000 simulated outcomes per matchup to map the margin and total distributions (the p16/p50/p84 bands); the win probability itself is closed-form. Combined with Kelly criterion for sizing, we know when to bet, how much, and crucially when to refuse to bet.