Posterior.
A Bayesian almanac of the Major League season.
Latest board has 6 qualified picks across 6 games. The public current-policy ledger includes 150 actionable picks: live-tracked rows plus 132 picks from the locked May 21-June 3, 2026 backfill. Older pre-policy rows remain audit history, not part of this headline count.
§When the model says 60%, the event happens roughly 60% of the time
A Bayesian MLB odds desk
with the record attached.
Built for bettors, analysts, and builders who want model probabilities, market odds, line movement, confidence tiers, closing-line tracking, Brier score, and calibration in one accountable MLB desk.
Posterior is not a sportsbook, bookmaker, or guaranteed-picks service. It is a probability research product: every actionable entry is time-stamped, settled, and graded.
Hierarchical Bayes
Isotonic · Beta · BBQ
Graded daily, risk-referenced
One season, fully instrumented.
Posterior started as a calibration curve. It is now a working desk — prices, proof, a news wire, a personalized feed, and an open API. Everything below reads from the same nightly model.
Nightly model prices
Calibrated player-prop probabilities for every slate, de-vigged against the sharpest consensus across five books.
Tonight's board →A live calibration curve
When the model says 60%, the event lands near 60%. The reliability curve is public and refreshes nightly.
See the curve →Graded by morning
Every STRONG and LEAN settles against the market the next day — wins and losses both, never a curated subset.
The record →Ask the model
Grounded Q&A on tonight's board. It explains the picks the model made and refuses to invent ones it didn't.
Ask a question →The wire
Every club's news, pulled through the day and read by the model into structured, dated signals.
Read the wire →Your club, followed
Pick a team and players. Get a personalized desk, a filtered wire, and alerts when lines move.
Follow a team →Open data + API
Machine-readable picks and the resolved ledger. pk_live keys put the whole desk behind one request.
Developers →Live game board
Win probability and prop resolutions as the night unfolds — no login, free to watch.
Watch live →An honest tally, refreshed nightly.
Every STRONG and LEAN entry settles against the public market the morning after. The live board only shows entries that qualify for publication, while the public current-policy ledger includes live-tracked rows plus the labeled May 21-June 3, 2026 backfill. Older pre-policy rows remain audit history. Live entries, pending rows, and full model internals stay inside the subscriber almanac and API, while the delayed audit ledger is published at /data.
Perfectly calibrated (y = x)
Posterior, season-to-date
- Reading the chart.
- When the model says 0.70, the underlying event should occur roughly seventy percent of the time. The closer the curve hugs the diagonal, the more our stated confidence matches the world.
- Why it matters.
- A miscalibrated model can post a healthy hit-rate while sizing bets at the wrong confidence and bleed bankroll silently. We refit isotonic, Beta, and BBQ scalers each morning and ship the winner.
- Hover a point.
- Each bucket displays its predicted probability, the empirical hit-rate, and the sample size — the bigger the n, the more reliable the bucket.
Six architectural pillars.
We don't publish the signal recipe. We publish the architecture that contains the signals — load-bearing decisions that make every downstream choice tractable.
Hierarchical priors over partial pools.
Each batter's hit rate is shrunk toward the league mean with a strength inverse to their plate-appearance count. New call-ups don't get over-confident point estimates; veterans don't drown in noise from a slow week.
De-vig before edge.
Sportsbook lines carry a vigorish that distorts implied probability. We multiplicatively scale every market against its complement to recover a fair-line probability, then compute edge against that — not the raw quote.
Calibration as a first-class artifact.
The active production calibrator is conservative and versioned. Candidate calibrators — isotonic regression, Beta calibration, Bayesian Binning into Quantiles — are evaluated daily against resolved entries, but a candidate is only promoted after minimum-sample, holdout, and stability gates clear. Reliability is a model, not a hope.
A signal stack, one consensus.
Umpire zone tendencies, park handedness, pitcher arsenal vs. batter platoon, weather and elevation, lineup slot, bullpen fatigue, sharp money line-movement, opener vs. starter. Each contributes a small probabilistic adjustment to the prior.
Gap-graded confidence, conservative risk reference.
Gap above the STRONG threshold ships as STRONG; above LEAN as LEAN. Below-threshold observations stay internal for calibration and research. The reference size scales with the gap under 1/4 Kelly, confidence dampening, and caps — a displayed risk metric, not an instruction to wager.
Settle in the morning, in public.
Every entry resolves against the closing market the following morning. Hit-rate, units, closing-line EV, and Brier score update on the public dashboard before the kettle boils. Nothing is curated, nothing is hidden — even the losing days.
Make it your desk.
The almanac covers the whole league, but a subscriber reads it one club at a time. Pin your team, follow the players you actually watch, and let the wire do the reading.
Thirty teams. Pin the one that ruins your evenings.
Choose a favorite team and a roster of players. Your team desk collects their props, their wire, and their line movement in one place — and alerts you when a number moves or news breaks.
Pick your team →The whole league's news, read by the model.
Beat reporters file all day. Posterior pulls the feeds hourly and runs each item through the model, which turns prose into dated, sourced signals — and flags the ones that actually move a price.
On the difference
between a model and
a hot streak.
— THE EDITOR
There is a small but durable industry built around men who picked the same side as the closing line for four games in a row and then sold a newsletter. Posterior is not that. The whole point of publishing the architecture — pillars, calibration curve, full slate ledger — is to remove every reason a reader has to trust us at our word. Trust the curve, not the copy.
What the model does well: it prices markets where the public posts a slow, sentimental probability and the book lays a thin vigorish on top. That gap is small — a handful of percentage points at most — and exists only in markets where the priors are dense and the signals are decomposable. Player-prop hits, total bases, walks. It does not do moneylines well. It does not pretend to.
What the model does badly: anything where the underlying causal chain is short and chaotic. One-game pitcher results, no-hitters, weather-cratered totals. We don't post entries there. The discipline of the EV threshold means most nights we ship a small handful of entries, not forty. That's the product.
If you came here expecting certainty and hype, this is deliberately not that. If you'd like to know what a probability actually is, and what it costs to be wrong about one, you're in the right almanac.
What you'd ask a friend.
No marketing. Plain answers.
What books do these picks work on?+
Any licensed operator available in your jurisdiction can be used as a comparison point. Posterior publishes calibrated probabilities, consensus context, and conservative risk references; it does not place, route, or encourage wagers. If you choose to act, you are responsible for checking the current line and local rules.
Do I have to bet every entry?+
No — and Posterior never tells you to. STRONG entries are the model's highest-confidence signals; LEANs are smaller-gap signals. Below-threshold observations stay internal and are not presented as picks. Some slates have several entries; some have none. Ignoring the board entirely is always a valid choice.
How is this different from a tipster?+
Tipsters give you a name and a vibe. Posterior gives you a calibrated probability, a conservative risk reference, and a public track record. Every entry is logged the moment it's made and graded against the actual outcome the next morning. The calibration curve above is live — you can see for yourself whether the 60% model entries actually hit 60% of the time.
What's a STRONG actually mean?+
STRONG = the model's calibrated probability is at least 7 percentage points higher than the sharp consensus (no-vig fair price across Pinnacle / Bet365 / DraftKings / FanDuel / BetMGM). LEAN = 3pp edge. Anything below that stays out of the customer pick ledger. Tiering against the actual market means a 22% prediction at a +400 line can be a real edge while a 50% prediction at a -150 line is not — both are graded honestly.
When do you publish tomorrow's picks?+
We publish entries across four scheduled windows daily — 11:00 AM, 12:15 PM, 3:00 PM, and 6:30 PM ET — as MLB posts each game's lineups. Most evening-slate entries are filed by 6:30 PM ET, about 30 minutes before first pitch on the 7:00 PM ET games. Subscribers get a notification when each publish window completes.
- 1Statistics. the distribution of a belief once the evidence is in — what Bayes' theorem turns a prior into. — this is the one we mean.
- 2Anatomy, informal. the hindquarters; the seat; the rump. — yes. we are aware.
“We named it for the math. The second reading is a happy accident we've chosen to wear with dignity — cheek included.”
Two ways in.
Aggregate accuracy is public. Tonight's top STRONG signals are free at /today and source-linked MLB news briefs live at /news and the complete historical pick log is downloadable at /data. The full live slate at publish time — every entry with conservative risk reference and real-time sharp consensus — is the subscriber product. No affiliate program, no Discord, no Telegram, no sportsbook referral, no sliding pricing model.
Pay as you go. Cancel anytime.
- →Full slate · every actionable signal, every night
- →Conservative risk reference against a hypothetical bankroll
- →Two actionable confidence tiers (STRONG / LEAN)
- →Morning recap by 5:30 AM ET · postmortem of every entry
- →Closing-line snapshot stored on every entry the morning after
- →Pre-game line snapshots + movement deltas
- →Cancel anytime · Stripe-secured · no app to install
Founding annual billing at $199/yr while the subscription remains active. Save $149 vs monthly.
- →Full slate · every actionable signal, every night
- →Conservative risk reference against a hypothetical bankroll
- →Two actionable confidence tiers (STRONG / LEAN)
- →Morning recap by 5:30 AM ET · postmortem of every entry
- →Closing-line snapshot stored on every entry the morning after
- →Pre-game line snapshots + movement deltas
- →Cancel anytime · Stripe-secured · no app to install
Stripe-secured checkout · cancel anytime · existing subscribers stay at their current price