How Mixl works

Twentyseasonsin.Onehonestnumberout.

Mixl is a data pipeline with a scoreboard: ingest everything, train nightly under rules that make cheating impossible, price the slate, and grade every call in public.

[ 2002 ]
seasons on file
[ 10k+ ]
sims per matchup
[ 5+ ]
markets priced daily
[ 100% ]
calls logged first
scroll

The pipeline

From raw pitches to a priced slate

01 · Corpus

The corpus

Every game, box score, play, pitch and plate appearance since 2002 — plus posted lineups, probable starters, injuries, park weather and sportsbook lines — lands in Mixl's own store. Live games merge in within seconds of the final out.

It isn't a feed we rent by the call; it's a warehouse we own and control, so a model can look back across two decades in one query and a live game becomes history the moment it ends.

  • Every pitch, play and plate appearance — not just box scores.
  • Lineups, probables, injuries, park weather and book lines, joined.
  • Live games become history within seconds of the final out.
[ 2002 ]
seasons on file, to today
[ 2002 ]
first season on file
[ 6+ ]
data layers, one store

the corpus

gamesbox scores
playspitches
PAscontext

every pitch, play & PA — one store, one query

02 · Point-in-time

Point-in-time discipline

Every row is stamped with when it was actually knowable. Models can only train on what existed at decision time — no peeking at the future, the single most common way sports models silently cheat.

Rolling stats are shifted, forecasts are never backfilled, and econ-style vintages are honored. A feature that couldn't have been seen before first pitch is structurally unavailable to the model that predicts it.

  • Every row stamped with the instant it became knowable.
  • Rolling stats shifted; forecasts never backfilled.
  • Backtests read as-of — the model can't see its own future.
[ 0 ]
rows a model sees from its own future
[ 0 ]
future rows a model can see
[ as-of ]
how every backtest reads

point-in-time

knowable · trainable the future · off-limits

shift(1) · as-of reads · no backfilled forecasts

a model can't see data it couldn't have known

03 · Training

Nightly training, walk-forward only

Models retrain on fresh data every night. A challenger replaces a champion only by beating it out-of-sample, walk-forward across seasons — never by fitting the past harder.

The champion is the number you see; a challenger has to win on data it was never trained on before it's allowed to take the mound. Overfitting the past earns nothing here.

  • Retrains on fresh data every single night.
  • A challenger must win out-of-sample to take over.
  • Walk-forward across seasons — never refitting the past.
[ nightly ]
champion vs. challenger, out-of-sample
[ nightly ]
retrain cadence
[ OOS ]
the only bar that promotes

nightly training

champion72%
challengerpromoted ↑81%

out-of-sample window →

tested out-of-sample, walk-forward across seasons

04 · The board

The board

Each day the engine prices the slate — game winners, totals, run lines, strikeout and batter props — and quotes our probability next to the sportsbook consensus and the live Kalshi price.

You don't get a number in a vacuum. Our model sits side-by-side with the market so the disagreement — the edge — is the thing you actually read.

  • Winners, totals, run lines, strikeout and batter props.
  • Our probability beside the sportsbook consensus.
  • And beside the live Kalshi price — the edge is the gap.
[ 5+ ]
market types priced every day
[ 5+ ]
market types priced daily
[ 3 ]
numbers, side by side

the board

Yankees to win+8% edge
Mixl model62%
Kalshi market54%
Yes54¢No46¢

the gap between the two is the edge

05 · Receipts

Receipts, in public

Every model call is logged before first pitch and scored against the settlement. The track record page shows the hits, the misses and the calibration — not a highlight reel.

Calls are stamped and locked before the game, benchmarked against de-vigged closing lines, and graded on calibration nightly. Settled misses stay up — there's no quiet deleting of the ones that didn't land.

  • Every call locked and logged before first pitch.
  • Scored against the settlement, graded on calibration.
  • Settled misses stay up — nothing quietly deleted.
[ 100% ]
of calls logged before first pitch
[ 100% ]
calls logged before first pitch
[ close ]
the benchmark we're graded on

receipts, in public

Yankees MLloggedhit
Over 8.5loggedhit
Ohtani 2+ Kloggedmiss
Dodgers -1.5loggedhit

graded against the close — misses stay up

06 · You

You, in the loop

Search the corpus in plain English, run your own Monte Carlo simulations, and see where our numbers disagree with the market. The data argues; you decide.

Ask a question and get the game log behind the answer. Simulate a matchup ten thousand times — fifty thousand on Pro — and watch the distribution form. Every conclusion traces back to a row you can check.

  • Ask in plain English, get the game log behind the answer.
  • Simulate any matchup 10,000 times — 50,000 on Pro.
  • See exactly where our number disagrees with the market.
[ 10k+ ]
games you can simulate per matchup
[ 10k+ ]
sims free · 50k on Pro
[ 1 ]
unlocked model pick daily

you, in the loop

ohtani vs padres, last 10
ask
simulate
compare
10ksims free
50ksims Pro
1pick daily

the data argues; you decide

The data argues

Watch the number sharpen itself

This is one real matchup, simulated ten thousand times. The win-probability line settles and its 95% confidence band collapses like 1/√N — the reason a distribution beats a hot take. Not told; drawn.

New York Yankees @ Los Angeles Dodgers10,000 games · seeded replay
Run your own simulation

The discipline

Rules that keep the numbers honest

Walk-forward or it doesn't ship

A model only sees the past when it predicts. Backtests that leak tomorrow's data into today's features are how 90% accuracy claims are minted — ours are structurally impossible to mint that way.

Why it matters: leakage is the difference between a number that looked great in a slide and one that survives contact with a live game.

Benchmarked against the market

Every model is scored against de-vigged sportsbook closing lines — the hardest baseline there is. Beating a coin flip is easy; the honest question is beating the close.

Why it matters: the closing line already absorbs the sharp money, so beating it is the only bar that means anything.

Calibration over bravado

A 60% call should land 60% of the time. We publish calibration, and our models are graded on it nightly — confident and wrong is the one thing the pipeline won't tolerate.

Why it matters: a well-calibrated 58% beats a loud, miscalibrated 80% every time you have to act on it.

Demo money only

The autonomous trading side runs on demo funds while the track record accrues. No real-money bravado, no pressure to overclaim.

Why it matters: nothing about the numbers you see is bent to defend a position we're holding.

The proof is on the track record page — every call, scored after settlement.

Seethereceiptsforyourself.

Free accounts get the search, three simulations a day, and one unlocked model pick daily.

Free account · 3 simulations a day · no card required