Sparrow Data
Methodology
The model estimates one thing: the probability that a given starting hitter hits a home run in a given game. It outputs a calibrated probability, never a yes/no call. About 12% of starter-games end in a home run; a pick at 20% is a strong call, not a certainty.
What goes into it
Three groups, deliberately described as groups rather than as a feature list: talent (career and season contact quality — barrel rate, exit velocity, hard-hit rate), form (recent 30-game and 7-day deviations from that bat's own norm), and matchup (the starting pitcher's home-run and contact profile, the park's handedness-specific home-run factor, and weather). Predictions are anchored to each bat's own trailing baseline, so the model is always answering 'how is tonight different for this hitter' rather than 'who is good'.
The rules we hold ourselves to
- Boards are published before first pitch and graded the next morning against what happened.
- Grading reads the boards as posted. A pick cannot be quietly dropped from its own scorecard.
- Board rules are frozen quarterly. Selection rules never change mid-quarter in response to a bad run.
- Nothing is published without two-season retrospective receipts on untouched holdout data first.
- A candidate model is never promoted on its own numbers — promotion runs against criteria written down before the candidate was trained.
How to read our numbers
The boards and the nightly card are written in plain English. Every one of those sentences has an exact definition, and this is it — nothing on this site is stated one way to readers and another way in the data.
| What we say | What it means exactly |
|---|---|
| “homered 36% more often than these bats normally do” | Realized home runs divided by the sum of those same bats' own baseline probabilities, for that cohort: a ratio of 1.36. The baseline is a shrunk trailing-365-day per-game rate for that hitter, not a league average. |
| “A typical night for these bats is 2” | The sum of those bats' baseline probabilities for that night, rounded. It is what you would expect from this exact set of hitters with no model at all. |
| “enough that it's probably real, not luck” | The 95% bootstrap confidence interval on the ratio excludes 1.0. If it spans 1.0 we say so instead. The intervals themselves are printed on the record. |
| “not enough nights yet to call it” | Sample gating: below 10 graded dates in the window we publish no ratio at all rather than a noisy one. |
| “they hit the ball like guys who should have homered” | xHR — expected home runs from each batted ball's exit velocity and launch angle, on a grid built from every batted ball since 2021. It measures contact quality independently of whether the ball left the yard, so good process with bad results is visible as variance rather than hidden. |
| Tier labels (T1 Sluggers … T6 Glove-First) | Frozen probability bands. They are a display label only and never reorder a board. |
The model itself is scored on Brier score against a naive per-hitter baseline, with purged walk-forward validation and a 15-day embargo; calibration is checked against the reliability curve on the record. Those are the numbers that decide whether a model version ships.
What we tried and threw away
Nine pre-registered tests, eight of them dead. Each had its decision rule written down before the result was known, so a null could not be argued away afterwards. Publishing the failures is the honest half of publishing the record.
| Idea | Verdict |
|---|---|
| Weather (air density, wind to centre) | The one that lived. Passed its gate and is in the candidate model now, awaiting a scheduled review. |
| Batter age | Dead. Added nothing measurable; the run-to-run noise of the training itself exceeded the effect. |
| Bullpen exposure | Dead. No replication on any cut, and the only significant result was the placebo — a sign of confounding. |
| Batter profile x pitcher profile | Dead. The construction was contaminated by batter quality; the signal did not survive removing it. |
| Spray angle x park shape | Dead. The variant that looked best predicted the opposing team's home runs — it was reading the park, not the hitter. |
| Pulled barrels | Dead. It appeared to work, then a control showed it was plain barrel rate wearing a direction label. |
| Expected plate appearances | Dead. Team plate appearances turned out to be barely forecastable, and batting order already carries the part that is. |
| Day-game-after-night fatigue | Dead before it was even tested properly — the raw data pointed the wrong way. |
| Starting-pitcher recent workload | Dead. One season looked convincing; the next was exactly zero. |
Known limitations, stated plainly
- Predictions in the top band (25%+) have run hot — we have called those slightly more often than they landed. Sample is small; it is on a monthly watch list rather than patched.
- Left-handed bats against left-handed pitching are over-predicted by roughly 10%. The model has no way to express an asymmetric platoon effect, so this is structural rather than a tuning miss.
- A month of live data produces wide confidence intervals. Where a ratio appears on this site it appears with its interval; if the interval crosses 1.00x, the honest reading is 'not yet demonstrated'.
- FAVORITES claims no edge at all. It is published for transparency and should land near 1.00x.
If we get something wrong
Boards are never edited after the fact. A pick that was published stays published, exactly as it was posted, whether it hit or missed — that is what makes the record on this site worth anything. If a scoring error is ever found that changed a number a reader saw here, the correction is stated on the affected night's own page, alongside the original, rather than quietly folded into a total.
Auditing this yourself
The data directory carries the same picks and grades as CSV. Every figure on this site is computed from those rows by the same code that produces the nightly scorecard — there is no second pipeline that could quietly disagree.