How It Works
signal vs noise · the verdict engine · interpretable, not consensus · as of June 13, 2026
Signal vs Noise
Not all stats are equally trustworthy. A player's barrel rate stabilizes in about 60 batted-ball events — roughly two weeks of regular playing time. His BABIP takes over 800 balls in play before it tells you anything meaningful about his true talent. Everything in between lands on a spectrum.
Daily Delta flags every metric as either signal (the sample has cleared the point at which that stat reliably reflects true talent) or noise (it hasn't, so you're looking at variance, not skill). The thresholds come from public reliability research (Carleton et al.: split-half reliability ≈ 0.70) and are wired into the catalog — the same numbers the pipeline uses to flag cards are the ones shown in the table below.
The product differentiator is not the chart — Baseball Savant has charts. It is the verdict-on-the-chart and the movement board: what just became believable today, and what does it mean for your roster.
Stabilization Thresholds
The table below shows the real thresholds from the live catalog — the exact sample sizes at which each metric crosses from noise to signal. Read them top-to-bottom: fastest-stabilizing first, so the story of "what you're allowed to believe yet" flows naturally.
Hitters
| Metric | Stabilizes At | Unit | 2026 Lg Avg |
|---|---|---|---|
| Avg Exit Velo | 50 | BBE | 89.0 |
| Hard-Hit % | 50 | BBE | 40.0% |
| Barrel % | 60 | BBE | 8.0% |
| Strikeout % | 60 | PA | 22.0% |
| Walk % | 120 | PA | 8.5% |
| xwOBA | 160 | PA | .315 |
| wOBA | 200 | PA | .315 |
| BABIP | 800 | BIP | .295 |
Pitchers
| Metric | Stabilizes At | Unit | 2026 Lg Avg |
|---|---|---|---|
| Hard-Hit % Allowed | 50 | BBE | 40.0% |
| Barrel % Allowed | 60 | BBE | 8.0% |
| Strikeout % | 70 | TBF | 22.0% |
| Walk % | 170 | TBF | 8.0% |
| xwOBA Allowed | 200 | TBF | .315 |
| xERA | 200 | TBF | 4.10 |
| ERA | 200 | IP | 4.10 |
| BABIP Against | 800 | BIP | .295 |
PA = plate appearances · AB = at-bats · BBE = batted-ball events · BIP = balls in play · TBF = total batters faced · IP = innings pitched · PITCHES = pitches thrown. Thresholds are conservative defaults from published research. League averages are the 2026 MLB baseline.
The Verdict Engine
The verdict is deterministic. Rules in the analytics layer compute Buy / Buy-Low / Hold / Sell-High / Sell from the observed metrics, their signal state, and a player's distance from the league average. The language model in the narrative layer writes prose explaining that verdict — it never overrides it.
If the narrative model asserts a different action than the rule engine computed, the card is flagged Needs Review and held back from auto-publish. The rule engine is always the decision; the narrative is the explanation.
- BUY
- Stable signal metrics above league average — the underlying skill is real and the surface stats should follow.
- BUY-LOW
- Strong underlying skill metrics but surface stats (AVG, ERA) have underperformed — the discount is temporary.
- HOLD
- Mixed signal or insufficient sample — the picture isn't clear enough to act. May carry a tilt (leans buy, leans sell, wait & see, or steady).
- SELL-HIGH
- Surface stats are ahead of underlying skill — the current value is unlikely to hold. Sell before regression.
- SELL
- Stable signal metrics below league average — the weakness is real, not a slump.
Projection: Interpretable, Not Consensus
Daily Delta does not try to out-accuracy THE BAT X or ZiPS. It ships the interpretable projection: a forward estimate computed by shrinking the observed value toward the player's true-talent prior, in proportion to how far the sample is from stabilization.
A player with 30 PA has a barrel rate that's half noise — so the projection shrinks it 50% toward his career baseline. At 60 BBE (the stabilization threshold) the shrinkage drops to zero and the observed value speaks for itself. You can see exactly why the projection is what it is; there is no black box.
This model is not built for season-long precision. It is built to answer: given what we can trust today, what's the honest rest-of-season estimate? That is a different question than the consensus projections answer, and it is the right question for a daily tool.