Scoring Guide

Score (0-100)

A weighted algorithm that grades the batter across 15 factors — recent performance, contact quality, matchup against the opposing pitcher, ballpark, and more. Think of it as a “matchup quality” rating. Higher is better, but the number alone doesn’t account for how many plate appearances the batter will get.

Hit Probability (50-90%)

The score converted to “what are the odds this batter gets at least one hit today?” This is the primary metric for BTS picks. It accounts for lineup position — a leadoff hitter with the same score as a 9-hole hitter will have a higher hit probability because they get more plate appearances.

ML Probability (55-75%)

An independent machine learning model (LightGBM) trained on 180,000+ historical batter-games across 5 seasons. It uses 23 raw features and doesn’t know about the score — it’s a completely separate second opinion. When Hit Prob and ML Prob agree, you can be more confident in the pick.

Contact rate is the single strongest predictor of getting a hit. Batters who put the ball in play give themselves a chance on every swing. K rate is inverted — lower strikeout rates score higher. Both are further adjusted by chase rate and whiff rate percentiles (see Discipline Adjustments below).

Who’s hot right now? L7 and L15 batting averages are rolling windows that capture streaks and slumps — they use raw points with no smoothing, so what you see is what the batter is actually doing. Career BA provides a stable baseline. Season BA is blended with prior-season data early on (see PA Confidence below).

Statcast data based on exit velocity and launch angle. These measure how hard and how well a batter is hitting the ball, regardless of whether fielders happened to catch it. They strip out luck. These require the most PA to stabilize (see PA Confidence).

Hard-hit balls (95+ mph exit velo) and line drives are the types of contact most likely to become hits. A batter who barrels everything but hits it right at fielders will still show up well here.

The opposing pitcher’s recent form matters most here. Sprint speed uses raw points (it’s a physical trait, not sample-size dependent) with a ground ball rate multiplier — fast runners who hit grounders benefit more. BvP uses historical head-to-head data (requires 5+ at-bats). Hitter-friendly parks like Coors Field provide a small boost.

BvP Redistribution

When the batter has fewer than 5 career at-bats against today’s pitcher, the BvP data is unreliable. Its 4 points are zeroed out and 1 point each is added to K rate, contact rate, pitcher form, line drive rate, and L15 BA — factors with more stable data. This adds 5 points to the pool (net +1), so the total possible score is 101 when BvP is insufficient.

Training data

180,000+ historical batter-game records from 2019-2023, tested against 2024-2025 holdout data. Each record has 23 features built from rolling stats (not end-of-season aggregates) to avoid lookahead bias. The model learned what patterns predict hits from real in-season data.

What it knows

23 features including batting averages (L7, L15, season, career), prior-season pitcher stats (ERA, WHIP, FIP, K/9, IP), Statcast metrics (xBA, xwOBA), hard hit rate, sprint speed, contact rate, line drive rate, rest days, home/away, lineup position, ballpark factor, and platoon matchup.

Score

Sorted by the raw 100-point score. Best for seeing overall matchup quality without lineup position adjustment.

Hit Prob (default)

Sorted by hit probability. The recommended primary sort because it accounts for lineup position and expected plate appearances.

ML Prob

Sorted by the independent LightGBM model’s prediction. Useful when you want a pure data-driven ranking with no formula tuning.

Combined

A 3-way rank average: each batter’s rank position in Score, Hit Prob, and ML Prob is averaged. The lowest average rank (best consensus across all methods) sorts to the top. This is the best way to find batters that all three methods agree on.

Confidence badges

Each AI pick shows a confidence level: High (strong agreement across methods), Medium (reasonable signal), or Low (methods disagree or marginal stats).

Streak tracking

The AI picks card tracks two streak types: 1x (singles) where only the #1 pick needs a hit each day, and 2x (doubles) where both picks must hit or the streak resets. These mirror the BTS game’s single pick and double down modes.

Five methods tracked

Score, Hit Prob, ML Prob, Combined, and Consensus (a method that only picks batters appearing in the top N across multiple methods). Each gets a hero card with cumulative accuracy and head-to-head day wins.

Filters and views

Filter by pool size (#1, Top 3, Top 5, Top 15) and date range. Toggle between daily accuracy line chart and $1 bankroll simulation. Expand any day in the table to see exactly which batters were picked and whether they got hits.

L7 BA, L15 BA, career BA, sprint speed, ballpark factor, pitcher recent, BvP, bullpen quality.

These either don’t depend on current-season sample size (career BA, sprint speed, ballpark) or are rolling windows already bounded by their time period (L7, L15). What you see is what the batter is actually doing.

~2-3 weeks of games

Season BA, K rate, contact rate, hard hit rate, line drive rate. These need some data to be reliable but stabilize relatively quickly. Blended with prior-season values until 60 PA.

~8-9 weeks of games

xBA and xwOBA only. Statcast expected metrics need the most data because they rely on batted ball distributions that take longer to stabilize.

Late scratches & pinch hitters

If a batter is scratched after lineups are confirmed, or gets pinch-hit for early, they may get fewer PA than expected. The system can’t predict mid-game decisions.

Weather & delays

Rain delays, postponements, and shortened games can reduce plate appearances or cancel games entirely. The system doesn’t factor in weather forecasts.

Unreported injuries

A batter playing through a minor injury that hasn’t been reported won’t trigger the injury modifier. The system only knows what the API reports.

Pitcher changes

The score is calculated against the announced starting pitcher. If the starter gets pulled early or an opener is used, later at-bats may be against a different pitcher than the one scored.

Early-season noise

In the first 2-3 weeks, current-season stats are based on small samples. The PA confidence system smooths this for some metrics, but L7/L15 averages can be volatile. A 3-for-5 day can swing a batter from .200 to .333 in L7.

The 59% ceiling

The baseline hit rate for any MLB starter is ~59%. Even the best picks only push that to ~67-70%. Getting a hit in any single game is inherently uncertain — the edge comes from making consistently good picks over many games.