Per-opportunity, not per-game
Per-game totals confuse pace with quality. Per-possession (basketball), per-drive (football), per-shot (hockey/soccer), and per-plate-appearance (baseball) metrics strip pace out and expose true strength.
This is the single most important shift a bettor can make in how they read stats. A team scoring 28 points per game can be more efficient than one scoring 32, simply because they have fewer possessions. Every modern public analytics site (Cleaning the Glass, FiveThirtyEight, FanGraphs, Natural Stat Trick) builds outward from per-opportunity rate stats.
Expected-value metrics - the closest thing to fundamentals
Expected goals (xG), expected points added (EPA), expected weighted on-base (xwOBA), and similar 'expected' metrics estimate the average outcome of a play type given league-wide data. They strip finishing luck from process quality and are the closest thing sports have to fundamental valuations.
xG-based ratings now form the backbone of soccer and hockey betting models because actual goals are too rare and too noisy to use directly. EPA per play has done the same for the NFL - pure record and yards-per-game tell you almost nothing that EPA doesn't tell you better.
Key metrics by sport
A working bettor's vocabulary across the five major US/global sports. Each metric below is publicly available, well-documented, and a meaningful upgrade on its box-score equivalent.
- NBA - Offensive/Defensive rating (per 100 possessions), True Shooting %, Effective FG%, Usage rate, Net Rating, On/Off splits.
- NFL - EPA per play, Success rate, DVOA, Pass-block/rush win rate (ESPN), Pressure rate, ANY/A (adjusted net yards per attempt).
- MLB - wOBA, wRC+, FIP, xFIP, SIERA, xwOBA (Statcast), barrel rate, hard-hit %, WAR (fWAR/bWAR).
- NHL - Expected goals (xG), Corsi For % (shot-attempt share), Fenwick, High-danger chances per 60, PDO (luck indicator).
- Soccer - xG, xA (expected assists), npxG (non-penalty xG), shot-creating actions, progressive passes/carries per 90.
Where advanced stats break down
Every metric has a stable range and a noisy range. xG over 5 NHL games is essentially noise; over 20 games it stabilizes. NBA on/off splits are heavily affected by lineup combinations and require 1,000+ minutes to be reliable. NFL EPA over 4 games can be entirely driven by 3–4 explosive plays.
Treat advanced stats as Bayesian priors that update slowly, not as crystal balls. The smaller your sample, the more you should regress toward the league average - by a lot. Public models that don't regress aggressively in small samples overfit dramatically.
Process stats vs outcome stats - repeatability
Process stats (shot quality, EPA generated, hard-hit rate) are sticky season-to-season - they tell you what a team is likely to do next. Outcome stats (actual goals, yards, runs) are partly luck - they tell you only what already happened.
When evaluating a team's hot stretch, ask whether their underlying process stats moved or only their outputs. If only outputs moved, the hot stretch is variance and will regress. If process stats moved too, something real changed - scheme, personnel, health.
Stabilization rates - when each stat starts to mean something
Advanced stats stabilize at different sample sizes. Using a metric before it stabilizes is how 'small sample noise' becomes a confident bet. The figures below come from publicly available research (FanGraphs, Football Outsiders, Cleaning the Glass, Evolving-Hockey).
- MLB - strikeout rate stabilizes ~70 PA, walk rate ~170 PA, BABIP ~820 PA, ERA effectively never (use FIP/xFIP/SIERA instead).
- NBA - 3-point shooting rate ~750 attempts, true shooting ~400 shots, on/off net rating ~1,000 minutes shared.
- NFL - EPA per play stabilizes around 250–300 plays (~7 games of offense), DVOA around mid-season for the offense, later for defense.
- NHL - individual xG rates stabilize ~250 shots; team-level xG share is reliable around the 20-game mark.
- Soccer - npxG per 90 stabilizes ~12–15 matches at team level; individual finishing rates take a full season or more.
Composite ratings - what public power ratings actually do
Public composite ratings (FiveThirtyEight Elo and RAPTOR, ESPN FPI/BPI, Massey/Sagarin, KenPom, EvanMiya) combine multiple advanced inputs into single team-strength numbers and adjust for opponent quality. They are useful as priors, not as final answers - most are calibrated within ~2 points of market spreads on average.
Differences between composite ratings often reveal where models disagree about a team. A team rated 3 points apart by KenPom and Bart Torvik in college basketball is usually a team whose defense or pace is being measured inconsistently - worth investigating before betting either side.
- NFL - FPI, DVOA, EPA-based ratings, FiveThirtyEight Elo, PFR SRS.
- NBA - RAPTOR, EPM, LEBRON, BPI, Cleaning the Glass team ratings.
- College football - SP+ (Bill Connelly), FPI, Massey, Sagarin.
- College basketball - KenPom, Bart Torvik, Haslametrics, EvanMiya.
- Soccer - FiveThirtyEight SPI (archived), Opta power rankings, club Elo (clubelo.com).
How to read a stat line like an analyst
When a team's record diverges sharply from its underlying metrics, look for the gap. A 10-2 team with a -1 point differential is almost certainly a regression candidate; a 6-6 team outscoring opponents by 7 points per game is usually underrated. Pythagorean expectation (used in MLB and adapted across sports) formalizes this gap and is one of the most robust simple models in sports analytics.
Pythagorean win % = (points scored^x) / (points scored^x + points allowed^x), with x ≈ 2.37 for the NFL, ~14 for the NBA, ~1.83 for MLB. Teams meaningfully above or below Pythagorean expectation tend to regress within 20–30 games.
Frequently asked questions
What are the most useful advanced stats for sports betting? Per-possession or per-opportunity efficiency in every sport, plus expected-value metrics (EPA, xG, xwOBA) that strip finishing luck from process quality.
What does EPA mean in football? Expected Points Added - the change in a team's expected points before and after a single play, based on down, distance, field position, and time. Aggregating EPA per play gives a far more accurate efficiency picture than yards per play.
What is xG in soccer and hockey? Expected goals - the probability a shot becomes a goal given its location, type, and situation, summed across all shots. xG is more predictive of future goal-scoring than past goals scored.
Are advanced stats public or paid? Most are freely available. NBA.com/stats, Basketball Reference, Pro Football Reference, FanGraphs, Natural Stat Trick, FBref, and Cleaning the Glass cover almost everything a recreational analytical bettor needs.