How EyeBlackIQ Works

Primary focus. Team ML signals using ELO ratings, ISR (Iterative Strength Ratings), and SP ERA adjustments. Updated daily with the latest D1 rankings.

• ELO head-to-head probability (HFA = +25 pts)
• ISR normalized win probability blend (40%)
• Starting pitcher ERA adjustment (±run line shift)
• Run line alternative surfaced when ML > +250

Team ML and player prop signals for NHL (SOG on slip · PTS/AST tracked silently), MLB (pitcher Ks, F5 results), and Soccer (1X2, SOT, Goals). Same quantitative rigor as NCAA baseball. NHL Team ML now uses ESPN Core API for live odds.

• NHL: MoneyPuck xG data + NHL API stats + ESPN Core API (live odds)
• MLB: Statcast / pybaseball (Spring Training calibration through 3/26)
• Soccer: Club Elo API + Understat xG + football-data.co.uk

Resource-Value model. Par Score (venue-adjusted expected score), Venue Z-Factor (ground-specific scoring deviation), ZIP wickets model (innings survival probability), T20 middle-order compression, ELO + Par blend 55/45. Separate ELO pools for franchise (IPL) vs national (ICC) teams.

• Cricsheet T20/IPL parser (cricsheet.org)
• API Sports cricket fixtures

Efficiency-Flow model. Possession-based Strength of Schedule (SOS), Adjusted Expected Goals (adj_xG = raw_xG × league_avg_def / opp_def), Poisson goal distribution, ELO + Poisson blend 55/45, Usage redistribution for injury absences, Platt scaling calibration infrastructure.

• API Sports EHF/Bundesliga
• ESPN sports API

Two independent rating systems are computed for every team:

• ELO — Elo rating updated after every game. Head-to-head win probability uses P = 1 / (1 + 10^(-Δ/400)). Home field advantage = +25 ELO pts.
• ISR — Iterative Strength Rating converges to a stable power-ranking via repeated margin-of-victory regression. Updated weekly.

Blend: 60% ELO + 40% ISR. The gap between ELO and ISR directly drives the confidence level.

Model probability is compared to the no-vig (fair) market probability.

Devig method: imp = abs(o)/(abs(o)+100) then normalize two-way to 100%.

Edge = Model% − Market%

Only positive edge signals advance to tier assignment. The market isn't always wrong — the model only acts when there's a measurable discrepancy.

Starting pitcher quality shifts team win probability based on ERA vs. league average (5.70).

Formula: r = 5.0 × SP_ERA / 5.70 (capped 1–9 runs)

ERA adjustment: (r_away − r_home) / 20 added to home win probability.

A dominant home SP (ERA 2.50) vs. shaky away SP (ERA 6.20) can shift p_home by ~3-4pp.

Every signal must pass all five gates in sequence. Any failure blocks the pick.

1. Pythagorean gate — team performance vs. expected (GREEN or cleared YELLOW)
2. Edge ≥ 2% above no-vig fair probability
3. ELO and ISR models agree within 15pp
4. Line has not moved >0.5 units sharp-direction against model
5. ETL data timestamp < 4 hours old

Tier	Edge	Units	NCAA / MLB	NHL	Soccer
T1	≥12%	2.0u	FILTHY	SNIPE	UPPER 90
T2	≥5%	1.5u	WHEELHOUSE	SLOT MACHINE	CHEEKY
T3	≥2%	1.0u	SCOUT	SCOUT	SCOUT
T4	<2%	0u	BALK	ICING	PARK THE BUS

All units are sized using Quarter Kelly with a hard bankroll cap.

• Kelly fraction: 0.25 (1/4 of full Kelly)
• Hard cap: 3% of bankroll per play
• Daily maximum: 15 plays
• 1 unit = 1% of bankroll (e.g. 2.0u = 2% of bankroll)

Units are set by tier, not raw Kelly — the tier floor ensures minimum sizing when edge is confirmed. Kelly is used as a ceiling check only.

Confidence is a grounded signal about how well-supported this particular signal is — not whether it's a good bet. It's built from objective inputs, not gut feel.

Critical distinction: Tier (edge%) determines your stake. Confidence tells you how reliable the underlying model inputs are. A FILTHY pick at MED confidence still has 12%+ edge — that's a significantly larger expected value than a SCOUT at HIGH confidence with 2.4% edge.

Example: BYU +1400 FILTHY 2.0u LOW ●○○ (23.9% edge) vs. Auburn +105 WHEELHOUSE 1.5u HIGH ●●● (6.2% edge). BYU's edge is nearly 4× larger. LOW confidence signals that SP data is limited — not that the edge is fabricated. Auburn's HIGH confidence reflects strong model agreement and confirmed SP starts. Both are legitimate plays; size them accordingly.

POD Rule: Pick of the Day requires HIGH confidence + WHEELHOUSE or better. PODs are the face of the system — they should reflect our highest-conviction signals where both the edge AND the model certainty are strong.

When the model finds significant edge on a big underdog (ML > +250), the system flags a run line alternative. Instead of betting BYU at +1400 ML (where you need them to actually win), you can bet BYU +1.5 runs at a much lower price (typically +300 to +500) and still cash if they keep it close.

This is especially valuable for teams like BYU where our model says 30% win probability but the market says 6% — the "surprise competitive game" outcome is more likely than outright victory. The run line captures value from the team being competitive without requiring an upset.

Future feature: When ML edge >15% on a big dog, the system will automatically surface the run line as a separate bet with its own edge calculation against the offered run line odds.

• 3-month rolling CLV < 0 → full suspension
• 25 consecutive losses → 72-hour review period
• 7 consecutive losing days → stakes reduced to 25% for 5 days

No model goes live until all four thresholds are cleared:

• Backtest ROI > 3%
• CLV (Closing Line Value) ≥ 55%
• Calibration within ±2% across all probability bins
• Minimum 300–500 sample plays in paper trading

CLV measures whether we're beating the closing line — the final market price after sharp money has moved the line. Pinnacle closing line is our gold standard benchmark.

Beating the closing line at 55%+ rate is the strongest long-term profitability signal — more reliable than win/loss record over small samples.

Phase: Active paper trading — calibrating models and tracking CLV.

NCAA Baseball: Live signals since March 21, 2026. Tracking real-time.

NHL: Active signals via ESPN Core API + FanDuel player props. ELO+Pythagorean blend.

Cricket: Data phase complete (4,390 T20 matches loaded). Signals in forward-looking mode.

Handball: Data phase complete (973 matches). Forward picks posted for EHF CL QF (Apr 2, Apr 9).

Soccer: Active signals via Club Elo + Understat xG.

Go-live target: 30 days paper trading + all 4 thresholds cleared per sport.

EyeBlackIQ posts all plays with measurable edge (EV ≥ 2%). There is no daily cap on number of plays. The market doesn't limit its inefficiencies to 5 per day — neither does the model.

The tier system handles sizing: T1 plays get 2.0u, T2 get 1.5u, T3 get 1.0u. Volume is managed through edge threshold, not an artificial daily quota.

Kill switches handle drawdown protection automatically (25-loss streak, 7 losing days, 3-month CLV negative) — not volume caps.

• It does not predict game scores directly — it computes win probability vs. market implied probability
• It does not claim to be right every day — edge means winning over large samples, not game-by-game
• It does not chase losses — kill switches trigger reduced sizing automatically
• It does not post picks without data — both TBD starting pitchers blocks NCAA signal generation
• It does not use subjective gut reads — every signal is fully model-derived
• It does not guarantee future results — past edge calibration doesn't guarantee forward performance

• Signal layer: Every pick posted with tier, edge %, EV, confidence, and line at time of signal
• Results layer: Every outcome recorded publicly — wins and losses, unit P&L per pick
• Model layer: Full methodology published here — no black boxes, no cherry-picked records

The full market view section on the Results page shows every game analyzed — not just picks with positive edge. This lets you see what the model passed on and why.

Each sport gets at most one POD per day. POD selection algorithm: highest ev_pct × win_probability_gap with minimum 7% EV.

POD picks require HIGH confidence (●●●) plus WHEELHOUSE tier or better. This ensures the system's flagship pick represents both strong edge AND strong model certainty.

PODs are tracked separately in the Results page. This provides a clean comparison between the high-conviction signal and the full model record.

• ELO ratings — updated weekly (Mar 21 as of today)
• ISR ratings — updated weekly (Mar 21 as of today)
• ISR ratings from Boyd's World (boydsbets.com)
• ELO ratings from Warren Nolan (warrennolan.com)
• Conference ISR — context and matchup annotation
• SP ERA data — scraped daily from ESPN / Warren Nolan
• Market odds — TheRundown API (Pinnacle included)

• NHL: MoneyPuck CSV (xG, 2007+) · NHL official API
• NHL props: FanDuel public sportsbook API · ESPN Core API (free, no key)
• MLB: pybaseball / Baseball Savant · Statcast
• Soccer: Club Elo API · Understat xG · football-data.co.uk
• Live odds: TheRundown API free (20K calls/day, Pinnacle + 15 books)
• Props: OddsAPI (MLB Ks, NHL SOG)

• Cricket: Cricsheet T20/IPL parser (cricsheet.org) · API Sports cricket API · ICC/IPL fixtures
• Handball: API Sports handball API (EHF CL, Bundesliga) · ESPN sports API · EHF official calendar

Base URLs: v1.cricket.api-sports.io / v1.handball.api-sports.io

EyeBlackIQ integrates API Sports for expanded sport coverage. API Sports provides official fixture data, results, and odds across cricket, handball, soccer, MLB, and NHL.

• Cricket: IPL, ICC T20 World Cup, ICC ODI World Cup fixtures and scores
• Handball: EHF Champions League, German Bundesliga Handball schedules and results
• Soccer: Supplemental to football-data.co.uk for live fixture tracking
• MLB / NHL: Supplemental odds data when official APIs lack market lines

When primary paid APIs (OddsAPI, TheRundown) are unavailable, EyeBlackIQ uses a cascade of free public endpoints. ESPN Core API provides ML, spread, and O/U for NHL, MLB, soccer leagues, and NCAA baseball. FanDuel's public sportsbook endpoint provides player prop markets (SOG, Goals, Points, Anytime Goal Scorer) for NHL with no API key. This ensures model continuity regardless of credit cycles.

• ESPN Core API: https://sports.core.api.espn.com/v2/sports/hockey/leagues/nhl/events/{id}/competitions/{id}/odds
• FanDuel Public: https://sbapi.nj.sportsbook.fanduel.com/api/content-managed-page?page=CUSTOM&customPageId=nhl&_ak=FhMFpcPWXMeyZxOx&includeMarkets=true

Picks sourced from pick'em platforms — PrizePicks, Underdog Fantasy, ESPN Pick'em, DraftKings Pick'em — are labeled Pick'em throughout the site. Pick'em lines represent player props or game outcomes priced at platform-specific formats (no vig, over/under style). These are tracked separately from traditional sportsbook lines to allow independent analysis.

• When our model also shows roughly equal probability, the pick'em badge is a calibration check — both model and market agree, which increases confidence in the no-vig probability estimate
• When our model significantly disagrees with a pick'em line, the edge reading may be inflated relative to higher-priced plays — the badge surfaces this for manual review
• Pick'ems are tracked separately in the Results page to allow independent calibration analysis

You can filter for Pick'em picks on both the Today and Results pages using the Pick'ems filter button.

The Full Market View section (visible at the bottom of the Today page) shows every game the model processed that day — including games where edge fell below the 2% threshold or other gates were not met. These are informational outputs only with no unit sizing.

Why this matters:

• Transparency — you can see what the model passed on and why, not just what it liked
• False positives — occasionally a model outputs a low-confidence edge that doesn't survive full gate filtering; the Full Market View shows these without dressing them as recommendations
• Long-term tracking — all Full Market View picks are graded in the Results page under the same methodology as recommended picks, allowing unbiased record-keeping across the full model output