Your idea, sharpened: not a sportsbook — a decision layer that sits on top of every book. It pulls every price, strips out the vig, computes its own probability from team history, player history and current-season form, and only surfaces a bet when your probability beats the best price available anywhere. Everything else — most games, most weeks — it should honestly say “no bet.”
Steal: their prop menus (as a scan list), line-movement charts, side-by-side odds boards, bet tracking & automatic grading.
Refuse: odds boosts and same-game parlay pushes. Those are the books’ highest-vig products — a platform in favour of the bettor prices most SGPs as negative-EV by construction, and says so.
Build verdict: feasible cheaply for the NFL. Free nflverse data for the model + The Odds API free tier (500 credits/mo) for prices prototypes the entire loop. The realistic win is small edges, best prices, and discipline — not a money printer. Evidence in §8.
You must win 52.38% just to break even. The −110/−110 market implies 104.76% — the 4.76 pts is the vig.
Unbiased over 7,276 games — but mean absolute miss is 10.27 pts (SD 13.20). Calibrated, yet very noisy.
8,400+ games. A good public model trails the market by ~2.6 winners per season — with no injury awareness at all.
Titans moneyline range across books in the Week 8 snapshot. Best price alone is worth more than most model tweaks.
Sources for strip: MarketMath devig guide · NFL Analytic market efficiency (nflverse, 1999–2025) · Independent NFL Elo FAQ · RG.org Week 8 cross-book table. Full sourcing throughout; research conducted Thu Oct 29, 2026.
The nearest slate when the research ran — Thu Oct 29 to Mon Nov 2, 2026. Byes: Saints, Giants, 49ers, Texans. Search it, filter by day, and note how far prices stray from each other.
| Game (away @ home) | Spread | Total | Moneyline A / H | Covers consensus |
|---|
Covers (lines as of 10-28): NE −6.5 · total 43.5 | FanDuel Research: NE −7 (−105/−115) · total 44.5 · ML −330/+265 | bet365 (Oct 26): NE −7.5 · ML −380/+300.
Three spreads and a 50-cent moneyline range for the same kickoff. Any app that shows one book’s price as “the odds” is hiding free edge.
Sources: Covers Week 8 · FanDuel Research · Geneva Times / FOX bet365 lines, Oct 26
Titans ML +420 (Fanatics) to +475 (BetRivers); Bengals −588 (DK/Fanatics) to −667 (BetRivers); spread −10 at DK/BetRivers vs −10.5 elsewhere; total 46.5–47.5 by book. RG.org’s own note: DK/BetRivers dealing −10 “rather than the −10.5 at the other three books, and that half-point is worth taking.” Bengals were the week’s biggest favourite; Cardinals @ Cowboys carried the highest total (54.5).
Ravens −4.5 vs Jaguars (total 49.5, ML +170 / −205) · Seahawks −13.5 vs Cardinals (total 45.5, ML +575 / −850) · Bills −3.5 at Vikings (total 48.5, ML −170 / +140). Byes: Steelers, Titans.
Sources: SportsBrackets Week 9 · Covers Week 9
Sample player props for Thursday Night Football, Panthers @ Packers (Oct 29). Each line comes from a different publisher/book — the small disagreements between them are the point.
Passing yards: Over 253.5 (−111) at BetDecider; BetMGM lists the line at 252.5. Rushing: Over 10.5 (−112) vs BetMGM’s 9.5. Passing TDs 1.5 at both; SI’s DK-sourced pick: Over 1.5 (−111).
Season at snapshot: 1,712 yds (285.3/g), 12 TD, 4 INT; over 253.5 in 5 of 6 games. BetMGM career context: 253+ yards in only 12 of 53 career games, career avg 193.7; 2+ pass TD in 17 of 53; 10+ rush yards in 27 of 53 (avg 14.6).
Love pass yards 262.5, Over −113; pass TDs 1.5, Over −148; rush yards 5.5, Over −102. Season: 1,958 yds (279.7/g), 13 TD, 5 INT.
Christian Watson receiving: 72.5, Over −113 (595 yds, 6 TD). Josh Jacobs rushing: Over 58.5 (−110, via DraftKings). SBR’s TNF card (Oct 26): Chuba Hubbard Over 14.5 receiving (−112, DK), Jacobs Over 77.5 rush+rec (−113, FanDuel), Watson Under 26.5 longest reception (−112, FanDuel).
Prop boards sell hit rates: e.g. Jared Goff Over 263.5 (−112, DK) with an “80% L10 hit rate” and 301.8 L10 average; Mark Andrews Under 29.5 receiving (−113, BetRivers), 80% L10.
A hit rate is not a probability. One prop site states it plainly: “The historical hit rate is context, not a forecast of an 80% chance to win.” Your app should show a distribution — projected mean + probability of clearing any line — with sample sizes, not a dressed-up hit rate.
Into Game 5 (Wed Oct 28, Tropicana Field): BetMGM MIL −160 / TB +135, run line −1.5, total 6.5. Second aggregator: MIL −156 / TB +131, −1.5 (+119), O/U 6.5 (−119/−101); series MIL −1429 / TB +850 (Sports Interaction, updated Oct 29). Covers after Game 4 quoted Polymarket at Brewers 91% / Rays 10%.
Verify before reuse: Sportradar’s results list shows Game 5 finished Rays 1–4 Brewers — this market was settling as the snapshot was taken. Sources: Milwaukee Journal Sentinel · Canada Sports Betting · Covers
Hawks −1.5 vs Cavaliers, total 238.5 — numberFire prediction Hawks 55.7%. Pelicans −5.5 vs Kings, ML −225/+188, total 237.5 — prediction Pelicans 63%.
Wild −192 / Islanders +158, total 5.5, puck line Wild −1.5 (+128) — prediction Wild 61.2%. Blues −114 / Bruins −105, total 5.5.
31 states + DC have live statewide online betting (39 + DC have some legal betting, as of Aug 26, 2026). Footprint leaders: DraftKings (27 + DC), BetMGM (23–24), FanDuel (24 + DC), Caesars (21–24), Fanatics (19–24, contested), bet365 (17), BetRivers (15), Hard Rock (10, sole book in Florida), theScore Bet (20). Filter by what each does best.
Across every review and first-party page in the research, in-app research stops at: (1) raw team leaders / season averages, (2) unadjusted ATS & over-under trend counts, (3) bet% vs handle% splits, (4) prop hit rates in editorial content. No major book shows a devigged fair probability next to its own price, an opponent-adjusted EPA rating, a player distribution, a quantified points impact for an injury or 18-mph wind, or your own CLV history. A probability-first app that shows its work on exactly those five items is differentiated from every book above.
Caveat: sportsbook apps are geo/age-walled to a signed-out browser, so per-book stat descriptions rest on first-party web surfaces (FanDuel Research, BetMGM blog, DK Network) and dated third-party reviews — not hands-on app inspection. Features also vary by state.
These are the exact calculations the app runs on every game. All maths runs in your browser on the numbers you type — nothing is fetched live.
Enter both sides of a two-way market in American odds. Proportional devigging divides each implied probability by their sum.
Worked example pre-loaded from the research: −152/+138 → raw 60.32% / 42.02%, vig 2.33%, fair 58.94% / 41.06%, fair line ≈ Side A −143.6 / Side B +143.6 (the −152 favourite stays the favourite; the +138 underdog stays the underdog). Note: proportional, additive and Shin methods disagree — the choice matters most on lopsided moneylines; Shin exists because books shade favourites less. Source: TeachersBet no-vig calculator · MarketMath
Enter your model’s probability and the best price you can actually get. Full Kelly assumes your probability is exact — it never is — so the product default is ¼ Kelly with a hard cap.
Formulas: EV = (Pwin × profit) − (Plose × stake); Kelly f* = (bp − q)/b. Practitioners default to quarter/half Kelly (half Kelly ≈ 75% of full-Kelly growth with far less variance); pros commonly stake 1–3% of bankroll per game. Sources: MarketMath EV/Kelly · Closer · SmartBet value guide
1) Sharp-book benchmarking — a stale soft-book price vs Pinnacle’s devigged price. Needs no model at all; it’s most of what OddsJam-style tools sell.
2) A custom model that disagrees with the sharp consensus.
3) Promo / boost extraction.
Your price vs the price at kickoff. Beating the close consistently is “the strongest single proxy for long-run skill,” because the closing line has absorbed all information. Books themselves use CLV to identify and limit sharp accounts. PFF data (2007–2020): closing spreads picked winners at 65.9% vs 63.5% for openers — lines sharpen through the week, so model value decays toward kickoff.
Academic anchor usually cited: Levitt (2004); Shin (1991/92) — via secondary summary, cite as such. SmartBet · PFF
Bin predictions (0.05-wide), compare predicted vs observed, hold out one season at a time, score with Brier score & log loss, recalibrate (Platt / isotonic). Sobering benchmark: adding the pregame spread to nflfastR’s in-game win-probability model cut cross-validated error from 27% → 23% and log loss 0.52 → 0.44 — the betting line is itself one of the strongest features available.
The standard pipeline, in the order you’d build it. Picking winners straight up is the easy, mostly worthless part — favourites win most of the time and the price already knows it. The game is calibration and price.
Documented FiveThirtyEight-style settings: K=20, home advantage = 48 Elo points, between-season mean reversion of ⅓ toward 1505, spread ≈ diff / 25. A maintained independent Elo (v4.2026.07) adds margin, rest/travel, overtime, parity reset and “meaningless game” flags — and states plainly it has no awareness of injuries, rosters or coaching changes.
EPA = expected points after the play minus before it, given down, distance and field position. Team offensive + defensive EPA/play correlates with wins at r ≈ 0.75–0.85 in-season and is more stable year-to-year (r ≈ 0.40–0.50) than yards. Production models use rolling, opponent-adjusted splits (pass vs run, early-down, success rate blended in because EPA is explosive-play noisy) and weight pass EPA more — rushing efficiency is less stable week to week.
Logistic regression or gradient boosting on rating diff, EPA splits, QB features, rest/travel and weather — or two-stage: project each team’s points, then convert projected margin + total into win/spread/total probabilities via a margin distribution (final margins cluster with SD ≈ 13–14 pts, piling up on 3 and 7). FiveThirtyEight-style systems Monte-Carlo the game thousands of times off the adjusted spread; Dimers describes the same commercially. Leakage warning: the honest worked example (10,923 games, 1980–2024) chose its model on pre-Dec-2003 data only, then evaluated walk-forward on 5,460 games.
Quarterback (biggest): Nate Silver’s ELWAY layers a QB rating (QBERT) onto Elo, scaled by how QB-dependent the offence is, with return dates treated probabilistically. Rule of thumb: an elite QB absence moves the line 3+ points.
Other injuries: weight by role & production share (WR by target share, RB by carry+target share, defender by tackle/sack share), capped, QBs handled separately.
Home field: historically ~3 pts, now ~1.5–2 pts league-wide with large venue variance (ELWAY uses 1–3.5 and argues the decline is overstated).
Rest/travel: bye ≈ +1 pt, each extra rest day ≈ +0.3 pts, plus time-zone/distance effects.
Weather: mainly a totals/passing adjustment (see §6), not a spread adjustment, except extreme wind.
Prototype stack from the research: nflverse (free) for the model, The Odds API (free → $30/mo) for prices, Pinnacle/Circa as the sharp devigging anchor. Move up only when latency or prop depth — not headline book count — becomes the bottleneck.
| Provider | Coverage | Free tier | Paid (as published, Oct 2026) |
|---|---|---|---|
| The Odds API the-odds-api.com ✓ verified live | 100+ sports, 100+ books; US: DK, FD, BetMGM, Caesars, Bovada, MyBookie + more; EU incl. Pinnacle; h2h, spreads, totals, outrights, props/alts for selected sports; history back to 2020 on paid | FREE — 500 credits/mo, all sports, most books, all markets, no historical | $30/mo = 20k credits · $59 = 100k · $119 = 5M · $249 = 15M. ⚠ Credits burn per market × region — broad polling costs far more than the headline number. |
| OddsJam (API) | 100+ (consumer pages say 150+) books incl. FD, DK, BetMGM, Caesars, BetRivers, Hard Rock, Fanatics; NFL/NBA/MLB/NHL/NCAAF/NCAAB/soccer/golf/MMA/tennis | No self-serve API free tier — sales-gated, pricing on request | Consumer subs are separate: Gold ~$199.99/mo, Global $399.99–$499/mo, Platinum $999/mo — funnels disagree, confirm at checkout. Budget ~$200–$500/mo. |
| SportsDataIO ✓ verified live | 13 sports; play-by-play, injuries, lineups, odds from 10+ books, projections, player news; Vault = 10+ yrs history | Free trial = scrambled (fake) data — unusable for analysis. Discovery Lab free = last season only | Discovery Lab (personal, next-day delayed, 100–1,000 calls/day, NOT licensed for commercial redistribution): $99/mo Fantasy or Odds, $149/mo both. Commercial Leagues API: real-time, unlimited — sales-only, price by use case. |
| SportsGameOdds | 85+ books, props & alt lines, settlement, deeplinks; WebSocket on top tier | Yes — 2,500 objects/mo (card required) | Rookie $149/mo ($99 annual) · Pro $499/mo ($299 annual), unlimited objects. Transparent, self-serve — credible step-up. |
| SharpAPI | 43 books incl. exchanges (Betfair, ProphetX) & prediction markets (Polymarket, Kalshi); Pinnacle as +EV/no-vig reference; SSE/WebSocket | Free $0: 12 req/min, 2 fixed books (DK, FD), pre-match, 60-sec delay | Hobby $79/mo · Pro $229/mo · Sharp $399/mo (25 books, sharp books); streaming +$99/mo |
| TheRundown | 25+ books, 550+ market types, 23 leagues; unified schema incl. Kalshi/Polymarket/Novig/ProphetX | Free: pre-match REST from DK/FD/BetMGM, 5-min delay, 20,000 data points/day | $49 Starter (60-sec delay) → $149 Pro (30-sec) → $399 Ultra (real-time + WebSocket) → $2,499 |
| Unabated / OpticOdds / Sportradar | Unabated: vig-free “Unabated Line,” props, SGP pricing, WebSocket. OpticOdds: 200+ books, enterprise desks | Unabated constrained free API (15-sec delay, deeplink-traffic condition) | Unabated ~$799/mo (5-league) to ~$2,999/mo all sports — treat as ~$1–3k. OpticOdds $5,000+/mo per sport. |
⚠ Lookalike warning: theoddsapi.com (no hyphen) publishes different tiers ($0/25 req/day limited NFL/NBA/MLB h2h; $29 Pro; $99 Business). Do not build procurement on it — the ladder above is the-odds-api.com, verified live and corroborated by independent comparisons. Sources: the-odds-api.com · SharpAPI pricing · TheRundown comparison · SportsGameOdds alternatives · SharpAPI docs
The default NFL build block. Play-by-play 1999–present, 370+ columns incl. EPA, WPA, air yards; player & team stats 1999–present; schedules incl. Vegas lines & weather fields; rosters 1920–present; injuries 2009–present; snap counts 2012–present; Next Gen Stats 2016–present; PFR advanced 2018–present; FTN charting 2022–present; participation 2016–2024; draft picks 1980–present; contracts, trades, combine, depth charts. Via nflreadr (R) / nflreadpy (Python).
Licence caution: “free to download does not always mean free for commercial use or redistribution” — underlying sources (PFR, FTN, OTC) have their own terms. nflreadpy · sportsapis.dev datasets
Historical stats since 1920; scoring complete all-time, yardage official from 1932, advanced stats since 2018. Aug 2025 expansion pushed play-by-play, targets and expected-points summaries back to 1978 (from 1994). Caveat from PFR itself: its play-by-play is unofficial and diverges on sacks/tackles due to scorer inconsistency. Site free; Stathead is the paid query/export layer. Use for research/validation — scraping at scale is against typical terms, not a production feed unless licensed.
Undocumented JSON: site.api.espn.com/apis/site/v2/sports/football/nfl/scoreboard, /teams, /roster, /schedule, /statistics; core API for season & career player stats. Widely used for live scores and current-season stats. No contract, no SLA, endpoints die (“many of the previously listed URLs are now dead”). Prototype & fallback only — never the sole production source.
Sportradar (official partner NFL/NBA/NHL/MLB/NASCAR) and Genius Sports (many NCAA / international rights): real-time feeds, PBP, props data, advanced metrics. Third-party estimates conflict by an order of magnitude — “$10,000+ annually for research tiers” vs “$10,000+/mo” vs €1,300–4,990+/mo — no public rate card; treat all figures as unverified estimates. Get a quote only if official-data licensing becomes a requirement. SportsDataIO Leagues is the commercial middle ground. Also useful: MySportsFeeds (free for personal use), PFF grades (paid feature source), NFL Next Gen Stats (in nflverse from 2016), NBA nba_api.
For Sunday games: Practice Reports Wed–Fri use DNP / Limited / Full; Friday’s Game Status Report uses Out / Doubtful / Questionable. Game-day inactives land ~90 minutes before kickoff. Machine-readable: nflverse load_injuries() (2009–present), SportsDataIO Leagues, FantasyPros /players /news /injuries (with practice-report probabilities).
Design implication: refresh & re-version the probability at Wed, Fri-final, and 90-minute-inactives points. Late status changes are where both edges and embarrassments live. Cautionary case this week: a star QB listed as a full participant Friday was ruled out Saturday, prompting a league review — status fields can be wrong, so show when a probability was computed and on what snapshot.
WebMD injury-report explainer · FantasyPros API · TheSpreadSports stack notes
National Weather Service api.weather.gov — free, official, public domain, matched to kickoff time & venue; Open-Meteo as non-US fallback; roof status & schedule from nflverse (which also carries historical game weather for training). A live NFL weather board flags wind impact at 15 mph. Relevance is mainly totals & passing/kicking props: wind >15 mph “significantly impacts passing games”; rain/snow raise fumble risk and cut passing efficiency. One DFS model: 20 mph + rain ≈ −15–20% passing yards (directional, lower-authority sources — encode as a soft adjustment, not gospel).
RotoWire player news (the FanDuel Research injury items are this format), ESPN + Sleeper designations, SportsDataIO player news (commercial), FantasyPros news endpoint. Beat-writer/social news is faster than official reports but unstructured and unverifiable at scale — use it to trigger a re-check of official status, never as a direct model input.
The +EV / CLV tooling cluster sits at $200–$500/mo. The $15–$30/mo cluster lacks either the model or the pricing engine. Nobody sells both, transparently, at a consumer price.
| Product | What it does | Published price | Limit for this build |
|---|---|---|---|
| Dimers (Pro) — closest existing product | Simulation/ML picks, prop projections, parlay picker, trends, futures line shopping, “Signals” edge ranges, Dimebot AI. Models “refresh daily with injuries, weather, news”; “every event stress-tested with thousands of simulations” ✓ verified live | $14.99/wk · $29.99/mo · $199.99/yr; 3-day free trial; partner offers give 1–3 free months via book sign-ups (one listed until Nov 29, 2026 via DraftKings) | Probability-first like this brief — so differentiation must be transparency: publish the model’s own calibration & CLV per market, show devig-vs-sharp and model-vs-market separately, explain each pick. None of which Dimers markets. |
| OddsJam | +EV finder (vs sharp consensus), arbitrage scanner, line shopping 150+ books, prop tools, parlay builder, bet tracker | 7-day trial; Gold ~$199.99/mo (40+ US books) · Global $399.99–$499/mo · Platinum $999/mo — funnels differ | +EV is market-relative (stale price), not a game model. At $199/mo you need $200+/mo of edge just to cover the sub — implying a $1,000–$3,000+ bankroll for the maths to work. |
| Action Network (PRO / Labs) | BetSync tracking/grading, public betting splits, line movement, PRO signals/projections; Anytime TD Machine — live TD odds vs 5 seasons of graded hit rates (75,090 bets), CSV export | PRO $119.99/yr · $24.99/mo · $14.99/wk (no standard free trial). Labs $249/mo or $1,499.88/yr — several tools assumed to be in PRO actually live in Labs | “Stronger as a workflow tool than as a pure ‘find me winners’ app”; its reports are “trend data rather than a modeling edge.” |
| Betstamp / Betstamp PRO | Free app: odds comparison 50+ books & prediction markets, best-line spotting, prop shopping, auto-graded tracking, ROI & CLV sliced by sport/league/type/side/book. PRO: real-time screen, “True Line,” Edge, 45+ prop markets | App free. PRO from $249/mo (application-gated, no free trial) + alts $129/mo + PPHs $99/mo = $477/mo fully loaded; annual saves up to 40% | Pro-priced, not consumer-priced: at $249/mo, a 2% edge needs ~$12,500/mo handle just to cover the screen. |
| Pikkit — closest free competitor for grading | Free auto-sync tracking from 30+ books/DFS/prediction markets (read-only BookSync), P&L / ROI / CLV breakdowns, line shopping across synced books, AI queries over your own history | Free tier. Pro $29.99/mo or $199.99/yr (7-day trial) adds CLV analysis, SGP line shopping, odds alerts, web version | Tracking-first, not a model: tells you whether you beat the close, not what the true probability is. |
| Also in market | Juice Reel (free sync + social record) · Rithmm (build-your-own AI models, ~$29.99/mo per OddsJam comparison) · Unabated (vig-free line, pro tier) · prop hit-rate boards (Win Daily, FastStat, ProPropLabs: L5/L10/season hit rates + best price + a model score) | Mostly free–$30/mo at the prop-board end | The tail proves demand for “hit rate + best price + model probability on one screen” — and most of it is thin: hit rates dressed as probabilities. A calibrated, sample-size-aware version is a real gap. |
Sources: Dimers subscription (verified live) · BetSmart — Dimers · Fault.bet — OddsJam · BetSmart — Action PRO · BetSmart — Betstamp PRO · Betstamp tracking · BetSmart — Pikkit
1. Two probabilities, labelled, never blended silently — (a) devigged sharp-consensus and (b) your own model, edge decomposed into “stale price” vs “model disagreement.” Incumbents sell one or the other.
2. A public, graded model record — every probability timestamped, graded for Brier/log-loss, calibration & CLV, per market. Nobody above markets its own calibration curve.
3. Explainability — name the factors that moved the number (QB status, EPA split, rest, wind) with point values.
4. Props as distributions, not hit rates — projected mean + probability of clearing any line the user enters, sample sizes shown.
5. Consumer price for the full loop — a $15–$25/mo tier with both model + pricing engine, plus a genuinely useful free tier (one game fully worked per week), undercuts the $200–$500 cluster.
6. No affiliate conflict — most incumbents monetise book sign-up offers. A subscription-only product can say “no edge — don’t bet” on a full slate. That honesty is also the only defensible marketing given §8.
Measurement and line shopping deliver value on day one, before any model exists. The model arrives into an app that can already prove whether it works.
The Odds API free tier (500 credits/mo), strict credit budgeting per market × region. Multi-book board with timestamps, best price highlighted, sharp-anchor (Pinnacle/Circa) devigged fair line next to every soft price. Useful with zero model.
nflverse pipeline: Elo + opponent-adjusted EPA logistic model. Walk-forward backtest, hold-one-season-out calibration, Brier/log-loss dashboard. Publish the calibration plot in-product from the start.
QB starter-vs-replacement adjustment first (worth 3+ pts), then production-share injury weights. Wed / Fri / 90-min refresh & versioning. Prop projections as distributions with sample sizes. Weather (NWS, free) into totals & passing props, wind flag at 15 mph.
Bet log with automatic grading, ROI by market, your CLV vs the close, and the model’s own graded record side by side. Fractional-Kelly (¼, capped) stake suggestions throughout — never full Kelly.
Step up to SportsGameOdds / SharpAPI when latency or prop depth bites. Add NBA/NHL only after the NFL model has ≥1,000 graded probabilities. It all runs comfortably on a home stack at prototype scale.
NFL lines are among the most efficient prices on Earth. This section is not a disclaimer bolted on — it’s the design constraint the whole product is shaped around.
Elo author’s own warning: the ratings are “not intended for use in informing sports betting decisions.” Sources: NFL Elo FAQ (8,400+ games, 1994–2025) · PFF
The walk-forward model (leakage-free, bet only above break-even) landed at 51.7% overall — below break-even — with unders at 56.0% in its sample. One project isn’t proof, but it’s the shape of honest out-of-sample results from public data: at or just under the vig line. Only 21.5% of games land within 3 points of the closing spread; every spread bucket’s average margin lands within ~1 point of its line — there is no bucket-level free lunch.
Early-week stale lines (closers beat openers on winner-picking) · player props (posted late, thinner, more role/injury-sensitive — but lower limits) · promo/boost value. Not Sunday-afternoon main lines. Practitioner-claimed sustainable ROI for disciplined value bettors: 2–4% per bet — “anyone promising 10%+ ROI is selling something” (a value-betting guide’s claim, not audited data — label it as such in-product too).
With outcome SD ~13 points and true edges of 1–3%, hundreds of bets are needed before results distinguish skill from luck — one +EV tool’s own FAQ concedes a +5% EV bet “still loses more often than it wins,” and its illustrative simulation includes a >25-unit drawdown. Hence: fractional Kelly only, CLV (not W/L) as the in-product skill gauge, ≥1,000 graded probabilities before claiming any model edge. And if it works: soft books limit consistently profitable / CLV-positive accounts “within weeks to months” — sharp books (Pinnacle, Circa) are the exception. Plan churn around degrading access to soft prices.
Bottom line: build it as a measurement and line-shopping product with a model attached — devigging, best-price surfacing, timestamped probabilities, CLV and calibration grading are deliverable value on day one. Market it on transparency. Do not market it on beating the NFL market long-term; the sources above don’t support that claim for retail models, and the one number every sophisticated user will check — CLV against Pinnacle/Circa closes — will expose it if it isn’t true.
Sportsbook in-app screens: geo/age-walled to a signed-out browser — §1 rests on first-party web surfaces and dated reviews, not hands-on app inspection; features vary by state.
OddsJam API & SportsDataIO commercial pricing: both sales-gated, nothing published; consumer OddsJam prices conflict across funnels ($199.99 Gold / $399.99 Global vs a $499/mo Global page).
Sportradar / Genius pricing: third-party figures conflict by an order of magnitude ($10k/yr vs $10k/mo) — all unverified estimates.
The Odds API identity: a second, conflicting pricing ladder exists at the lookalike theoddsapi.com — confirm the legal entity / API keys at sign-up.
World Series Game 5: sources dated Oct 28–29 disagree in tense (pregame odds vs a listed 4–1 Brewers final) — re-check before reuse.
Levitt (2004) / Shin CLV claims: cited via a secondary guide, not the papers themselves.
Landscape source: state counts & footprints via Lines.com sportsbook reviews (Aug 26, 2026 ranking).