EPLValue
Arsenal — Chelsea
1 — home win
Model54%
Odds2.05
Edge+4.2%
DATA · ML0 matches · 175 features · one verdict
The model has seenmore matchesthan an analyst sees in a career
ScalpBet computes probabilities for 50+ markets on every match — built on big data and an ensemble of ML models. Not an opinion. Not a “lock”. A number with a clear method.
18+Analytics is a tool for decisions, not a guarantee of results. The responsibility for any bet is always yours.
INFERENCE · 175 FEATURES LIVE · 67
0
matches in training · 1993→2026
Home win (1)54%edge +7.3%
Over 2.5 goals62%edge +4.1%
Both teams to score58%edge +1.8%
Handicap (−1.5) home44%edge −0.4%
Calibration: Brier 0.187model computes +50 markets →
edge = model probability − market-implied probability. Examples to illustrate the format.
Under the hood
01377 901unique matches at the core — 33 years of football, 1993→2026
0210.3Modds (open + close) — the model sees how the market moves
0312.2MStatsBomb events with coordinates and 360° — what really happened on the pitch
04487Kshots with real xG — who deserved the goal, not who got lucky
0550+markets computed — win, totals, handicaps, any score
Figures come from a data audit. Deduplicated across 7 sources: 562,536 source records merged without duplicates.
Methodology
From raw data — to one number
Four steps. 7 sources in. A probability you can compare to the odds out.
1
Collect
Data
7 sources: history, odds, advanced stats, events with coordinates, market-implied probabilities. 377,901 matches, merged without duplicates.
2
Compute
175 features
Every match becomes a vector of 175 features: attack and defence strength, form, context, odds movement. Machines need numbers — we prepare them.
3
Run
Ensemble
Not one model, but a team: CatBoost experts, a neural net, Poisson, league specialists. Each looks from its own angle, the meta model combines them.
4
Deliver
Probabilities
Probabilities for 50+ markets plus edge vs the market. You see not just “who's the favourite”, but where the odds lag the model.
One pipeline, no manual tuning to fit the result.
Under the hood
Not one model, but an ensemble of specialists
Each method covers what the others miss. On the left — what it's called. On the right — what it gives you.
Ensemble = lineup
Each one covers its zone
Like a 4–3–3: the defence holds the teams' base strength, the midfield connects, the attack hunts for patterns. The meta model is the playmaker in the centre: it gathers every “player's” opinion into one final probability.
- GK Poisson — base of any scoreline
- Defence 4× CatBoost — attack & defence strength
- Centre MoE · META hub · league specialists
- Attack Neural net · value · calibration
expert model meta model
4× CatBoost experts →
Base attack and defence strength of the teams — the foundation everything else rests on.
Neural net (TFT / Transformer) →
Catches hidden form patterns invisible in the table or the season average.
Bivariate Poisson + Dixon-Coles →
Probability of any scoreline, not just “who wins” — hence honest totals and handicaps.
Meta model →
Combines every expert's opinion into one final probability, without bias.
Mixture-of-Experts →
A dedicated specialist for each match type: top derbies and mid-table games are scored differently.
League specialists →
A separate model for each league's character: the EPL and a lower division live by different rules.
Value / Kelly →
edge = model probability − implied odds; Kelly suggests a sensible stake size.
Calibration →
When the model says “60%”, it happens roughly 60% of the time. Brier 0.187.
We kept the jargon for those who check. If it doesn't matter to you — read the right column.
Data
Seven sources. Zero blind spots.
One source lies — the others correct it. We merge them into a single picture and deduplicate: 562,536 records → 377,901 clean matches.
01
AF API-Football
Lineups, injuries and live — the model knows who actually takes the pitch.
02
FD football-data.co.uk
33 years of match history — the long memory the models learn from.
03
FB FBref
Advanced stats — deeper than “shots and possession”.
04
U Understat
xG at the level of each shot (487K) — who deserved the goal, not who got lucky.
05
SB StatsBomb
Events with coordinates and 360° (12.2M) — what really happened on the pitch.
06
OP OddsPortal
Odds movement (10.3M, open + close) — where the market shifted and why.
07
PM Polymarket
Market-implied probabilities — the “crowd with money” as one more cross-check.
Dedup
7 sources · merged without duplicates · 562,536 → 377,901.
LIVE BROADCAST67
The model has seenmore matchesthan an analyst sees in a career
An ensemble of ML models turns 377,901 matches into calibrated probabilities for 50+ markets. No emotion, no hype — just data.
SCOREBOARD · PROBABILITIES0matches in training · 1993→2026CALIBRATION · Brier 0.187
1 · WIN54%edge +7.3%
O2.562%edge +4.1%
BTTS58%edge +1.8%
HCP −1.544%edge −0.4%
BTTS AT HT31%edge +5.0%
Preview
This is the feed
Every signal is the model's probability against the market's odds. Open cards are real in format. The rest — inside.
Serie AValue
Inter — Lazio
Over 2.5 goals
Model58%
Odds1.85
Edge+3.1%
No value under these filters today.
That's normal: the model shows edge only where it exists. No signal — no forced bet. Check back tomorrow or turn on alerts.
Numbers in the open cards illustrate the signal format.
Honestly
Backtest on history
We have few live bets yet, and we won't draw a pretty curve out of thin air. We show what we have, honestly: how the strategy would have behaved on historical data.
Historical simulation, not real bets.
This is a backtest on past matches. Past results do not guarantee future ones. A real track record from live bets will appear here as it accumulates — without embellishment.
Simulation period
2019 → 2025
Matches in backtest
118 430
Signals with edge > 3%
9 120
Return in simulation
+6.4% sim
On principle we don't write “ROI 0%” and we don't pass a simulation off as a live result. When real bets come, you'll see real numbers.
Calibration chart
When the model says 60% — it happens ~60% of the time
ideal ScalpBet model
EDITION · MLNo. 377.901 · 1993→2026
The modelhas seenmorematches.
Probabilities for 50+ markets on every match — built on big data and an ensemble of ML models. Not an opinion. Not a “lock”. A number with a clear method.
0matches at the core
0features per match
50+markets in the models
0.187Brier · calibration
Access
Choose your depth
No payments yet — we open access by waitlist. The earlier you join, the sooner you get the feed.
Free
Try it
$0
A lead magnet — to get the format.
- Yesterday's results and review
- 1 free signal per day
- Access to methodology and data
Popular choice
Pro
For value hunters
Soon · locked in for early birds
The main thing: a real-time feed.
- Everything in Free
- Full signal feed in real time
- Stats and track record per signal
- Catch value before the odds close
- Alerts on new edge
VIP
Maximum control
Soon · locked in for early birds
For those building their own.
- Everything in Pro
- AI analysis of any match on demand
- Export predictions to your sheet or bot
- Priority support
We open access in waves by waitlist. Prices will be locked before payments launch — early members get the best terms.
Questions
The essentials, briefly
Is this a betting service or a casino platform?
No. ScalpBet is data analytics. We compute probabilities and show where the odds diverge from the model. The decision to bet — and the responsibility — is yours.
Do you guarantee profit?
No, and anyone who guarantees it is lying to you. We give probabilities and edge — a statistical advantage over the long run, not a “lock” on a single match. Losing streaks are inevitable in betting.
Where do numbers like 377,901 matches come from?
From a data audit across 7 sources (history, odds, events, xG, market-implied probabilities), merged and deduplicated: 562,536 source records → 377,901 unique matches.
What do edge and Kelly mean?
edge = model probability − implied odds: how much the market underrates an outcome. Kelly is a formula for a sensible stake relative to your bankroll, so you don't burn out on emotion.
Why do you show a backtest, not real ROI?
Because it's honest. We have few live bets yet — we don't pass a historical simulation off as a real result and we don't draw “ROI 0%”. Real numbers will appear as they accumulate.
How much does it cost and when do you launch?
Right now access is by waitlist, there are no payments yet. Leave an email or Telegram — we'll invite you in a wave and lock in the best terms for early members.
18+ Responsible gaming
ScalpBet is an analytical tool, not a call to gamble and not a guarantee of winning. Sports betting carries a risk of losing money. Analytics helps you make informed decisions, but does not remove the randomness of a single match. Play responsibly, stake only what you're ready to lose. If gambling stops being entertainment — stop and seek help.