How do football clubs like Brighton & Hove Albion and Brentford consistently identify undervalued talent before the rest of the world catches on?
The answer lies in quantitative modeling—specifically Poisson goal intensity models and probabilistic modeling pioneered not in scouting departments, but inside private sports betting syndicates like Starlizard and Smartodds.
However, mathematical models have a well-known blind spot: they look strictly at historical numbers. A Poisson model has no idea if a top striker suffered an injury during yesterday's training session, if a manager is facing immediate dismissal, or if a team is resting key players ahead of a European semifinal. On the other hand, relying purely on human intuition or generalist AI prompts leads to emotional guesses without any mathematical grounding.
To solve this, I built a serverless pipeline that bridges hard quantitative data with real-time qualitative match intelligence.
Today marks Day 1 of a live benchmark test: pitting seven frontier AI models against one another using identical match data to see which one delivers the sharpest betting analysis.
The entire workflow runs in the cloud via GitHub Actions without requiring any dedicated local server infrastructure. It operates across three distinct stages:
With a single manual trigger in GitHub Actions, a Python script requests precomputed model outputs from OddsLine API. Rather than looking at superficial standings, it extracts:
-
True Win/Draw/Loss Probabilities: Stripped of public sentiment and bookmaker bias.
-
Poisson Strength Ratings: Attack and defence capabilities relative to league averages ($1.00 = \text{league average}$).
-
Elo Ratings: Adjusted for tactical home-field advantage.
-
Goal Distributions: Independent probabilities for Over/Under 2.5 goals and Both Teams To Score (BTTS).
Using live search retrieval, current decimal odds are pulled from sharp bookmakers (such as Pinnacle or Bet365) across 1X2, Asian Handicap, and Over/Under lines. The market-implied probability is calculated directly alongside the bookmaker's overround (margin), establishing the exact benchmark that our quantitative model needs to beat to prove positive expected value ($+\text{EV}$).
The combined data is fed into a structured prompt designed to force the LLM to think like a syndicate risk manager rather than a casual fan. The prompt directs the model to:
-
Conduct live web searches: Scour local media for verified injuries, suspensions, lineup changes, fixture fatigue, and dressing-room morale.
-
Synthesize numbers with context: Determine whether bookmaker pricing is a genuine market inefficiency ($+\text{EV}$) or merely a rational reaction to a missing star player.
-
Predict game flow & exact scorelines: Project tactical dynamics (e.g., low defensive block vs. vulnerable transition defense).
-
Issue a graded decision: Assign a strict classification (Grade A, Grade B, Grade C, or PASS / False Value).
<
Hot Forum Topics
