Open-source AI sports betting projects are proliferating on GitHub, and three in particular demonstrate the range of techniques builders are applying — from deep learning match prediction to ensemble classifiers for totals markets. Here is what each project does, how it works, and where the underlying math connects to production betting agent design.
BettingAI: Deep Learning Football Prediction
BettingAI is a four-module Python pipeline that ingests football data from fotmob.com into PostgreSQL, trains a TensorFlow/Keras neural network on historical match features, and serves predictions through a FastAPI interface.
The architecture maps cleanly to the Agent Betting Stack:
| Module | Function | Stack Layer |
|---|---|---|
| Module 1 | Data ingestion and PostgreSQL storage | Layer 1 — Identity |
| Module 2 | Feature engineering and preprocessing | Layer 4 — Intelligence |
| Module 3 | Neural network training with K-Fold CV | Layer 4 — Intelligence |
| Module 4 | Live odds fetching and prediction serving | Layer 3 — Trading |
The model uses StandardScaler normalization, Batch Normalization, and Dropout to prevent overfitting — standard techniques covered in our regression models guide. Module 4 fetches upcoming matches within a five-day window and pulls live odds across leagues, making it a functional (if basic) end-to-end prediction agent.
What builders can learn: The modular separation between data ingestion, model training, and prediction serving is the right pattern. Where BettingAI stops short is bet sizing — it predicts outcomes but does not calculate expected value or apply Kelly Criterion to determine optimal stake. Adding an OpenClaw EV Calculator skill or Kelly Sizer skill would close that gap.
DGFantasy Optimizer: +EV Player Props at Scale
DGFantasy takes a different approach. Rather than predicting match outcomes, it focuses on player prop markets — specifically finding lines where DFS platforms like PrizePicks price a prop differently than the broader sportsbook market.
The core logic is straightforward expected value math: if multiple sportsbooks have a player’s strikeout line at Over 5.5 at -145, but PrizePicks still shows 5.5 at even money, the Over side has positive expected value. DGFantasy automates this comparison across hundreds of props daily.
This is a direct application of the concepts in our expected value guide — the same framework applies whether you are comparing DFS lines or building an autonomous agent that trades prediction market contracts. The key insight is that you do not always need a sophisticated predictive model. Sometimes the edge is simply comparing prices across fragmented markets, which is exactly what an OpenClaw Odds Scanner skill does for sportsbook odds and an Arb Finder skill does for cross-market arbitrage.
What builders can learn: DGFantasy shows that +EV discovery does not require deep learning. Cross-market price comparison — effectively an odds arbitrage scanner — is one of the highest-ROI tools in the stack. Our Vig Index applies the same principle to sportsbook vig comparison at scale.
AIFootballPredictions: XGBoost for Over/Under 2.5 Goals
AIFootballPredictions by MauroAndretta is purpose-built for a single market: predicting whether a football match will have over or under 2.5 total goals. It covers five European leagues (Serie A, EPL, Bundesliga, La Liga, Ligue 1) using Scikit-Learn and XGBoost.
The pipeline follows a clean four-stage structure:
- Data acquisition — Downloads and merges historical match data across leagues and seasons
- Preprocessing — Feature engineering, missing value handling, feature selection
- Model training — Multiple models with hyperparameter tuning, combined into a voting classifier
- Prediction — Generates formatted predictions for upcoming matches
The Over/Under 2.5 market is particularly well-suited to Poisson distribution modeling because goal-scoring in football follows approximately Poisson-distributed patterns. Our Poisson guide walks through building the exact same type of match probability matrix this project targets. The project’s use of XGBoost gradient boosting also connects to the broader regression models covered in the Math Behind Betting series.
The expected goals (xG) guide goes even deeper into the football-specific modeling that projects like this depend on — from shot-level logistic regression to match-level Poisson predictions.
What builders can learn: Specializing in one market (Over/Under 2.5) rather than trying to predict all outcomes is a smart constraint. Voting classifiers that ensemble multiple model types (XGBoost, random forest, logistic regression) typically outperform any single model — a principle that maps directly to multi-agent architectures where different intelligence layer agents vote on a final decision.
From Projects to Production Agents
All three projects demonstrate real ML techniques applied to real betting markets. But none of them are production betting agents yet. The gap between “model that predicts” and “agent that trades” requires the remaining layers of the Agent Betting Stack:
- Identity and wallet infrastructure to manage keys and funds
- Trading execution to place and manage positions
- Risk management via Kelly sizing and drawdown controls
The OpenClaw framework provides modular skills for each of these layers, and the full Math Behind Betting series covers the quantitative foundations — from expected value to Poisson modeling to Kelly Criterion — that turn raw predictions into sized, risk-managed positions.
If you are building on any of these projects, the next step is connecting your model’s probability outputs to an EV calculator and a position sizer. The math is straightforward. The open-source prediction market bots guide covers additional repos worth forking.