A machine learning AI to predict NBA game outcomes—money lines (winners) and over/unders—integrating XGBoost models with daily-scraped DARKO data. Builds off an original project by kyleskom, expanded to include deeper synergy logic, advanced metrics (Weighted DPM, Off/Def splits, best lineups), and daily caching of scraped data.
Achieves ~69% accuracy on money lines and ~55% on under/overs (XGBoost). Outputs expected value (EV) for each team’s money line and optionally a Kelly Criterion fraction. We also incorporate DARKO (Daily Adjusted and Regressed Kalman Optimized) data for more robust synergy and clarity in final predictions.
- XGBoost for Money Line & Over/Under predictions, updated daily.
- Expected Value & optional Kelly Criterion bet sizing, factoring in the model’s win probability vs. odds.
- DARKO daily data for more accurate synergy:
- Summed team points to see if it aligns with the Over/Under or the chosen side.
- Weighted DPM & Off/Def splits to highlight team strengths.
- Best lineup ratings to see top net rating lineups.
- Player-level daily insights (top scorers, etc.) to see who’s driving projected performance.
- Caching of daily-scraped data (one date-based CSV per day) to avoid repeated Selenium calls. Optionally force re-scrape if needed.
- ASCII color-coded synergy output:
- “Moneyline synergy” (XGBoost side vs. Darko side),
- “EV synergy” (whether the chosen side’s EV is positive),
- Weighted DPM, advanced stats, top scorers—all displayed for easy interpretation.
- TensorFlow (optional, if you want the old NN approach)
- XGBoost – The main gradient boosting framework
- Numpy – Scientific computing
- Pandas – Data manipulation & analysis
- Colorama – Color text output in console
- Tqdm – Progress bars
- Requests – HTTP library (for odds or other data)
- Scikit-learn – Additional ML tools
- Selenium & webdriver_manager – For scraping the DARKO Shiny app pages
Python 3.11 is recommended.
- Clone the repo and install dependencies:
git clone https://github.com/notoctosting/NBA-ML-with-DARKO.git
cd NBA-ML-with-DARKO
pip3 install -r requirements.txt
-
Fetching Odds / Running XGBoost:
- Example:
This pulls the day’s betmgm odds from sbrodds if available, prompts the script to do XGBoost picks, and prints money lines, over/unders, expected value, etc.
python3 main.py -xgb -odds=betmgm
- Example:
-
Scraping DARKO Data:
- By default, the script will attempt to scrape the Daily (Player Per-Game), Lineup, and Current Skill CSVs for the day.
- The cache is date-based: e.g.,
darko_data/daily_projections_YYYYMMDD.csv. If that file already exists, we skip re-scraping unless you pass-force_dark.
-
Optional Flags:
-kc=> Show Kelly Criterion fraction on EV lines.-force_dark=> Force re-scrape of Darko CSVs even if they exist.
-
Output:
- After XGBoost lines and EV, you’ll see a synergy block with color-coded ASCII boxes, showing:
- XGBoost-chosen side & probability,
- EV for both sides,
- Darko daily sums (margin),
- “Moneyline synergy” (AGREE or DISAGREE) vs. Darko side,
- “EV synergy” (POSITIVE if your chosen side’s EV is > 0),
- Weighted DPM, Off/Def splits, best lineup rating,
- Top daily scorers from Darko’s player-level projections.
- After XGBoost lines and EV, you’ll see a synergy block with color-coded ASCII boxes, showing:
------------------betmgm odds data------------------
Los Angeles Lakers (-235) @ Charlotte Hornets (195)
... [XGBoost Predictions & EV lines] ...
Daily Adjusted & Regressed Kalman Optimized - DARKO
Scraping daily data for 20250127 => darko_data/daily_projections_20250127.csv
Scraping lineup data for 20250127 => darko_data/lineup_projections_20250127.csv
Scraping skill data for 20250127 => darko_data/current_skill_20250127.csv
================================================================================
===== EXPANDED DARKO + XGBOOST ANALYSIS WITH PLAYER-LEVEL CONTEXT =====
--------------------------------------------------------------------------------
| [Charlotte Hornets] @ [Los Angeles Lakers] |
--------------------------------------------------------------------------------
| XGB => side=HOME (home=58.3%, away=41.7%) |
| EV => Los Angeles Lakers:-16.85, Charlotte Hornets:22.92 |
| Darko => Los Angeles Lakers:102.2, Charlotte Hornets:107.5 (margin= -5.3 ) |
| Moneyline synergy => REDDISAGREE |
| EV synergy => GREENPOSITIVE |
--------------------------------------------------------------------------------
| Los Angeles Lakers => WDPM= 7.20, Off= 2.50, Def= 4.50, BestLU= 12.00 |
| Charlotte Hornets => WDPM= 0.50, Off= -0.80, Def= 1.40, BestLU= 0.50 |
--------------------------------------------------------------------------------
| Top scorers Los Angeles Lakers: [LeBron James=28.5pts] [A. Davis=24.3pts] |
| Top scorers Charlotte Hornets: [LaMelo Ball=25.9pts] [T. Rozier=23.1pts] |
--------------------------------------------------------------------------------
... [other games displayed similarly] ...
If you want to explore a Flask interface:
cd Flask
flask --debug run
This can be extended to display synergy data in a browser, though it’s primarily a console-based workflow.
- Gather the newest data:
cd src/Process-Data python -m Get_Data python -m Get_Odds_Data python -m Create_Games - Train or re-train XGBoost models:
cd ../Train-Models python -m XGBoost_Model_ML python -m XGBoost_Model_UO - Optionally do advanced synergy metrics with DARKO by scraping daily with:
scrape_dark_data_for_date(teams, ...)in your code.
All contributions and improvements are welcomed—particularly around advanced synergy logic, new metrics from the DARKO data, or improved scraping reliability.
This project is an expanded version of kyleskom’s NBA-Machine-Learning-Sports-Betting. Feel free to clone/fork and propose changes. All standard disclaimers: this is for educational and informational use; no guarantee of profit or success. Bet responsibly.