🏏 IPL Analytics Platform

An end-to-end cricket analytics pipeline built on ball-by-ball IPL data (Cricsheet, 2008–present). Covers data parsing, advanced metric engineering, three ML models, and a multi-page Streamlit dashboard.

🔗 Live app: iplanalysis-ja3jmqhz5kra7wfydke9xw.streamlit.app

---

Contents

• What it does
• How it's built
• Tech stack
• Project structure
• Running it locally
• Data source
• Status

---

What it does

The dashboard has 7 pages:

Page	What it shows
📊 Season Overview	Run rate, average innings score, and wickets trends across seasons
🏆 Team Performance	Win %, toss impact, chasing vs defending trends, phase-wise team strength
🧍 Player Analysis	Player impact scores, strike rate vs runs, phase-wise top batters, death-overs specialists
🏟️ Venue Insights	Par scores, chasing win % by venue, scoring patterns by ground
📈 Win Probability	Over-by-over win probability curve for any match in the dataset
🎯 Score Predictor	Predicts final innings score from the first 10 overs of input stats
🎲 Tournament Simulator	Monte Carlo simulation of championship and playoff probabilities

Open the Live app link above and click through these directly — that's faster than reading about them.

How it's built

Data pipeline

• Source: Cricsheet ball-by-ball JSON (json_files/), parsed into structured CSVs (parsers/)
• Covers 294K+ deliveries across 1,239+ matches (2008–present)
• Processed output feeds a reports/ directory split into tables/, advanced_metrics/, and modeling/

Models (trained via model1.py, model2.py, model3.py, win_probablity_model.py, tournament_simulator.py)

• Win Probability — RandomForestClassifier predicting batting-team win probability at each over checkpoint. Trained on engineered features: current score, wickets lost, run rate, target, required run rate, chase indicator. Evaluated on accuracy, ROC-AUC, and log loss — exact numbers live in reports/modeling/win_probability_metrics.json, generated at training time. Check there rather than this README for current values.
• Score Predictor — regression model predicting final innings score from the first 10 overs (runs, wickets, boundaries, dot balls, run rate). Reports MAE, R², and RMSE at runtime via reports/modeling/score_model3_metrics.json.
• Tournament Simulator — Monte Carlo simulation (1,000 runs) estimating per-team championship and playoff probabilities for an upcoming season.

Advanced metrics (advanced_metrices.py)

• Player impact score (role-aware, eligibility-filtered)
• Death-overs batting/bowling index
• Venue par score and team phase-strength scoring

Frontend

• Streamlit, dark theme, Plotly charts (bar, line, scatter, gauge)
• Custom CSS for metric cards, sidebar, and typography (Bebas Neue / DM Sans)

Tech stack

Python · Streamlit · pandas · NumPy · scikit-learn · Plotly · joblib

Project structure

ipl_analysis/
├── app.py                      # Streamlit dashboard (7 pages)
├── advanced_metrices.py        # Player/venue/team advanced metrics
├── insights.py
├── model1.py / model2.py / model3.py
├── win_probablity_model.py     # Win probability RandomForest training
├── win_probablity_curve.py
├── tournament_simulator.py     # Monte Carlo tournament simulation
├── warehouse_and_reports.py
├── parsers/                    # Cricsheet JSON → structured CSV
├── json_files/                 # Raw Cricsheet match data
└── requirements.txt

Running it locally

git clone https://github.com/arka562/ipl_analysis.git
cd ipl_analysis
pip install -r requirements.txt
streamlit run app.py

Note: the dashboard expects pre-built CSVs/models under reports/ and models/. If these aren't present, run the relevant pipeline script first (e.g. python win_probablity_model.py, python tournament_simulator.py) to generate them — the Tournament Simulator page tells you explicitly if its output file is missing.

Data source

Cricsheet — ball-by-ball IPL data, 2008 to present.

Status

Actively developed as a personal analytics project. Not production-hardened — no tests, no CI, and directory paths currently have fallback logic to handle inconsistent local vs. deployed folder structures (see BASE resolution in app.py). Treat model metrics as a snapshot of the latest training run, not guaranteed constants.