Project Title	CreaseIQ: IPL Match Intelligence Platform
Submitted by	Jyotshna Payasi
Registration number	26BCE10651
Faculty	Dr. Pradeep Kumar Mishra
Submission date	30 September 2026

---

🏏 CreaseIQ

IPL Match Intelligence Platform

Clean data · honest analytics · calibrated win probabilities

_{Build Your Own Project (VITyarthi) · Python Essentials · Jyotshna Payasi (26BCE10651) · VIT Bhopal}

Overview

CreaseIQ turns a raw IPL match file (1,243 matches, 2008–2026, Cricsheet-derived) into:

1. A clean, canonical, validated dataset.
   • 19 team strings map to 15 franchises, and 60 venue strings to 37 venues.
   • Super-over runs are removed from tied innings.
   • D/L results and a voided match are flagged.
   • Every change is counted in docs/data_quality_report.md.
2. Honest analytics.
   • Does the toss matter? A causal binomial test with confidence interval (CI) and minimum detectable effect.
   • Is chasing easier? How much did the Impact Player rule change scoring?
3. A win-probability model built the right way.
   • Every feature is computed as of the day before the match, and five automated leakage tests check this.
   • Orientation is symmetric, so which team you list first does not matter.
   • Walk-forward validation and calibration.
   • Baselines and bootstrap CIs, with a holdout touched only after the model selection was frozen and hashed.

The headline finding is that pre-match IPL outcomes are close to a coin flip. CreaseIQ measures that honestly instead of claiming 90% accuracy.

Results (generated from reports/metrics.json)

Analytics.

• Toss winners won 51.6% (95% CI 48.8%–54.4%, p = 0.29). There is no detectable toss advantage.
• The chasing side won 54.7% (p = 0.001).
• First-innings scores rose by +27.1 runs in the Impact Player era (Cohen's d = 0.82).

Win-probability models. Walk-forward validation used development seasons ≤ 2024; the holdout, 2025, 2026, was evaluated only after each tier's selection was frozen and hashed (one selection per tier; every recomputation is logged in reports/holdout_ledger.json). A coin flip scores log-loss 0.6931.

Tier	Model	Walk-forward log-loss	Holdout log-loss [95% CI]	Holdout accuracy	Holdout AUC
post_toss	logreg	0.6857	0.6974 [0.6818, 0.7136]	47.9%	0.489
pre_toss	elo_logit	0.6900	0.7060 [0.6889, 0.7234]	43.0%	0.453

Honest verdict. The models edge a coin flip in walk-forward validation, but none of them beats it on the 2025–26 holdout: relationships learned before 2023 weakened. See docs/model_card.md.

Performance. The full pipeline takes 6.6 s. A warm prediction takes 9.3 ms, and the slowest dashboard page renders in 0.57 s. Peak memory is 40 MB.

Features

Module	What it does	Requirements
M1 Data engineering	Schema and row-rule validation with quarantine; fail-loud canonicalisation; tie, D/L and voided-match handling; quality report and data dictionary; normalised SQLite; validated CSV append	FR-01…05
M2 Analytics & visualisation	Team records with Wilson CIs, titles, head-to-head, Elo history; causal toss test; chasing and era tests; shrunk venue profiles and k-means clusters; scoring trends; POTM leaderboards; filters and safe CSV export	FR-06…11
M3 Prediction engine	As-of FeatureBuilder; margin-aware Elo; 5 model families plus 5 baselines; walk-forward CV; one-SE selection; time-ordered calibration; bootstrap and paired tests; hash-verified registry; explanations; score regressor	FR-12…18
M4 Scenarios	What-if tornado (toss, venue, opponent); Monte Carlo season simulation	FR-19, FR-20
M5 Ops & reporting	key=value logs with run ids; prediction log; PSI drift; model card, figures and NFR evidence	FR-21, FR-22

Dashboard pages: Home · Data Explorer · Team Analytics · Venue & Toss · Predict Match · Model Lab · What-If.

Technologies

• Language: Python 3.12/3.13.
• Data and statistics: pandas 3, NumPy, SciPy, statsmodels.
• Machine learning: scikit-learn (logistic regression, random forest, histogram gradient boosting, k-means).
• Validation: pandera.
• Storage: SQLAlchemy 2 with SQLite (Postgres via URL).
• Interfaces: Streamlit and Plotly (dashboard), Typer and Rich (CLI).
• Figures: matplotlib.
• Quality: pytest, Hypothesis, ruff, mypy, import-linter, bandit, pip-audit, GitHub Actions.
• Documentation: Mermaid (diagrams), Playwright/Chromium (screenshots and the PDF report).

Install & run

Needs Python ≥ 3.12. Node ≥ 22.13 is optional and only needed for re-rendering diagrams and the PDF.

Windows (PowerShell)

py -3.12 -m venv .venv

.venv\Scripts\activate

macOS / Linux

python3.12 -m venv .venv

source .venv/bin/activate

Install (both platforms)

pip install -r requirements.txt -r requirements-dev.txt

pip install -e . --no-deps

Run the whole pipeline (validate → build-db → analyze → train):

creaseiq all

Start the dashboard:

streamlit run src/creaseiq/app/Home.py

Useful CLI commands (creaseiq --help lists them all):

creaseiq predict mi csk --venue wankhede

creaseiq predict mi csk --venue wankhede --toss-winner csk --toss-decision field

creaseiq whatif rcb gt --venue narendra_modi

creaseiq ingest --append new_matches.csv --dry-run

creaseiq benchmark

Testing

pytest

pytest --cov

ruff check . && mypy src && lint-imports

The test suite contains several hundred tests (the exact count is in report §11.1):

• Unit: including Hypothesis property tests.
• Data contract.
• Leakage: future deletion, future mutation, same-day isolation, label shuffle and the column allow-list, plus an orientation-signal check.
• Reproducibility.
• Security: secrets scan, SQL-injection strings, tampered artifacts, CSV injection, upload guards.
• Streamlit: AppTest for every page.
• CLI: CliRunner.
• Performance.

Coverage and NFR evidence are in docs/nfr_verification.md.

Screenshots

Home	Venue & Toss
Model Lab	What-If

Architecture

Dependencies point downward only: app/cli → services → domain → data → core. This is enforced by import-linter and an AST test. All diagrams (use case, workflow, sequence, class, component, ER, deployment) are in docs/diagrams/. The rationale is in docs/architecture.md and docs/decisions/.

Project structure

configs/        config.yaml · team_lineage.yaml · venue_canonical.yaml · home_grounds.yaml
data/           raw/ (immutable CSV + SHA-256) · external/ (verified facts) · processed/ (generated)
src/creaseiq/   data · db · analytics · features · models · simulation · services · viz · reporting · app · cli
tests/          unit · integration · validation (leakage, security, reproducibility) · fixtures
docs/           research notes · plan review · ADRs · diagrams · model card · data dictionary · viva prep
reports/        metrics.json · analytics.json · perf.json · coverage.json · figures/
models/         registry.json + hash-verified artifacts
report/         PDF report builder and CreaseIQ_Project_Report.pdf

Disclaimer

CreaseIQ is an educational analytics project. Its statistics and probabilities are retrospective, uncertain and close to a coin flip.

• They are not betting, fantasy-sports or financial advice. Do not use them to stake money on any match. In India, the Promotion and Regulation of Online Gaming Act, 2025 prohibits online money games.
• CreaseIQ is not affiliated with the BCCI, the IPL, any franchise or Cricsheet.

Data attribution & license

• Match data: Cricsheet, compiled by Stephen Rushe and made available under the Open Data Commons Attribution License v1.0. It was probably obtained through a Kaggle re-packaging ("IPL Complete Cricket Dataset (2008–2026)"). CreaseIQ transformed the data; any errors are ours.
• Code: MIT. The licence does not cover data/raw/.

AI assistance & sources

• Tools: this project was built with AI coding assistance (Claude Code) for scaffolding, research and review.
• Oversight: the author reviewed the design and code, and can explain every part (see docs/viva_prep.md).
• Sources: every external source is cited in docs/research_notes.md.

Author

Jyotshna Payasi · Reg. No. 26BCE10651 · Vellore Institute of Technology, Bhopal