Reviewing every contract with equal depth wastes limited legal capacity. Analysis of 510 real commercial contracts shows risk is highly concentrated: the top 15 contracts (3% of the portfolio) carry an average risk score 3.8x higher than the rest of the sample, and just the top 10% of contracts by score account for 29.3% of total portfolio risk. Two clauses - Uncapped Liability and Non-Compete - are reliably elevated in high-risk contracts despite being relatively rare overall (present in ~22–23% of contracts, but in 80–100% of the highest-risk group), making them useful early triage flags.
Recommendation: prioritize legal review using the risk score, and treat the presence of Uncapped Liability or Non-Compete as an early flag worth a closer look — even before a full score is calculated. This lets a legal/compliance team focus limited review time where risk actually concentrates, instead of spreading it evenly across the portfolio.
Companies signing dozens or hundreds of commercial contracts cannot give each one an equally in-depth legal review. As a result, contracts with dangerous terms (uncapped liability, exclusivity, volume restrictions, etc.) sometimes get the same attention as routine NDAs, and legal risk quietly accumulates. This project builds a lightweight, transparent prioritization model - a risk score - that a legal or compliance team can use to identify which contracts need priority review.
1. Risk is concentrated in a small group of contracts, not evenly distributed. The average risk score across the full sample was 14.1 (max: 57, for a Content License, Marketing and Sales Agreement). The top 15 highest-risk contracts (under 3% of the sample) averaged 49.1 — 3.8x higher than the remaining 495 contracts (13.0). Only 43.7% of contracts scored above the sample average, meaning the average is pulled up by a small number of outliers rather than reflecting a typical contract. Looking at the full distribution rather than just the top 15: the top 10% of contracts by score account for 29.3% of total portfolio risk. Implication: reviewing all 510 contracts equally is inefficient — prioritizing by score concentrates review time where risk actually lives.
2. Uncapped Liability and Non-Compete are reliably elevated in high-risk contracts - not the most common clauses. Uncapped Liability appears in only 21.8% of contracts overall, but in 80.0% of the top-15 highest-risk contracts (3.7x more often). Non-Compete shows an even stronger pattern: 23.3% overall vs. 100.0% in the top 15 (4.3x more often). Notably, neither ranks in the top 10 most frequent clauses overall - unlike Anti-Assignment (73.3%) or License Grant (50.0%), which are common but not discriminative, appearing in high-risk and routine contracts alike. I checked this pattern isn't an artifact of the specific top-15 cutoff by re-running it at top-25 and top-30: Non-Compete stays consistently elevated (4.2–4.3x) across all three window sizes, while Uncapped Liability is more moderate but still consistently 3x+ above baseline. Implication: the presence of Uncapped Liability or Non-Compete can be used as a fast triage flag, even before a full risk score is calculated.
3. A complementary check - ranking clauses by lift rather than raw frequency gap - surfaces additional candidates for early red flags. Lift measures how many times more often a clause appears in the high-risk group relative to its overall base rate, which rewards rare-but-concentrated clauses more than the frequency-gap view above. By this measure, the strongest signals are Third Party Beneficiary (6.5% overall → 40.0% in the top 15, 6.2x), Source Code Escrow (2.5% → 13.3%, 5.3x), and Liquidated Damages (12.0% → 60.0%, 5.0x). These clauses are rarer than Uncapped Liability or Non-Compete, so they contribute less to any single contract's total score, but their presence is disproportionately concentrated in the highest-risk contracts - worth treating as secondary flags alongside the two primary indicators above.
4. Grouping contracts into score-based tiers gives an operational review framework. Splitting the sample into quartiles by risk score produces four tiers: Low (135 contracts, avg. score −1.1), Medium (130, avg. 7.5), High (119, avg. 17.7), and Critical (126, avg. 33.7). A legal/compliance team could route Critical-tier contracts for review first, Low-tier through standard/fast-track processing, without relying on a single arbitrary top-N cutoff.
Dataset: CUAD (Contract Understanding Atticus Dataset) - 510 commercial contracts from SEC EDGAR filings, annotated by practicing attorneys across 40 labeled clause categories. Licensed under CC BY 4.0.
Approach:
- Flag the presence/absence of each of the 40 clause categories per contract.
- Calculate each clause's frequency across the full sample.
- Assign each clause category a risk weight from 1-10, based on its potential legal/financial consequence for the signing party (e.g. Uncapped Liability: 10; Governing Law: default weight). A score of 0 represents a contract with no flagged clauses at all; negative scores indicate contracts where protective clauses outweigh risky ones - this is intentional, not an error.
- Compute each contract's risk score as the sum of weights of the risky clauses present.
- Visualize the distribution and flag the highest-risk contracts; cross-check the clause-level findings across multiple top-N window sizes and against an alternative (lift-based) ranking.
Limitations (by design, at this stage):
- Contracts are drawn from public SEC filings of publicly traded US companies — the sample is not representative of private or non-US businesses.
- Risk weights are expert-assigned based on legal/compliance judgment, not statistically derived - this is an expert-based model, not machine learning, by deliberate choice at this stage.
- 81 contracts (15.9% of the sample) have partially redacted text for individual clauses.
- The top-15 cutoff used for the headline finding is a small group (3% of the sample), so single-clause percentages there move in ~6.7-point increments; findings were checked against top-25 and top-30 to confirm they aren't an artifact of that specific cutoff.
- Several contracts are tied on risk score right at the top-15 boundary; the script breaks ties deterministically (by contract name) so results are reproducible across runs and environments.
- Risk scores can be negative by design - protective clauses (e.g. liability caps) subtract from score, so a negative total reflects a net-protective contract, not an error.
Tools: Python, Pandas, NumPy, Matplotlib/Seaborn, Streamlit.
- Replace expert-assigned weights with weights statistically derived from historical contract dispute/loss data, if available.
- Add contract type classification (NDA, license, M&A, etc.) and compare risk profiles across types.
- Validate weights with input from 2-5 practicing lawyers to reduce model subjectivity.
- Add an NLP module to surface risk-relevant language not yet captured by the labeled clause categories.
git clone https://github.com/JuliaZimenina/Contract_Risk_Scoring.git
cd Contract_Risk_Scoring
pip install -r requirements.txt
# Download CUAD_v1_master_clauses.csv from:
# https://www.kaggle.com/datasets/theatticusproject/atticus-open-contract-dataset-aok-beta
# and place it in data/raw/
python risk_analysis.pyJulia Zimenina - Data Analyst with 10+ years of experience in legal and compliance (banking, energy, healthcare, manufacturing, agriculture), now applying that domain expertise and analytical mindset to data. This project reflects how I approach analysis: start from a real business problem, stay honest about a model's limitations, check whether findings hold up under different assumptions, and translate results into a recommendation someone can act on.