Mixed & Real-World Problems
Challenge yourself with these 25 integrated problems that combine multiple concepts. These prepare you for complex exam questions and real-world ML projects.
Real-World Scenarios (10 Problems)
Problem 1.1: Customer Churn Prediction
Difficulty: Hard | Time: 30 min | Topics: Classification, Metrics, Cost-Benefit
Business Context:
Telecom company wants to prevent customer churn
- 100,000 customers
- 5% churn rate (5,000 churners)
- Intervention cost: $50 per customer
- Revenue from retained customer: $5,000
- Customer lifetime value if saved: $5,000 (net benefit)
Problem Statement:
You train two models on historical data:
Model A (Simple):
Precision: 0.60
Recall: 0.80
Cost to deploy: $10,000
Model B (Complex):
Precision: 0.75
Recall: 0.85
Cost to deploy: $50,000
Questions:
a) For Model A: How many interventions? True positives? False positives? False negatives?
b) For Model B: How many interventions? True positives? False positives? False negatives?
c) Calculate ROI for each model: (Benefit - Cost) / Cost
d) Which model should you deploy?
e) What other factors would you consider?
Problem 1.2: Medical Diagnosis System
Difficulty: Hard | Time: 25 min | Topics: Imbalanced Data, Metrics, Threshold Tuning
Business Context:
Rare disease detection system (0.5% prevalence)
- 10,000 patients screened
- 50 have disease, 9,950 don't
- False negative: $100,000 (miss disease, patient dies)
- False positive: $1,000 (unnecessary treatment)
- True positive: -$100,000 (save life)
- True negative: $0
Problem Statement:
Model produces probabilities. Compare 3 thresholds:
Threshold 0.5:
TP: 40, FN: 10, FP: 200, TN: 9,750
Threshold 0.3 (lower threshold):
TP: 45, FN: 5, FP: 500, TN: 9,450
Threshold 0.7 (higher threshold):
TP: 35, FN: 15, FP: 50, TN: 9,900
Questions:
a) Calculate total cost for each threshold
b) Which threshold minimizes cost?
c) Calculate sensitivity (recall) for each
d) Calculate specificity for each
e) How does disease prevalence affect threshold choice?
Problem 1.3: Fraud Detection at Scale
Difficulty: Hard | Time: 25 min | Topics: Large-Scale Classification, Metrics
Business Context:
Credit card fraud detection:
- Process 1,000,000 transactions/day
- Fraud rate: 0.1% (1,000 frauds)
- False positive cost: $20 (investigation)
- False negative cost: $5,000 (fraud loss)
- Detection latency constraint: <50ms
Problem Statement:
Three model options:
Model A: Accuracy 99.8% (catches mostly non-frauds)
Precision: 0.50, Recall: 0.60, Inference: 5ms
Model B: Accuracy 99.5% (balanced)
Precision: 0.75, Recall: 0.80, Inference: 50ms
Model C: Accuracy 98.0% (high recall)
Precision: 0.40, Recall: 0.95, Inference: 200ms
Questions:
a) Calculate daily costs for each model (1M transactions)
b) Which model respects the latency constraint?
c) Best model considering latency + cost?
d) What if latency constraint relaxed to 100ms?
e) How does model choice change at scale?
Problem 1.4: Recommendation System Evaluation
Difficulty: Hard | Time: 25 min | Topics: Multi-Class Metrics, Ranking
Business Context:
E-commerce recommendation system
- Users rate products (1-5 stars)
- Goal: Recommend products user would rate ≥4 stars
- Two types of errors:
* Recommend bad product (user rates ≤3): Wastes attention
* Don't recommend good product: Misses sale opportunity
Reclassify as binary:
- Class 1: Would rate ≥4 (good recommendation)
- Class 0: Would rate ≤3 (bad recommendation)
Problem Statement:
Tested on 1,000 products:
Model A (Conservative):
TP: 300, FN: 100, FP: 50, TN: 550
Model B (Aggressive):
TP: 350, FN: 50, FP: 150, TN: 450
Questions:
a) Calculate precision and recall for each
b) Which shows more recommendations?
c) For business: missing good product = 2× worse than showing bad product
d) Which model maximizes user satisfaction?
e) How would you evaluate rankings (not just classification)?
Problem 1.5: Email Spam Classification
Difficulty: Medium | Time: 20 min | Topics: Classification, Threshold, Precision-Recall
Business Context:
Gmail-like spam filter:
- 100 million emails/day
- Spam rate: 30% (30M spam, 70M legitimate)
- False positive (mark legitimate as spam): User loses important email
- False negative (let spam through): Annoyance
Metric: Users prefer false negatives to false positives
Problem Statement:
Model with different thresholds:
Conservative (threshold 0.9):
Precision: 0.95, Recall: 0.50
Moderate (threshold 0.5):
Precision: 0.85, Recall: 0.80
Aggressive (threshold 0.1):
Precision: 0.70, Recall: 0.95
Questions:
a) Which threshold maximizes precision? Recall?
b) Which would miss the most spam?
c) Which would incorrectly flag the most legitimate?
d) Which would you choose? Why?
e) Could you use precision-recall curve?
Problem 1.6: Loan Approval System
Difficulty: Hard | Time: 25 min | Topics: Fairness, Metrics, Multiple Objectives
Business Context:
Bank loan approval model
- Model optimizes for profit (minimize defaults)
- But must also be fair across demographic groups
Prediction task: Will applicant default?
- Class 0 (no default): Approve loan
- Class 1 (default): Reject loan
Hidden issue: Historical data shows racial bias
- Model achieves high accuracy but discriminates
Problem Statement:
Model results by demographic group:
Group A (protected group):
Approval rate: 60%, Default rate: 10%
Group B (majority):
Approval rate: 80%, Default rate: 8%
Model accuracy: 92% overall
But different false negative rate by group!
Questions:
a) Is the model "fair"? Why/why not?
b) Which group faces discrimination?
c) What metric captures fairness?
d) Could you achieve both high accuracy AND fairness?
e) What's your recommendation?
Problem 1.7: Manufacturing Quality Control
Difficulty: Medium | Time: 20 min | Topics: Classification, Cost-Benefit, Metrics
Business Context:
Automated defect detection on assembly line
- Production: 100,000 items/day
- Defect rate: 2% (2,000 defects)
- False positive (scrap good item): $10 cost
- False negative (ship defect): $500 liability cost
- Inspection staff: $50/unit inspected
Problem Statement:
Model: Automatically flags suspected defects
Two strategies:
Strategy A: Model flags, then human inspects flagged items
Precision: 0.80 → 20% flagged are actually good
Recall: 0.90 → 10% of defects slip through
Expected: 12,500 items flagged/day
Strategy B: Model flags with higher threshold
Precision: 0.95 → 5% flagged are good
Recall: 0.70 → 30% of defects slip through
Expected: 5,000 items flagged/day
Questions:
a) Calculate daily cost for Strategy A (inspection + scrap + liability)
b) Calculate daily cost for Strategy B
c) Which strategy minimizes cost?
d) What threshold trade-off would you accept?
e) Could you improve without changing model?
Problem 1.8: Student Dropout Prediction
Difficulty: Hard | Time: 25 min | Topics: Early Prediction, Multi-Class, Intervention
Business Context:
University early warning system:
- Predict student will dropout BEFORE it happens
- Early intervention (tutoring, counseling): $500/student
- Cost of dropout (lost tuition, resources): $50,000
- True positive: Catch at-risk student early
- False positive: Spend $500 on student who would succeed anyway
- False negative: Miss opportunity to save $50,000
Problem Statement:
Baseline: No intervention, 5% dropout rate
Model predicts dropout probability:
With intervention on high-risk students:
50 identified as high-risk
- 40 actually would drop out (saved 40 × $50k)
- 10 would succeed anyway (wasted $5k)
Without intervention, among those 50:
- 40 would drop out (lost opportunity)
- 10 would succeed
Questions:
a) Calculate savings from intervention
b) What's the net benefit?
c) How many students should you intervene with?
d) Is perfect prediction necessary?
e) What's the break-even accuracy?
Problem 1.9: Click-Through Rate (CTR) Prediction
Difficulty: Medium | Time: 20 min | Topics: Regression, Classification, Ranking
Business Context:
Online advertising platform:
- 100,000 ad impressions/day
- Average CTR: 2% (2,000 clicks)
- Revenue per click: $5
- Model predicts: P(click | features)
Two strategies:
- Show ads with highest predicted CTR
- Show diverse ads (some uncertainty)
Problem Statement:
Model predictions on 100,000 impressions:
Strategy A (Highest predicted CTR):
Average predicted CTR: 0.040 (4%)
Average actual CTR: 0.038 (3.8%)
Calibration error: Small
Strategy B (Diverse):
Average predicted CTR: 0.022 (2.2%)
Average actual CTR: 0.025 (2.5%)
Calibration error: Moderate
Questions:
a) Which strategy generates more revenue?
b) Why might Strategy B be better long-term?
c) How do you balance exploration vs exploitation?
d) What's the right metric for this problem?
e) How would you validate model predictions?
Problem 1.10: Multi-Class Text Classification
Difficulty: Hard | Time: 25 min | Topics: Multi-Class Metrics, Class Imbalance
Business Context:
Customer support ticket router:
- Classify support tickets to 5 departments
- Misrouting costs: Delays, customer frustration
- Revenue implications: Lost sales, retention
Ticket types (imbalanced):
- Billing: 50% (high volume, easy to handle)
- Technical: 30% (medium volume, complex)
- Product: 10% (lower volume)
- Returns: 7% (lower volume)
- Other: 3% (rare)
Problem Statement:
Multi-class model results:
Per-class metrics:
Billing: Precision 0.90, Recall 0.95, F1 0.92
Technical: Precision 0.75, Recall 0.70, F1 0.72
Product: Precision 0.60, Recall 0.50, F1 0.55
Returns: Precision 0.40, Recall 0.30, F1 0.34
Other: Precision 0.50, Recall 0.20, F1 0.29
Questions:
a) Calculate macro-averaged F1
b) Calculate weighted F1 (by class frequency)
c) Which classes perform poorly?
d) Why might "Other" be hard to predict?
e) How would you improve rare class performance?
End-to-End Pipelines (8 Problems)
Problem 2.1: Complete ML Workflow - Binary Classification
Difficulty: Hard | Time: 30 min | Topics: All concepts integrated
Scenario:
A bank wants to predict loan defaults to reduce risk
Step 1: Problem Definition
- Target: Will customer default? (Yes/No)
- Dataset: 5,000 past loans with outcomes
- Features: Income, Credit Score, Loan Amount, Age, etc.
- Class distribution: 80% non-default, 20% default (imbalanced)
Step 2: Data Split & Validation
- Choose 80-20 train-test split? Or K-fold CV? Or stratified?
- Explain your choice
Step 3: Model Selection
- Candidate models: Logistic, Tree, RF, SVM, Neural Net
- Dataset size: 5,000
- Explainability needed? Yes (for regulatory)
- Speed needed? Yes (real-time decisions)
Step 4: Hyperparameter Tuning
- Ridge regression with λ selection
- Or: Tree depth, max_features for RF
Step 5: Evaluation
- Which metrics matter: Accuracy? Precision? Recall? F1? ROC?
- Why?
Step 6: Threshold Tuning
- Default cost: $50,000 loss
- False positive cost: $100 investigation
- Optimal threshold?
Step 7: Production Deployment
- Acceptable error rate?
- Monitoring plan?
- Retraining frequency?
Questions:
a) Explain your choice for train-test strategy
b) Which model would you choose? Why?
c) How would you tune hyperparameters?
d) Which metrics would you use? Why?
e) Calculate optimal decision threshold
f) Design monitoring & retraining plan
g) What would you document for stakeholders?
Problem 2.2: Feature Engineering Impact
Difficulty: Hard | Time: 25 min | Topics: Feature Engineering, Model Evaluation
Scenario:
House price prediction with different feature sets
Baseline (raw features):
- Square feet, Bedrooms, Bathrooms, Age
Model A (baseline):
R² = 0.65, RMSE = $50,000
Added features:
- sqrt(Square feet), Price/SqFt, Bedrooms×Bathrooms
- Neighborhood (categorical)
- Distance to transit, School quality
Model B (engineered features):
R² = 0.75, RMSE = $35,000
Questions:
a) How much did R² improve?
b) Is improvement significant? Statistically?
c) Did RMSE improve proportionally?
d) What could explain the improvement?
e) Risk of overfitting? How would you check?
f) Which features might be most important?
g) Would all features help equally?
Problem 2.3: Class Imbalance Handling
Difficulty: Hard | Time: 25 min | Topics: Imbalanced Data, Techniques
Scenario:
Fraud detection (99.9% non-fraud, 0.1% fraud)
Naive approach (standard model):
- Accuracy: 99.9% (predicting everything as non-fraud!)
- Useless for fraud detection
Solutions to explore:
1. Class weights: Weight fraud class higher
2. Resampling: Oversample fraud, undersample non-fraud
3. Different threshold: Lower detection threshold
4. Different metric: Use F1 instead of accuracy
Problem Statement:
Compare approaches on 10,000 test samples:
Naive Model:
Confusion matrix: TP=5, FP=0, TN=9995, FN=0
Accuracy: 99.95%
Precision: undefined (TP/(TP+FP) = 5/5 = 1.0 with no FP)
Recall: 1.0
With class weights:
Confusion matrix: TP=7, FP=50, TN=9945, FN=3
Accuracy: 99.55%
Precision: 0.12
Recall: 0.70
With threshold tuning:
Confusion matrix: TP=8, FP=100, TN=9895, FN=2
Accuracy: 99.03%
Precision: 0.07
Recall: 0.80
Questions:
a) Why is naive model accuracy "perfect" but useless?
b) Explain confusion matrix for each approach
c) Which approach would you use? Why?
d) How do false positives cost you at scale?
e) When is recall more important than precision?
Problem 2.4: Cross-Validation for Model Selection
Difficulty: Medium | Time: 20 min | Topics: CV, Hyperparameter Tuning
Scenario:
Predicting customer lifetime value (continuous target)
Grid search over hyperparameters:
Ridge λ values: [0.001, 0.01, 0.1, 1.0, 10]
Tree depths: [3, 5, 10, 15, 20]
Each combination evaluated with 5-fold CV
Total: 5 × 5 = 25 configurations
Each tested 5 times = 125 model trainings
Problem Statement:
CV Results (5-fold, RMSE metric):
Ridge λ = 0.1: CV RMSE = 2500 ± 100
Ridge λ = 1.0: CV RMSE = 2600 ± 80
Tree depth = 5: CV RMSE = 2400 ± 150
Tree depth = 10: CV RMSE = 2200 ± 400
Tree depth = 20: CV RMSE = 2000 ± 600
Best: Tree depth = 20 (lowest mean RMSE)
But highest std dev (600)!
Questions:
a) Which hyperparameter would you choose?
b) Why might depth=20 have high std dev?
c) Is depth=20 stable? Reproducible?
d) What's the trade-off?
e) Would you choose based on mean or std dev?
f) How would you break the tie?
Problem 2.5: Train-Test Leakage Detection
Difficulty: Hard | Time: 20 min | Topics: Data Leakage, Validation
Scenario:
Predicting if patient will get disease
WARNING SIGNS OF LEAKAGE:
Model A:
Training accuracy: 95%
Test accuracy: 91%
Gap: 4% (reasonable)
Model B:
Training accuracy: 92%
Test accuracy: 88%
Gap: 4% (reasonable)
Model C:
Training accuracy: 99%
Test accuracy: 52%
Gap: 47% (HUGE! Leakage suspected!)
Problem Statement:
Investigating Model C:
Features used:
- Age, Weight, Blood pressure
- Insulin level (MEASURED AT HOSPITAL VISIT)
- Prior medications
- Disease test result (MEASURED AT HOSPITAL VISIT)
Problem: Insulin & test results are MEASURED when
patient visits hospital, not predictions!
This is LEAKAGE if disease correlates with test visit.
Questions:
a) Identify the leakage source
b) Why doesn't it show in training accuracy?
c) Why does it crash in test accuracy?
d) How would you fix it?
e) What features should you actually use?
f) How would you catch leakage early?
Problem 2.6: Model Serving & Latency
Difficulty: Medium | Time: 15 min | Topics: Production, Trade-offs
Scenario:
Real-time recommendation system:
- User visits site, must get recommendation in <100ms
- Page loads: 50ms
- Available for ML: <50ms
Model options:
Model A (Simple):
Inference time: 5ms
Accuracy: 85%
Model B (Medium):
Inference time: 40ms
Accuracy: 90%
Model C (Complex):
Inference time: 150ms
Accuracy: 92%
Questions:
a) Which model meets latency constraint?
b) What's the accuracy-latency tradeoff?
c) Could you use Model C with caching?
d) What preprocessing could speed up inference?
e) Would batch processing help?
f) Is ensemble possible within latency?
Problem 2.7: A/B Testing New Model
Difficulty: Hard | Time: 25 min | Topics: Deployment, Experimentation
Scenario:
Current production model (Model A) vs. new model (Model B)
Current Model A:
- Conversion rate: 5.0%
- Deployed 1 year ago
- Known issues: Doesn't handle new products well
Proposed Model B:
- Testing shows: Conversion rate 5.3%
- More complex, riskier
- Better on new products
A/B Test Setup:
- 100,000 users
- 50% see Model A (control), 50% see Model B (treatment)
- Run for 2 weeks
- Goal: Detect 0.3% lift with 95% confidence
Questions:
a) What metrics would you track?
b) How would you detect if Model B breaks something?
c) What's the expected lift in revenue?
d) What would cause you to reject Model B?
e) How long should you run the test?
f) What if results are inconclusive?
Problem 2.8: Model Debugging Workflow
Difficulty: Hard | Time: 25 min | Topics: Troubleshooting, Diagnosis
Scenario:
Production model degradation:
Week 1: Model accuracy 90%, production deployed
Week 2: Accuracy drops to 85% (on same data!)
Week 3: Accuracy drops to 80%
Investigating:
Possibility A: Data drift (data distribution changed)
→ Check: Training data vs current data distribution
Possibility B: Label noise (ground truth changed)
→ Check: Are labels being recorded correctly?
Possibility C: Feature drift (features computed differently)
→ Check: Did upstream data pipeline change?
Possibility D: Bug in evaluation (measurement error)
→ Check: Is evaluation code correct?
Problem Statement:
Debug timeline:
Day 1: Accuracy 90%
Day 2: Accuracy 88%
Day 3: Accuracy 85%
Day 4: Accuracy 82%
Correlation check:
- No deployment changes (Possibility D ruled out)
- Feature computation same (Possibility C ruled out)
- Labels still being recorded (Possibility B seems OK)
- Training data: Income distribution shifted
* Previously: 40% < $50k, 60% > $50k
* Current: 60% < $50k, 40% > $50k
Questions:
a) What's the most likely cause?
b) How would you confirm data drift?
c) Should you retrain? When?
d) How do you prevent this?
e) What monitoring would help?
f) Is this a model problem or data problem?
Critical Thinking Problems (7 Problems)
Problem 3.1: The Accuracy Trap
Difficulty: Hard | Time: 20 min | Topics: Metrics, Trade-offs
Scenario:
Rare disease classifier:
- 99% of people don't have disease
- Model predicts everyone as "no disease"
- Accuracy: 99%
- But catches 0 actual cases (useless!)
Questions:
a) Why is 99% accuracy misleading?
b) Which metric would you use instead?
c) What if you weighted precision vs recall?
d) How would you explain this to non-technical stakeholder?
e) Design a better evaluation framework
Problem 3.2: The Interpretability-Accuracy Trade-Off
Difficulty: Hard | Time: 20 min | Topics: Model choice, Stakeholders
Scenario:
Bank loan approval model:
Option A (Linear):
- Accuracy: 85%
- Interpretable: YES
- Stakeholder feedback: "We can explain decisions"
Option B (Neural Net):
- Accuracy: 92%
- Interpretable: NO
- Stakeholder concern: "Why was John's loan rejected?"
- Legal requirement: Must explain decisions (Fair Lending Act)
Questions:
a) Which would you recommend?
b) Can you have both? How?
c) What's the true cost of accuracy gain?
d) Could you use both models (ensemble)?
e) Design a solution that satisfies both needs
Problem 3.3: The Bias-Fairness Trade-Off
Difficulty: Hard | Time: 25 min | Topics: Ethics, Metrics
Scenario:
Credit risk model trained on historical data:
Historical data shows: Past loans from Group A defaulted more
(But: This could be due to discrimination in past lending!)
Model learns:
- Default probability higher for Group A
- Denies more loans to Group A
- Perpetuates discrimination
Dilemma:
- Optimizing for accuracy → Model discriminates
- Optimizing for fairness → Lower accuracy, may reject worthy applicants
Questions:
a) Is high accuracy ethically acceptable?
b) How do you measure fairness?
c) What's the business impact of fairness constraint?
d) Can you optimize for both?
e) Who decides the fairness-accuracy tradeoff?
f) How would you present this to executives?
Problem 3.4: The Overfitting vs Real-World Gap
Difficulty: Hard | Time: 20 min | Topics: Validation, Deployment
Scenario:
Kaggle competition model:
Model achieves:
- 94% accuracy on competition test set
- Wins competition
- $50,000 prize!
But when deployed to real customers:
- Real-world accuracy: 72%
- Business impact: Terrible predictions
- Customer satisfaction drops
Why the gap?
- Competition test set ≠ real distribution
- Overfitting to competition quirks
- Real data has different patterns
Questions:
a) Why did test accuracy not predict real-world performance?
b) What would you have done differently?
c) How do you validate for real-world performance?
d) When is Kaggle not representative?
e) Design a better validation strategy
Problem 3.5: The Cost of False Positives vs False Negatives
Difficulty: Hard | Time: 20 min | Topics: Cost-benefit, Decision Making
Scenario:
Medical screening test:
False Negative: Patient has disease but test says "no"
- Risk: Disease progresses undetected, dies
- Cost to patient: LIFE
False Positive: Patient doesn't have disease but test says "yes"
- Risk: Unnecessary treatment, anxiety, cost
- Cost to patient: $10,000 + stress
Which error is worse?
Questions:
a) How do you quantify these costs?
b) Can you price a human life?
c) Who should decide the threshold?
d) Is 95% sensitivity (5% false negative rate) acceptable?
e) How would you communicate risk to patients?
f) Design a threshold that balances risks
Problem 3.6: The Automation Bias Trap
Difficulty: Medium | Time: 15 min | Topics: Ethics, Human-in-loop
Scenario:
Loan approval automated by ML model:
- Model rejects John's loan application
- John is human, appeals to human manager
- Manager says: "The model rejected it, sorry"
- No human review, no appeal process
Risks:
- Errors from model go unchallenged
- People trust model too much ("computer said no")
- Removes human judgment and empathy
- Legal liability if discrimination occurs
Questions:
a) What's wrong with full automation?
b) When should humans override models?
c) Design a better human-in-loop system
d) When is humans-in-loop too expensive?
e) How do you balance efficiency and fairness?
Problem 3.7: The Long-Term Impact of Model Choices
Difficulty: Hard | Time: 20 min | Topics: Strategy, Feedback loops
Scenario:
Hiring algorithm learns from historical data:
- Past hires from College A were more successful
- Model learns to prefer College A graduates
- Starts hiring more from College A
- Other colleges fall behind
- By year 5: College A has monopoly on hiring
- Diversity drops, innovation suffers
Was the model wrong?
- It predicted success correctly
- But created self-fulfilling prophecy through feedback loop
Questions:
a) Is the model's prediction inaccurate?
b) What went wrong?
c) How could you have prevented this?
d) When do models create feedback loops?
e) How do you detect when feedback loops are harmful?
f) Design monitoring for long-term impact
Final Exam-Style Integration Problems
Practice Test 1: Complete End-to-End (Integrated)
Difficulty: Hard | Time: 60 min | Topics: Everything
Problem: Design complete ML solution for real-world problem:
- Define the problem clearly
- Choose data collection strategy
- Design train-test split & validation
- Select model & tune hyperparameters
- Evaluate with appropriate metrics
- Tune decision threshold
- Plan for production deployment
- Design monitoring & retraining
- Address ethical considerations
Sample problem: Predicting which customers will churn to maximize retention ROI
Practice Test 2: Debugging Challenge (Integration)
Difficulty: Hard | Time: 45 min | Topics: Troubleshooting, Diagnosis
Problem: Model's production accuracy dropped. Debug:
- Identify possible causes (leakage, drift, bugs)
- Design experiments to test hypotheses
- Recommend solutions
- Create monitoring system to prevent recurrence
Study Path for Mixed Problems
Progression:
- Solve real-world problems 1-5 (get comfortable with integration)
- Solve end-to-end pipelines 1-3 (practice complete workflows)
- Attempt critical thinking 1-3 (develop judgment)
- Try practice tests 1-2 (full exam simulation)
Difficulty ladder:
- Easy → Medium problems → Hard → Real-world scenarios → Integration → Critical thinking
Time commitment:
- 4-5 hours for thorough understanding
- 2-3 hours for speed practice
Next Steps
Completed all 95 problems?
✓ You've covered all CT exam topics comprehensively
✓ Ready for exam with confidence
✓ Next: Timed practice test simulating actual exam conditions
Need solutions? See Solutions Guide
Questions or stuck? Review related Concept Pages
Good luck! You've got this! 🚀