KB 856266 - What is Model Development Life Cycle (MDLC)

KB 856266 - What is Model Development Life Cycle (MDLC)

Summary

The Model Development Life Cycle (MDLC) provides a structured path for turning a business problem into a reliable, monitored machine-learning solution. It helps teams keep projects transparent, reproducible, and compliant while ensuring that models remain useful long after deployment.

1. Problem Definition & Scoping

This is where the conversation begins:

  1. What decision are we trying to improve?
  2. What data supports that decision?
  3. What risks or constraints matter?
  4. How will we know the model is “good enough”?
Typical outputs from this phase includes: Business use cases description, success metrics, initial risk  assessment, feasibility notes.

2. Data Acquisition & Understanding

Before algorithms enter the picture, teams explore the landscape:\

  1. Identify internal/external data sources
  2. Assess data quality, availability, and lineage
  3. Perform exploratory data analysis (EDA)
  4. Determine ethical and compliance implications
Outputs from this phase will be: Data dictionary, EDA summary, data quality report.

3. Data Preparation

Most of the real work happens here. This step involves:

  1. Cleaning and standardizing raw data
  2. Feature engineering and selection
  3. Handling missing values, outliers, and data imbalance
  4. Creating training/validation/test splits ensuring no leakage
Outputs from this phase will be: Cleaned datasets, Transformation pipelines, Feature documentation

4. Model Development

Modelers choose techniques suited to the problem and constraints:

  1. Trying multiple candidate algorithms
  2. Hyperparameter tuning
  3. Cross-validation
  4. Evaluating fairness, robustness, and explainability
Outputs from this phase will be: baseline model, improved versions, comparative evaluation results.

5. Model Validation & Governance

Independent validation ensures the model is trustworthy:
  1. Review for accuracy, stability, bias, and interpretability
  2. Stress testing under edge conditions
  3. Compliance review and documentation
  4. Sign-offs from model risk and governance teams
Outputs from this phase will be:  validation reports, approval checklist, model risk rating.

6. Deployment & Integration

The model moves from experimenting to production:
  1.  Containerization or packaging
  2. API/service development
  3. Integration with bussiness systems
  4. Deployment approval and operational readiness checks
Outputs from this phase will be: Production ready model, deployment plan, rollback,procedures

7. Monitoring & Performance Management

A deployed model is never "finished". Teams watch for:
  1. Data drift and concept drift
  2. Model accuracy and stability degredation
  3. System performance issues
  4. Compliance and fairness  deviations
  5.  User feedback and operational anomalies
Outputs from this phase will be: Monitoring Dashboards, Alerts, Drift reports

8. Continuous Improvement & Retirement

Eventually a model changes or steps aside:
  1. Periodic retraining
  2. Revalidation after major updates
  3. End-of-life planning when the model is outdated or replaced
Outputs from this phase will be: retraining logs, model versioning records, retirement documentation.