This page is part of the ForgeSDLC knowledge base — an AI-assisted, human-directed methodology for taking product work from concept to production. For the core operating model and vocabulary, see Forge SDLC overview and What is ForgeSDLC?.
CRISP-DM: Cross-Industry Standard Process for Data Mining
Purpose: Project-agnostic reference for CRISP-DM — the most widely adopted methodology for data mining and applied machine learning. Use it to structure discovery, modeling, and handoff without tying the blueprint to a specific tool stack.
Audience: Teams following DATA-SCIENCE.md and approaches/README.md.
Overview
CRISP-DM (Cross-Industry Standard Process for Data Mining) is a process model, not a single algorithm or product. It organizes work from business intent through deployed models with explicit feedback loops. Its strength is shared vocabulary across business, analytics, and engineering: everyone can anchor discussions in the same six phases.
The process is iterative by design. Evaluation often sends you back to data preparation or modeling; deployment feedback revisits business understanding. Treat phase boundaries as decision gates, not linear milestones.
Six phases at a glance
| Phase | Objectives | Key activities | Typical outputs | Typical duration %* |
|---|---|---|---|---|
| Business Understanding | Align analytics with business goals and constraints | Stakeholder interviews; problem framing; success criteria; risk and scope | Problem statement, success metrics, project plan, feasibility note | ~15–20% |
| Data Understanding | Know what data exists and whether it can support the goals | Initial collection; EDA; quality checks; hypothesis sketches | Data dictionary, EDA report, quality assessment, gap analysis | ~15–20% |
| Data Preparation | Produce analysis-ready datasets and reproducible transforms | Selection, cleaning, integration, feature construction, formatting | Clean datasets, feature specs, transformation code/pipelines | ~25–35% |
| Modeling | Fit candidate models and compare approaches | Technique selection; design of experiments; train/validation protocol; tuning | Trained models, experiment logs, comparison tables | ~10–20% |
| Evaluation | Judge whether results meet business and technical bars | Metric review vs criteria; error analysis; cost/benefit; deploy decision | Evaluation report, go/no-go, documented limitations | ~10–15% |
| Deployment | Put the solution into operation and sustain value | Rollout plan; monitoring; maintenance; documentation; knowledge transfer | Deployed artifact, runbooks, monitoring dashboards, final report | ~5–15% |
*Percentages are indicative only; highly regulated or messy data domains shift time toward understanding and preparation.
flowchart TB
BU[Business Understanding]
DU[Data Understanding]
DP[Data Preparation]
MO[Modeling]
EV[Evaluation]
DE[Deployment]
BU --> DU
DU --> DP
DP --> MO
MO --> EV
EV --> DE
DE -.->|monitoring & new questions| BU
EV -.->|revisit features or data| DP
EV -.->|try other algorithms| MO
DU -.->|clarify goals or scope| BU
Phase deep dives
Business Understanding
| Sub-task | What to do | Output |
|---|---|---|
| Problem framing | Translate business language into an analytical or ML problem | Formal problem statement |
| Success criteria | Define measurable business and model-level success | KPIs, acceptance thresholds |
| Data mining goals | Specify modeling type (e.g. classification, forecasting) and constraints | Technical goal doc |
| Project plan | Resources, timeline, data access, ethics/compliance checkpoints | Plan with milestones |
Data Understanding
| Sub-task | What to do | Output |
|---|---|---|
| Initial collection | Acquire or catalog sources; document lineage where possible | Inventory, access agreements |
| EDA | Univariate/multivariate summaries; visualizations; segment checks | EDA notebook or report |
| Data quality | Missingness, duplicates, outliers, label noise | Quality scorecard |
| Early insights | Hypotheses that inform preparation and modeling | Short insight memo |
Data Preparation
| Sub-task | What to do | Output |
|---|---|---|
| Selection | Choose rows, columns, time windows aligned to the target | Subset definitions |
| Cleaning | Imputation, deduplication, outlier handling (with rationale) | Cleaning rules |
| Feature construction | Aggregations, ratios, encodings, domain features | Feature definitions |
| Integration | Joins across sources; consistent keys and grain | Integrated tables |
| Formatting | Types, schemas, splits compatible with modeling tools | Model-ready datasets |
Modeling
| Sub-task | What to do | Output |
|---|---|---|
| Technique selection | Match algorithms to data size, interpretability, latency | Shortlist with rationale |
| Train/test split | Respect leakage rules (time, groups, stratification) | Split protocol |
| Hyperparameter tuning | Search strategy; validation discipline | Tuned configs |
| Model building | Train final candidates; track seeds and code versions | Model artifacts + metadata |
Evaluation
| Sub-task | What to do | Output |
|---|---|---|
| Business criteria check | Compare metrics to thresholds from Business Understanding | Pass/fail vs criteria |
| Model review | Error analysis, robustness spot checks, fairness slices | Review notes |
| Deployment decision | Go, iterate, or stop — with documented trade-offs | Decision record |
Deployment
| Sub-task | What to do | Output |
|---|---|---|
| Deployment plan | Rollout, rollback, ownership | Runbook |
| Monitoring | Data drift, performance, latency, business KPIs | Dashboards, alerts |
| Maintenance | Retrain triggers, data refresh, dependency updates | Ops checklist |
| Final report | Lessons learned, limitations, handover | Closing documentation |
CRISP-DM vs other methodologies
| Dimension | CRISP-DM | TDSP (Microsoft) | SEMMA (SAS) | KDD Process |
|---|---|---|---|---|
| Origin / style | Cross-industry standard; phase-oriented | Team roles, agile sprints, code structure | Sample, Explore, Modify, Model, Assess | Research-oriented knowledge discovery |
| Business linkage | Strong explicit phase | Strong (business metrics, backlog) | Weaker; more data-centric | Strong on problem understanding |
| Iteration | Core to the model | Agile loops + phases | Implicit in Assess | Multiple loops |
| Typical audience | General analytics / ML | Enterprise Azure/data teams | SAS-centric shops | Academia / R&D framing |
Use CRISP-DM when you want a neutral, widely recognized scaffold; pair it with MLOps when production automation is required (see mlops.md).
CRISP-DM mapped to SDLC (high level)
Generic SDLC letters A–F here mean: A charter/requirements, B design, C build/integrate, D implementation detail, E test/verify, F release/operate — adjust to your org’s naming.
flowchart LR
subgraph CRISP["CRISP-DM"]
BU2[Business Understanding]
DU2[Data Understanding]
DP2[Data Preparation]
MO2[Modeling]
EV2[Evaluation]
DE2[Deployment]
end
subgraph SDLC["SDLC (illustrative)"]
A[A / B]
C[C / D]
D2[D]
E[E]
F[F]
end
BU2 --> A
DU2 --> A
DP2 --> C
MO2 --> D2
EV2 --> E
DE2 --> F
- Business Understanding → A/B (intent, scope, success definition).
- Data Understanding + Data Preparation → C/D (data products, pipelines, integration).
- Modeling → D (training code, experiments).
- Evaluation → E (verification against criteria).
- Deployment → F (release, monitoring, operations).
Adapting CRISP-DM for modern ML
- Deep learning: Modeling and Evaluation expand — longer training cycles, more hardware dependency, and need for baseline + ablation discipline. Data Preparation often includes large-scale labeling and augmentation policies.
- MLOps: Deployment is not a one-time step; it includes pipelines, registries, and continuous monitoring. Feedback loops are automated (retrain triggers) rather than only manual project reviews.
- Feature stores and real-time serving: Data Preparation and Deployment overlap with online/offline feature consistency and point-in-time correctness — classic CRISP-DM documents assumed batch analytics unless you extend them explicitly.
Anti-patterns
| Anti-pattern | Why it hurts | Better habit |
|---|---|---|
| Skipping Business Understanding | Optimizes the wrong metric; shelf-ware models | Lock success criteria before heavy modeling |
| Data leakage in preparation | Inflated offline metrics; production failure | Time-safe splits, group splits, pipeline encapsulation |
| Single evaluation metric | Hidden failure modes (e.g. good AUC, bad recall on minority) | Multi-metric dashboards + slice analysis |
| Treating Deployment as “IT handoff” | No monitoring owner; silent drift | Shared SLOs and runbooks from day one |
External references
- CRISP-DM 1.0 — IBM / SPSS consortium process model (widely cited standard document).
- Witten, Frank, Hall, and Pal — Data Mining: Practical Machine Learning Tools and Techniques — practical grounding for phases and evaluation.
Keep project-specific model documentation in docs/product/ and experiment logs in docs/development/, not in this file.