CDSP (DSP-110) study plan

Write a one-paragraph problem statement for three practice scenarios, each naming the decision being supported, the unit of analysis, and the outcome variable.

Distinguish predictive from inferential and descriptive framings by labeling ten short scenario descriptions and justifying each label in one sentence.

Draft a stakeholder-to-metric mapping table that lists who consumes each metric and what action it triggers.

Convert two vague business asks into testable hypotheses with explicit success thresholds and a stated time horizon.

Identify the data sources, grain, and join keys implied by each of your three week-1 problem statements, and flag gaps.

Enumerate at least eight data quality dimensions (completeness, uniqueness, timeliness, validity, consistency, accuracy, and others) and write one concrete check per dimension.

Design a sampling or cohort definition for one scenario and defend the inclusion and exclusion criteria.

Specify a leakage-prevention rule for each scenario, stating which fields are unavailable at prediction time.

Profile a public dataset end to end: row counts, null rates, cardinality, distribution shape, and duplicate keys, recording results in a table.

Implement at least six automated quality assertions in code and demonstrate each one firing on a deliberately corrupted copy.

Document a cleaning decision log that records each transformation, its rationale, and its reversibility.

Quantify the impact of one imputation or outlier treatment by comparing summary statistics before and after.

Build a train/validation/test split appropriate to the data's structure (time-based, grouped, or stratified) and justify the choice in writing.

Select two candidate model families for one scenario and state the assumptions, interpretability, and data-volume requirements of each.

Define the evaluation metric that matches the business cost structure, and explain why a naive accuracy baseline is inadequate.

Establish a trivial baseline and record its performance as the reference point for all later comparisons.

Train and tune at least two models on your prepared data, using cross-validation that respects the split design from week 4.

Compute and interpret a confusion matrix, precision, recall, and at least one threshold-independent metric such as AUC.

Sweep the decision threshold and show how precision and recall trade off against the business cost of each error type.

Run at least one sensitivity check, such as removing a feature or reseeding, and record whether conclusions hold.

Diagnose overfitting by comparing training and validation performance and state the remedy you would apply.

Interpret feature importance or model coefficients and separate correlational signal from claimed causation in writing.

Translate model outputs into business units (currency, customers, days) rather than raw scores.

List at least four threats to validity for your scenario, including selection bias, drift, and confounding, with a mitigation for each.

Draft an executive summary that leads with the decision and the recommendation, placing method detail after the conclusion.

Build a single chart that communicates the headline finding without requiring a caption, and critique it against a checklist for axis, scale, and color use.

Anticipate three stakeholder objections and write a concise, evidence-backed response to each.

State the limitations and confidence level of your findings in plain language a non-technical reader can act on.

Deliver a timed end-to-end case walkthrough covering framing, data preparation, model choice, interpretation, and communication within 45 minutes.

Complete a timed practice set or mock exam and tag every miss by domain to expose weak areas.

Rehearse verbal explanation of one technical result to a simulated non-technical audience and record the recording for self-review.

Consolidate a one-page formula and concept sheet covering metrics, split strategies, and validity threats for final review.