- Write and explain a basic SELECT, WHERE, GROUP BY, and ORDER BY query.
- Recognize a key, a missing value, a duplicate, and a data type.
- Filter, group, summarize, and make a simple chart in Pandas.
- Read an error message and reduce a problem to a smaller test.
SELF-GUIDED MANUAL / INDEPENDENT STUDY
Be your own program manager. Not your own answer key.
The Analyst can support serious independent practice, but it does not replace foundational instruction, feedback, or professional experience. This manual gives you a route, working discipline, and honest review process without turning the assignments into walkthroughs.
THE 15-MINUTE START
Start small enough to learn honestly.
- 01
Read the Workbench Guide once so you know how the workbench stores, runs, and exports work.
- 02
Open Assignment 01 in Supported mode. Treat it as a readiness diagnostic, not a speed test.
- 03
Block one uninterrupted 90-minute session. Finish with saved notes and an exported file even if the analysis is incomplete.
Know what this resource assumes.
You do not need every advanced prerequisite on day one. You do need enough fluency to investigate rather than spend the entire session decoding syntax.
- Joins and reusable Python functions before Assignment 02.
- Sampling and confidence intervals before Assignment 03.
- Missing-data and time-aware joins before Assignment 04.
- The stated modeling, forecasting, causal, NLP, or governance prerequisite before its advanced assignment.
Begin Assignment 01, but stop if syntax prevents you from investigating the business question. Spend focused time on introductory SQL and Pandas, then return. That pause is a readiness decision—not failure. The Analyst is a practice environment, not a from-zero programming course.
TECHNICAL PREFLIGHT / DO THIS ONCE
Open Assignment 01 in Supported mode. Wait for DuckDB to report ready and run the SQL starter. Open Python, allow the first runtime load to finish, run the starter, and confirm its figure renders. Download a test .analystcase file and restore it before investing substantial time.
LOCAL STORAGE IS NOT A BACKUP
Private browsing, clearing site data, managed-device resets, or changing browser profiles can erase local drafts. The first Python load and package installation also require a network connection. Export after every serious session.
DATA BOUNDARY
Use the fictional Meridian estate. Do not paste employer, client, patient, customer, or other protected data into the workbench.
Budget for revision, not just completion.
Hours describe investigation time, not mastery. Add foundation study where prerequisites are weak and a second pass after every review.
16-WEEK CORE
30–51 prepared-learner hours / 51–84 newcomer hours01 → 02 → 03 → 04 → one advanced elective
Use assignments 01–04 to build the common judgment core. Choose 05, 06, 07, 08, or 09 only after checking its prerequisites on the Assignments page. Reserve the final week for revision, not another first attempt.
24-WEEK EXTENDED
52–83 prepared-learner hours / 89–138 newcomer hours01–05 → one build practicum → one audit practicum
Complete the core through point-in-time modeling. Then choose Stock or Queue for a larger build and Orion or Model Risk for an audit. Put a recovery week after every two or three assignments.
ALL-NINE STUDIO
74–115 prepared-learner hours / 127–192 newcomer hours01 → 02 → 03 → 04 → 05 → 06 → 07 → 08 → 09
This is an intensive studio, not ordinary weekly homework. The ranges exclude foundation study, debugging, delayed self-review, and revision. Use it only when you can protect substantial time and tolerate unfinished work.
Use the Assignment Register to compare prerequisites, newcomer ranges, complexity dimensions, and the actual SQL/Python work before selecting an advanced assignment. Four carefully reviewed assignments will develop more judgment than nine rushed first drafts.
Reduce support deliberately.
The selector in the workbench changes the starting material, not the data or standard of judgment.
| MODE | WHAT YOU RECEIVE | WHEN TO USE IT |
|---|---|---|
| SUPPORTED | Assignment-specific SQL, Python, working prompts, and opening evidence. | Your first attempt at 01–02, or any assignment whose domain is new to you. |
| GUIDED | A general evidence frame and table-loading examples without assignment-specific analytical direction. | After one credible supported handoff; a sensible default for the middle sequence. |
| INDEPENDENT | The brief, source estate, required handoff, and nearly blank worksheets. | A fresh attempt after you can frame grain, cutoff, diagnostics, and deliverables yourself. |
Self-guided is not Independent mode. A self-guided learner may—and often should—use Supported mode. Each mode has separate local work, so choose before beginning a serious attempt and export before switching.
Use the same professional loop every time.
The loop is intentionally stable while the role, domain, methods, and ambiguity change.
- 01
Inbox
Before querying, write the decision, audience, cutoff, intended grain, and the evidence that could change your position.
- 02
Data register
Inspect keys, coverage, clocks, relationships, reliability notes, and dangerous joins before deciding what the tables mean.
- 03
Investigate
Reconcile the evidence layer in SQL. Use Python when visualization, inference, simulation, backtesting, modeling, calibration, or error analysis improves the decision.
- 04
Evidence
Record only claims you can trace to actual work. Include the relevant population and limitation rather than pasting a naked number.
- 05
Handoff
Draft a clear recommendation, qualification, or responsible refusal; bind the requested artifacts; and export the first complete attempt.
- 06
Review
Return after a break. Audit the work as another analyst, consult the post-attempt prompts, revise materially, and write down what changed and why.
Review the argument, not resemblance to a hidden answer.
Export version one first. Then interrogate the work across seven professional dimensions.
What does one row represent at every important step, and where could a join multiply it?
Could every feature, status, label, and comparison legitimately have been known at the assignment cutoff?
Did you test a credible rival explanation, not just confirm the first story that fit?
Can another analyst recover the population, exclusions, transformations, and result from the exported work?
Does the language distinguish descriptive, predictive, and causal claims—and match the strength of the design?
Is the recommendation proportionate to the stakes, constraints, and evidence? What would reverse it?
Can the intended audience find the decision, population, material limitation, owner, and next action quickly?
SPOILER CONTROL
OPTIONAL SPOILER NOTES The assignment prompts below are safe to use after exporting a complete first attempt: they name areas of inquiry, not planted truths, preferred methods, magic numbers, or conclusions. Full instructor notes disclose mechanisms and should remain closed until after revision.
POST-ATTEMPT REVIEW PROMPTS
Open only the assignment you completed.
Use these after version one. A self-review is a reflection structure—not a credential or correctness certificate.
01The Monday ScorecardCC-241202REVIEW PROMPTS
Lines of inquiry
- What population, time cohort, business grain, and scale does each opening number represent?
- Which repeated records are technical migration artifacts, and which could be legitimate repeat contacts?
- How does survey response coverage limit the population claim?
Evidence and diagnostics
- Score-scale and normalized-score distributions by source/cohort
- Ticket ID, interaction fingerprint, and pre/post-join count reconciliation
- Response numerator and eligible-population denominator with missingness profile
Decision and handoff qualities
- Defines the metric before naming the number
- Separates respondent evidence from all-customer claims
- Explains why the chosen one-number, paired-number, or refusal approach is fit for leadership use
02The Quarter That MovedCM-240708REVIEW PROMPTS
Lines of inquiry
- Which identifier is globally canonical, and where is source uniqueness only local?
- Which clock answers occurrence, source recording, warehouse availability, and financial close?
- Which captured transaction facts own revenue and fulfillment measures?
Evidence and diagnostics
- Key-collision and uniqueness profiles by source/period
- Clock-offset and late-arrival exception distributions
- Header/line/event/shipment row and amount bridges before and after joins
Decision and handoff qualities
- Issues a bounded certification with quantified exceptions
- Preserves raw fields and expresses reusable correction flags
- States a stable rerun and next-close policy
03The Navigation VoteGX-250505REVIEW PROMPTS
Lines of inquiry
- What was randomized, who was eligible, and what estimand was registered?
- Does event-level joining change the weight assigned to experimental units?
- What can an increase in navigation depth establish about value, friction, or conversion harm?
Evidence and diagnostics
- Assignment reconciliation, allocation balance, and missingness by variant
- Session-grain primary/guardrail effects with uncertainty
- Clearly labeled exploratory slices and repeated-account sensitivity
Decision and handoff qualities
- Separates confirmatory evidence from post-hoc discovery
- Does not translate nonsignificance into equivalence
- Connects rollout, guardrails, monitoring, and claim strength
04Rollback Before DawnOP-250320REVIEW PROMPTS
Lines of inquiry
- How do firmware, weather, telemetry availability, alerts, and field work align in time and geography?
- What would telemetry absence look like if failures themselves interrupt reporting?
- Which comparison can inform scope without pretending to be a randomized control?
Evidence and diagnostics
- Asset/channel expected-versus-observed reporting grid
- Storm/firmware/region trend and counterevidence views
- Cutoff audit separating physical failure, measurement change, alert, and work request
Decision and handoff qualities
- Names an actionable containment or monitoring scope
- Calibrates causal language to overlapping mechanisms
- Provides thresholds or evidence that would reverse the decision
05The 7:30 Capacity CallFO-250320REVIEW PROMPTS
Lines of inquiry
- What was knowable at 07:30 for each historical appointment?
- What operational promise defines the target, and when does its label become available?
- How should review capacity, asymmetric cost, and population shift shape the action policy?
Evidence and diagnostics
- Point-in-time one-row-per-appointment mart and leakage inventory
- Naive baseline, forward validation, calibration, and operational slices
- Threshold/coverage table under stated review capacity and cost assumptions
Decision and handoff qualities
- Separates risk prediction from intervention effect
- Keeps action at appointment/capacity grain and protects employee data
- Defines human authority, abstention, and shadow-monitoring gates
06Forty-Eight Hours of StockSP-251201REVIEW PROMPTS
Lines of inquiry
- How do technical events map to physical movements, accepted receipts, and complete daily demand?
- Which forecast loss reflects inventory decisions across intermittent and seasonal series?
- How do demand, lead-time, donor, lane, MOQ, and approval constraints interact?
Evidence and diagnostics
- Ledger/receipt reconciliation and all-days demand spine with explicit zeroes
- Multiple rolling-origin folds, untouched final origin, bias, and series-type performance
- Recipient/donor simulation plus referential and constraint checks on the action file
Decision and handoff qualities
- Distinguishes forecast accuracy, inventory risk, and action value
- Explains pooling, uncertainty, stress, and reproducibility assumptions
- Delivers feasible actions with monitoring, override, and expiration rules
07The Orion RenewalPR-260119REVIEW PROMPTS
Lines of inquiry
- What business unit owns each route, stop, work-order, visit, and workforce outcome?
- What does simultaneous adoption prevent the data from identifying cleanly?
- Which descriptive changes, attribution claims, and procurement actions remain distinct?
Evidence and diagnostics
- Stable-grain and fanout reconciliation including rescheduled work
- Pre/post trends, meaningful heterogeneity, placebo, or sensitivity views
- Outcome/guardrail definitions and evidence against individual ranking
Decision and handoff qualities
- Makes a procurement decision despite bounded attribution
- Refuses or safely constrains employee ranking
- Proposes prospective measurement or contract conditions for the next decision
08The Queue Nobody OwnsNL-241203REVIEW PROMPTS
Lines of inquiry
- What is one eligible intake example, and which text existed before the routing decision?
- Does the available label represent adjudicated correct ownership or eventual handling?
- Which populations and error types require abstention or mandatory human review?
Evidence and diagnostics
- Conversation-grain corpus, intake eligibility audit, and grouped forward split
- Transparent baselines, per-class/coverage results, calibration, and structured error analysis
- Sensitive/unsupported population slices and shadow-label event contract
Decision and handoff qualities
- Treats target validity as a product question, not just a modeling detail
- Refuses unsupported live automation while preserving a useful remainder
- Defines promotion, pause, rollback, access, and monitoring evidence
09Too Good to ShipMR-260120REVIEW PROMPTS
Lines of inquiry
- Which predictors were available at scoring time, including source-watermark constraints?
- Do accounts or future calendar regimes cross the submitted validation boundary?
- What evidence would authorize shadow, restricted evaluation, or customer action?
Evidence and diagnostics
- Complete predictor inventory including non-prefixed future outcomes
- Entity overlap audit and grouped temporal validation comparison
- Allow-listed baseline, calibration/segment audit, and access/export review
Decision and handoff qualities
- Explains why reproducibility does not establish validity
- Assigns a clear model disposition without claiming no future model can work
- Blocks unsupported action and names the smallest safe next gate
Diagnose the kind of stuck.
Mess is part of the assignment. Unstructured thrashing is not.
Reduce to one table and five rows. Inspect exact column names and types. Add one clause or transformation at a time, and keep the last query that ran.
Return to the Data Register. Write what one row represents, test key uniqueness, count before and after joins, and identify which clock each timestamp records.
Write two competing recommendations. Under each, list supporting evidence, counterevidence, consequence of being wrong, and the next fact that would change the decision.
Save and export. Write the smallest unresolved question in scratch notes. Review the prerequisite—not the assignment spoiler—or use AI Help Packet to create focused context for an external AI tool. Review the file before sharing, then verify any advice by running the work yourself.
Finish with work another person could use.
A completed worksheet is not automatically a completed assignment.
- The analytical population, grain, cutoff, and exclusions are explicit.
- SQL runs end to end and includes decision-material reconciliation checks.
- Python is reproducible and used where it adds analytical value.
- Important claims are recorded as evidence tied to actual runs.
- The final brief leads with a decision, qualification, or responsible refusal.
- Material uncertainty, risks, owners, next actions, and reversal conditions are visible.
- Every requested artifact is bound and the .analystcase file is downloaded.
- A delayed self-review produced at least one documented revision—or a defensible reason for none.
Keep the exports distinct: the .analystcase file is the restorable, reviewable record to save at every milestone. The portfolio .zip is a conventional human-readable copy to export after the work is polished. The optional AI-context .md is a learner-selected help brief—not a submission—and is shared only when you attach it to an external tool. Course, section, and learner-ID fields may remain blank for independent study.
READY TO BEGIN
START SUPPORTED