Your planning meeting
A teammate says,“Seasonal won, so use it to automate replenishment.” Write a one-page review using Cedar 2 and 3. Separate observed synthetic backtest evidence from deployment readiness. Deliver a forecast contract, frozen-method record, horizon table, uncertainty note, scenario assumptions and proposed next evidence.
Your attempt
Which planning question was evaluated? Does the model see genuine historical vintages? Are sales a valid proxy for demand under stockouts? What happens after a promotion or permanent operating change? Identify what the tiny seasonal fixture cannot answer.
Reference solution
The synthetic seasonal baseline wins the agreed development comparison and produces rolling MAE 2, but this does not validate a real replenishment policy. Actual stockouts can censor sales below demand; calendar changes and promotions can change the target process. First require representative historical data, issue-time availability, matched horizons, baseline comparisons and an agreed human-reviewed operating study. No real deployment is supplied by this project.
Forecast versus scenario
Base point forecast 20 with an assumed 10% promotion uplift gives a planning scenario 22. This is not a fitted causal effect, a newly validated forecast, or an uncertainty bound. State who provided the uplift, when it was known, and how sensitive the plan is to 0%,10%,20%. Those scenarios are 20,22,24; they are assumptions, not probabilities.
An evidence table
Report case count, units, horizon, split and evaluation type with each metric. Keep the method frozen for the reported test; do not pool a fixed-origin eight-quarter result with rolling one/two-step performance.
evidence | result | limitation
rolling seasonal h1/h2 | MAE2 each | constructed smooth fixture
fixed origin16 h1–8 | MAE3 | different forecast task
error-band test | 2/4 covered, width6 | tiny one-step sample
promotion scenario | 20 ->22 | assumed uplift, not causal estimateMonitoring proposal
Define target freshness, missing/zero handling, horizon-specific mature errors, signed bias, band hit counts and widths, plus stockout/promotion flags if available. Preserve forecasts as issued instead of overwriting them with revised predictions. A sudden break should trigger investigation and review under an agreed operating policy, not an automatic unreviewed model switch.
Honest portfolio statement
“Built reproducible baseline backtests and an error-band audit on original fictional fixtures, with chronology checks, horizon comparisons and a decision memo.” Do not claim real inventory savings, production model accuracy, guaranteed coverage, certification or employment outcomes.
Interview defense and rubric
Score 0–2 for chronological evidence, correct horizons, uncertainty limits, scenario distinction and honest communication. Answer: why can bias 0 coexist with MAE 2; when is a seasonal baseline inappropriate; why is realised future promotion spend unavailable to an earlier issue? Record one unresolved risk and the next evidence required.
Authoritative reference
3/simple-methods.html Rob J Hyndman and George Athanasopoulos, Forecasting: Principles and Practice, 3rd ed: benchmark methods. These public references remain public; supporter value is the authored offline project and review workflow, not exclusive access to a public book.
Try the activity before reading its reference. Keep your sources, commands and assumptions with your work. Companion activities are unsaved self-practice and do not award a certificate or change lesson progress.