Your assignment
Original fictional quarterly demand desk. A planner asks for “a better forecast.” Turn that into a contract: target and units, aggregation grain, frequency, issue time, horizon, revision policy, available inputs and intended decision. Prerequisites: averages, Python lists and basic chronological data. No real operational decisions are made.
The Cedar source
The supporter fixture has 24 consecutive quarterly observations with a constructed four-quarter seasonal pattern and a small upward step each cycle. The first two cycles are 10,20,30,40 then 12,22,32,42. This is a teaching construction, not evidence of real seasonality. A missing observation is unknown; a recorded zero is observed zero demand. Do not silently remove a missing quarter and shorten the seasonal period.
Your attempt
Define “forecast next quarter at the end of quarter 8” and “forecast eight quarters once at quarter 16.” Which inputs exist at those issue times? Can final realised promotion spending or a later revised demand value be used? How would you record a plan that was known before issue time?
Worked diagnosis
Use only information available at the issue time. A later actual can be used as an evaluation target, not as an input to that earlier prediction. A plan known in advance may be a legitimate input if its as-of version is documented. Keep initial and revised actuals separate when evaluating real forecasts; a current final extract alone cannot prove what was known historically.
Metric contract
Cedar uses actual minus prediction for error: positive bias means underforecasting. MAE averages absolute errors; RMSE square-roots average squared errors; mean signed error measures bias but can cancel over- and under-predictions. Report the units and horizon. MAPE has problems at zero actuals, so Cedar uses MAE/RMSE rather than silently dividing by zero.
Small exercise
Actual values are 0,4 and predictions 2,2. Compute signed errors, MAE, RMSE and bias. Explain why the mean error can look perfect while both predictions are wrong.
Worked solution
Errors are-2,+2; MAE 2; RMSE 2; bias 0. Zero signed bias does not mean accurate individual forecasts. One actual is zero, so the usual percentage-error expression is undefined for that row. Do not hide the undefined term by deleting it without an agreed evaluation contract.
Review and explain
Score 0–2 for target/time contract, available-information boundary, missing-vs-zero, metric definitions and honest limits. Write a 90-second explanation of why the same model can answer different planning questions at different horizons. This is unsaved practice, not a credential.
Authoritative reference
3/accuracy.html Rob J Hyndman and George Athanasopoulos, Forecasting: Principles and Practice, 3rd ed: evaluating forecast accuracy. The original Cedar examples are independent teaching constructions. Next: Cedar 2.
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.