Walk-Forward Validation for Time Series Quiz

Answers and explanations appear as you take the quiz.

1. A monthly model trains through June and forecasts July. In an expanding-window evaluation, which history is used for the next August forecast?
   • August outcomes combined with January through June
   • July alone, discarding all earlier months
   • January through July
   • January through June with months randomly reordered
2. Which distinction separates a rolling training window from an expanding one?
   • A rolling window retains a recent span while an expanding window accumulates earlier history
   • A rolling window accumulates all earlier history while an expanding window retains a fixed recent span
   • Both windows retain every historical row; only their forecast horizons differ
   • The distinction depends only on how often predictions are issued
3. A business makes decisions four weeks ahead. Why might next-day evaluation alone be insufficient?
   • Next-day error can be multiplied by 28 to determine four-week-ahead error
   • Strong next-day performance guarantees the same ranking of models at four weeks
   • The training-window length alone determines which horizon has been evaluated
   • The evaluated horizon does not match the actual decision
4. A feature averages sales through the same day whose sales are being predicted. Which correction addresses the direct leakage?
   • Include the next day's sales in the average as well
   • Shift the history so the target day's outcome is excluded
   • Fit a larger model to the unshifted feature
   • Randomize the order of the evaluation dates
5. Why should historical external data reflect publication delays and revisions in a forecast backtest?
   • Published revisions may be used whenever they are more accurate
   • A feature is safe whenever its observation date precedes the target, regardless of publication date
   • Later information may not have been available at the forecast origin
   • A time-ordered splitter automatically reconstructs historical releases
6. What additional check is needed even after using an ordered splitter such as TimeSeriesSplit?
   • Verify that feature preparation did not use future information
   • Assume every preprocessing operation is automatically safe
   • Use the whole dataset to calculate all target averages
   • Replace the forecast horizon with a random split
7. Overlapping evaluation horizons produce several forecasts for the same outcome. What should the report specify?
   • That every forecast of the same target should be counted as an independent outcome
   • That only the forecast with the smallest realized error is retained
   • That forecast origins can be discarded once the target date is known
   • Whether scoring is by forecast opportunity or aggregated by target time
8. Which data condition requires extra care when using TimeSeriesSplit to compare equal-duration test folds?
   • Observations arrive on a fixed daily schedule
   • Observations are irregularly spaced in time
   • Each fold contains the same fixed number of regularly spaced observations
   • The training window expands across successive folds