ORIGINAL COMPANY-CONTEXT PRACTICE
Choose forecast origins for electricity demand
Original practice scenario inspired by NextEra Energy, Inc.'s public business context; this is not a reported interview question or a description of its internal systems. A fictional utility evaluates hourly demand forecasts issued each morning for the next day. Its dataset includes delayed meter readings and revised temperature forecasts. Design the training and scoring boundaries.
Written for independent practice using company context; this is not a verified question from a company interview.
A REASONED APPROACH
Model answer.
Each simulated morning must expose only records and forecast vintages available by that issue time. The prediction horizon and the data availability delay together define the real training boundary.
Work through it
- Store event time and ingestion time separately. Exclude recent meter readings that had not yet arrived, even if their event timestamps precede the forecast origin.
- Fit each candidate on the same eligible history and score the same next-day hours. Compare with previous-day and previous-week baselines using documented daylight-saving handling.
- Aggregate errors over completed forecast days and inspect peaks, holidays, and missing-input periods. Keep the final next-day operational forecast separate from scored historical predictions.
Tradeoffs to discuss
- Retraining at every origin costs more than a coarse backtest but better reproduces the deployed update schedule. Use a smaller replay if necessary.
Common pitfalls
- Revised weather forecasts leak future knowledge.
- A final model fit on all available data must not be used to rescore earlier origins.