Original practice scenario inspired by TotalEnergies SE's public business context; this is not a reported interview question or a description of its internal systems. A fictional energy desk forecasts next-day renewable generation using weather forecasts, equipment status, and realized output. Describe an evaluation that reflects information available when the forecast would actually be issued.
Read model answer +
Replay each forecast origin using only data available at that origin. Realized weather and retrospectively corrected equipment status can make an offline model look stronger than its operational version.
Work through it
Record issue timestamps separately from valid timestamps for weather forecasts. Select the version published before each simulated decision cutoff.
Use rolling-origin evaluation, fit preprocessing within each training window, and compare against persistence and seasonally appropriate baselines.
Report errors by horizon, site, and low-generation periods. Include missing-input behavior and interval coverage if uncertainty estimates inform scheduling.
Tradeoffs to discuss
Historical forecast vintages may be incomplete. A smaller honest replay is more useful than a larger evaluation with silently substituted realized weather.
Common pitfalls
Random row splits leak nearby temporal structure.
Good aggregate error does not guarantee reliable performance during the operating conditions that matter most.
Original practice scenario inspired by TotalEnergies SE's public business context; this is not a reported interview question or a description of its internal systems. A synthetic logistics dataset shows that one energy terminal has longer average truck turnaround than another. Its trucks also arrive with different cargo types and appointment patterns. How would you investigate the difference?
Read model answer +
Separate a descriptive difference from a causal explanation. Cargo mix, arrival congestion, and incomplete visits can change the observed average without identifying a worse terminal process.
Work through it
Define start and end events, timezone rules, and how open visits are handled. Inspect duration distributions and timestamp quality before comparing means.
Compare like-for-like cargo and appointment strata, then standardize to a common mix. Show sample sizes and uncertainty rather than only a single adjusted estimate.
Examine time-of-day queues and process stages to locate where delay accumulates. If testing an operational change, choose a credible randomized or phased comparison.
Tradeoffs to discuss
Adjustment improves comparability only for measured differences. Median and tail metrics answer different service questions and should be reported together.
Common pitfalls
Dropping unfinished long visits biases results downward.
An adjusted association alone does not establish that terminal choice caused the delay.
