Listing Lifetime Value — Estimation (ML Design)

Estimate the lifetime value of a listing — projected revenue over the next 12 months. Used downstream for ranking, host outreach, and inventory acquisition decisions. The prompt is intentionally open and rewards strong scoping. Requirements Functional Predict per-listing 12-month booking revenue (or a related target the candidate proposes). Output consumed by ranking (boost high-LTV listings) and by host-success teams (prioritize outreach). Non-functional Batch inference is acceptable; daily refresh. Predictions should be calibrated — absolute dollar values, not relative scores. Notes Problem framing. Two-step: ask clarifying questions to pin down the target (calendar-year revenue? 12-month forward revenue?), then pick regression vs classification. The community version of this question was answered with "bucketize and classify" because the candidate did not feel comfortable with regression; the interviewer accepted but the regression framing scores higher. Target choice. Forward 12-month bookings × average nightly price. Censoring (new listings, host pauses) is the most subtle issue — handle with survival analysis or simply train on listings with ≥ 12 months of history. Features. Listing (price, location, amenities, host tenure, photo quality embedding), historical performance (last-12mo bookings, conversion rate, cancellation rate), market (geo demand index, seasonality). Model. GBDT (XGBoost / LightGBM) is the standard answer for tabular regression; argue against deep models on the basis of feature-density + interpretability. Eval. Regression: MAE / RMSE on a held-out future period (not random split — temporal split avoids leakage). Calibration plot. Per-segment error (new vs mature listings, low vs high LTV). Cold start. For new listings, fall back to a market-segment average with a confidence interval; flag as "low confidence" downstream. Production loop. Daily batch; write predictions + confidence to a feature store consumed by ranking and outreach teams. The interviewer pushed on the regression / classification choice — be ready to defend either, but the regression framing is the senior answer. Preparation Pre-script the regression-vs-classification trade-off so it lands in one minute. Drill the temporal-split eval argument (random split leaks future into past). Pre-write the cold-start fallback (market-segment average + low-confidence flag). Pair-prep with the search-ranking MLSD; both share the listing-features section.