Original practice scenario inspired by The TJX Companies, Inc.'s public business context; this is not a reported interview question or a description of its internal systems. A fictional off-price retailer sells products that may not be replenished. A model trained on long-lived items is proposed for new short-lived stock. Describe an evaluation and baseline strategy.
Read model answer +
Match the evaluation to new-item forecasting rather than repeated observation of familiar products. Assortment changes, stock limits, and short selling windows make a simple random-row split misleading.
Work through it
Hold out complete items and later arrival periods. Use features known at allocation time, and separate product availability from observed sales.
Compare with category-level and attribute-based baselines. Score cumulative demand over the intended selling horizon, alongside sell-through and uncertainty where supply constrains observation.
Inspect performance by category, price band, and item lifetime. State how the model handles unseen attributes and how often it abstains or falls back.
Tradeoffs to discuss
Pooling categories helps sparse new items but can miss important differences. More complex item embeddings need evidence that they generalize beyond familiar catalog records.
Common pitfalls
Treating sold-out days as true zero demand biases training.
Evaluating only items with long histories excludes the difficult cases the proposal is meant to serve.
Original practice scenario inspired by The TJX Companies, Inc.'s public business context; this is not a reported interview question or a description of its internal systems. A synthetic retail pricing service has effective intervals for markdown prices. Stores may be offline and later replay price updates. Explain representation, conflict detection, and tests for choosing the price at a transaction time.
Read model answer +
Represent price schedules as versioned intervals with a defined timezone and endpoint convention. A sale should retain the applied price and schedule version, rather than being repriced silently after a later update.
Work through it
Use half-open validity intervals and integer minor-unit prices. Validate that intervals for the same item and store do not overlap within a schedule version.
Distribute versioned snapshots or ordered updates that an offline store can apply idempotently. Detect missing versions instead of guessing that the latest received update is complete.
Test exact start and end boundaries, a gap, an overlap, daylight-saving changes, and replay after reconnection. Reconcile the applied version against the authoritative schedule.
Tradeoffs to discuss
Offline continuity may require a documented fallback price. That policy trades operational availability against freshness and must be explicit.
Common pitfalls
Wall-clock string comparison across timezones is unreliable.
Overwriting historical schedules makes disputed transaction prices impossible to reconstruct.
