Supply Chain Analyst Interview Questions for AI Training Work

I use a moving average for stable, low-variance products where simplicity and interpretability matter more than precision, and I move to a more advanced model, like exponential smoothing or a regression approach, for products with seasonality or trend. Model complexity should match the actual variability in the demand pattern, not just be the most sophisticated option available.

I check for duplicate transaction records, inconsistent unit-of-measure conversions across systems, and gaps in historical data caused by stockouts, since a stockout period looks like low demand but actually reflects unmet demand. Skipping this step tends to produce forecasts that quietly repeat the errors in the source data.

I look at the supplier's historical on-time delivery rate, geographic and geopolitical exposure, and the lead time to qualify an alternate source if something goes wrong. I combine that with the business impact of a disruption to prioritize which single-source relationships actually need a mitigation plan first.

I track forecast accuracy using mean absolute percentage error at the SKU level, but I also watch for bias, meaning whether the forecast consistently overshoots or undershoots, since a model can have low average error while still being systematically wrong in one direction.

I calculate safety stock based on the variability in both demand and lead time together, not just demand alone, since lead time variability is often the larger driver of stockout risk for imported or single-source components. I revisit the calculation whenever a supplier's performance shifts materially.

I'd first quantify how much buffer stock exists and how long it covers current demand, then identify whether alternate suppliers can be qualified in time. I'd flag the affected downstream orders to the planning team immediately so they can adjust customer commitments before the shortage actually hits production.

I'd check whether a real shift happened, like a competitor entering the market or a pricing change, versus a data issue, like a promotional period skewing the historical baseline the model was trained on. I would not just recalibrate the model without understanding the underlying cause first.

I'd weigh the cost of qualifying and maintaining a second supplier against the business impact if the current single source failed, factoring in how replaceable that component is and how long a disruption would take to resolve. Diversification isn't free, so I only recommend it where the downside risk clearly outweighs the added cost.

I underestimated demand for a seasonal product because I anchored too heavily on the prior year without accounting for a marketing push that increased visibility. I flagged the miss to stakeholders quickly, worked with procurement on an expedited order, and adjusted my model to weight recent trend signals more heavily going forward.

I needed to explain to a sales leader why we couldn't commit to a large order on the original timeline due to a supplier risk. I framed it in terms of delivery dates and dollar impact rather than the underlying data analysis, which made the tradeoff concrete enough for them to make a fast decision.