AI Demand Forecasting Specialist Interview Questions

A strong answer defines demand forecasting, explains limitations of naive methods (lag, no external signals), and highlights how AI captures nonlinear patterns, seasonality, and exogenous variables.

Cover temporal ordering, autocorrelation, non-i.i.d. nature, seasonality, and the need for time-aware train/test splits to prevent data leakage.

Explain that MAPE averages percentage errors per item (problematic near zero demand), while WAPE is volume-weighted; WAPE is preferred for SKU-level retail forecasting.

Mention trend (piecewise linear or logistic growth), seasonality (Fourier terms for weekly/yearly), holiday effects, and optional regressors for external signals.

Stationarity means constant mean/variance/autocorrelation over time; ARIMA requires it (or differencing to achieve it); non-stationary data leads to spurious model fits.

Cover calendar features (day-of-week, holidays, Ramadan/Chinese New Year), lag features, rolling statistics, promotional flags and depth, price elasticity, weather, and external economic indices.

Discuss top-down, bottom-up, middle-out approaches, and optimal reconciliation methods (MinT, OLS) that ensure forecast coherence across levels.

Mention Croston's method, SBA variant, zero-inflated models, two-stage models (classification + regression), and appropriate error metrics that don't penalize zero-demand periods unfairly.

Explain how demand signal amplification propagates upstream; AI forecasting reduces it by sharing accurate consumer-level signals, reducing order batching, and enabling demand sensing.

Discuss promotional uplift modeling, price elasticity features, baseline vs. incremental decomposition, cannibalization effects, and the challenge of unseen future promotions.

Explain expanding window or sliding window CV, gap periods between train and test to simulate forecast horizon lag, and why random k-fold is inappropriate for temporal data.

Bias indicates systematic over- or under-forecasting; over-forecasting causes excess inventory costs, under-forecasting causes stockouts; consistent bias direction matters more than occasional large errors.

Feature stores provide consistent, versioned, low-latency feature serving; prevent training-serving skew; enable feature reuse across models; centralize feature computation and governance.

Global models share patterns across SKUs (better for cold-start, fewer SKUs), local models capture idiosyncratic behavior; hybrid approaches and hierarchical models often work best; discuss scale trade-offs.

Cover missing data (zero-fill vs. null), outlier detection (COVID anomalies), stockout masking (censored demand), duplicate records, clock changes, and use Great Expectations or similar for automated validation.

Discuss variable selection networks, gated residual networks, multi-head attention across time steps, static covariate encoders, quantile regression outputs, and how attention reveals which features/timesteps drive predictions.

Discuss attribute-based similarity (transfer from similar SKUs), market-level analogues, launch curve templates, meta-learning approaches, few-shot learning, and using LLMs to extract product attribute embeddings.

Cover data pipeline orchestration (Airflow), feature store (Feast), model registry (MLflow), batch inference (SageMaker Batch Transform or Spark), monitoring (drift detection, accuracy dashboards), alerting, and automated retraining triggers.

Translate forecast error into inventory carrying cost reduction, stockout revenue loss avoidance, markdown reduction, and improved OTIF; use simulation or sensitivity analysis; present in P&L terms executives understand.

Discuss difference-in-differences, synthetic control methods, instrumental variables, uplift modeling (meta-learners), and how causal models improve counterfactual scenario planning.

Explain pre-training on large time-series corpora, zero-shot forecasting ability, reduced need for per-series tuning; limitations include domain shift, lack of exogenous variable support, and interpretability challenges.

Discuss demand unconstraining methods (EM algorithm, Kaplan-Meier), Tobit models, mixture models that separate latent demand from observed sales, and inventory record integration to detect stockout periods.

Cover quantile regression, conformal prediction, Monte Carlo dropout, distributional outputs (Negative Binomial for count data); connect to safety stock optimization, service-level-driven replenishment, and risk-aware planning.

Discuss randomized SKU or store assignment, pre-period matching, traffic/forecast splitting, interference risks, metric selection (accuracy + business KPIs), statistical power analysis, and minimum detectable effect calculation.

Discuss stakeholder trust requirements, regulatory constraints (GDPR right to explanation), SHAP for feature importance, attention visualization in Transformers, LIME as local surrogate, and when to prioritize accuracy over explainability.

Walk through data pipeline audit (upstream schema changes, missing feeds), model input drift analysis, structural break detection (COVID recovery, market shifts), feature staleness, and whether the issue is systemic vs. localized.

Discuss stratified error analysis, building specialized models for high-value SKUs, incorporating additional data (e.g., sell-through, web traffic), ensembling approaches, and aligning with business to understand revenue impact priorities.

Build scenario-based forecasts for multiple promotional configurations, use historical promotion response curves, create parameterized models that accept promotional inputs, and establish rapid re-forecast workflows once details are finalized.

Discuss anomaly detection on sentiment streams, separating genuine demand signal from noise during crises, short-term vs. long-term impact modeling, human-in-the-loop review, and fail-safe model overrides.

Discuss model simplification, incremental/online learning, distributed training (Spark, SageMaker), feature caching, pre-computation of expensive features, batch vs. streaming inference, and tiered model architectures.

Cover higher-frequency data ingestion, finer-grained seasonality modeling, daily anomaly handling, updated cross-validation windows, faster model serving requirements, and new data sources (weather, traffic) that matter at daily granularity.

Discuss distinguishing supply-side artifacts from demand-side signals, working with the client to get true consumer sell-through data, model debiasing, and the risk of modeling ordering patterns vs. actual demand.

Discuss lightweight stack (Prophet + LightGBM, open-source orchestration with Prefect, free-tier feature store, batch inference on spot instances), prioritizing high-impact SKUs, and phased scaling strategy.

Discuss building flexible scenario models, leveraging analogous product launches, using the new marketing plan as input features, communicating uncertainty ranges, and establishing rapid feedback loops post-launch.

Consider market-level demand shifts (recession, competitor entry), changing consumer behavior, model concept drift, feature distribution shift, and discuss implementing bias correction layers and adaptive retraining triggers.

Describe the RAG architecture: user question → LangChain agent → SQL tool querying forecast database → LLM synthesizes answer with context; include guardrails for accuracy, source attribution, and hallucination prevention.

Explain fine-tuning a sentiment/topic model on financial text, extracting named entities (product mentions, market segments), scoring demand sentiment, and joining these features to your time-series data at the appropriate lag.

Describe defining feature views (batch for historical training, online for real-time inference), registering features in a central registry, versioning, point-in-time correctness for training, and eliminating training-serving skew.

Describe task dependencies, sensor operators for data availability, parameterized DAGs for different forecast horizons, retry policies, Slack alerting on failure, and how to handle backfill scenarios.

Cover experiment naming conventions, logging metrics (WAPE, MASE, bias), parameters, artifacts (model files, SHAP plots), model registry stages (Staging → Production), and how to compare runs programmatically.

Describe containerizing the model, defining batch transform input/output formats on S3, scheduling with EventBridge or Step Functions, using SageMaker Model Monitor for data drift, and CloudWatch alarms for accuracy thresholds.

Define expectations (non-negative sales, reasonable ranges, completeness of date sequences, freshness), configure checkpoint runs in CI/CD, generate data docs, and set up alerting on expectation failures.

Explain staging models for raw data cleaning, intermediate models for joins and transformations, mart models for forecast-ready datasets, testing with dbt tests, documentation, and scheduling with Airflow.

Describe defining function schemas (retrain_model, get_accuracy_metrics, analyze_anomaly), the agent loop architecture, context window management, error handling, and human-in-the-loop approval for sensitive actions.

Define the sweep config (Bayesian vs. grid search), search space (max_depth, learning_rate, n_estimators), objective metric (validation WAPE), parallelism strategy, and how to analyze sweep results to select the best configuration per category.

Look for ownership, structured root-cause analysis, transparent communication with stakeholders, corrective action taken, and systemic improvements implemented to prevent recurrence.

Assess ability to translate technical concepts into business language, use of backtesting evidence, gradual trust-building (shadow mode, A/B), respect for domain expertise, and collaborative model refinement.

Look for pragmatic decision-making, stakeholder alignment, ability to articulate trade-offs clearly, and evidence of choosing the right level of sophistication for the problem rather than defaulting to the most complex approach.

Assess cross-functional collaboration skills, ability to speak the language of engineering (APIs, schemas, latency), handling differing priorities, and experience with CI/CD, code reviews, and production handoff.

Look for evidence of continuous learning habits (papers, conferences, communities like Kaggle or MLOps.Community), ability to critically evaluate new methods, and concrete examples of adopting or rejecting new approaches based on evidence.