Data Scientist: Demand Forecasting

Apply simple exponential smoothing and explain its exponentially decaying weight structure.

Extend exponential smoothing to Holt's linear trend and Holt-Winters seasonal methods.

Express ETS models in state-space form and navigate the error/trend/seasonality (ETS) taxonomy.

Identify candidate AR, MA, and ARMA orders using ACF/PACF inspection and information criteria (AIC/BIC).

Specify and estimate AR, MA, and ARMA models on a real demand series.

Extend ARMA to ARIMA by applying differencing to handle non-stationarity.

Incorporate seasonality with SARIMA and add exogenous regressors via SARIMAX and dynamic regression.

Validate model residuals using the Ljung-Box test and diagnostic plots, and discuss when to prefer ETS vs ARIMA.

Compute and interpret point forecast metrics: MAE, RMSE, MAPE, SMAPE, MASE, and WAPE.

Identify pitfalls of common metrics like asymmetry, scale dependence, and breakdown on intermittent demand.

Design time-series cross-validation strategies: rolling-origin, expanding-window, and multi-horizon evaluation.

Statistically compare forecast accuracy with the Diebold-Mariano test and model confidence sets.

Construct weighted, business-aware metrics aligned with revenue, stock-out, and overstock costs.

Translate forecast errors into downstream business KPIs such as service level and inventory cost.

Audit forecasts for bias, calibration, and systematic errors across product and time segments.

Construct lag, rolling-window, and expanding-window features without introducing target leakage.

Encode calendar effects, holidays, promotional events, and Fourier seasonal terms as features.

Reframe forecasting as supervised regression using direct, recursive, and multi-output strategies.

Train a global gradient boosting forecaster (e.g., LightGBM) across many series with categorical IDs.

Compare local vs. global modeling in terms of bias-variance, data efficiency, and cold-start behavior.

Apply Facebook Prophet for trend, seasonality, and holiday decomposition and articulate its limitations.

Choose appropriate target transformations (e.g., log, mean-scaling, asinh) for skewed count demand data.

Implement a feedforward MLP forecaster in PyTorch using engineered lag features.

Explain how RNNs share parameters across time and why vanishing/exploding gradients arise.

Implement an LSTM or GRU encoder for univariate forecasting in PyTorch.

Describe the seq2seq encoder-decoder framework for multi-step forecasting.

Explain scaled dot-product attention and multi-head self-attention.

Describe the transformer architecture and contrast it with recurrent models.

Implement per-series scaling and windowed batching to prepare time series data for deep learning models.

Apply teacher forcing as a training strategy for autoregressive sequence models and explain its trade-offs.

Describe DeepAR's autoregressive likelihood-based approach to probabilistic forecasting.

Explain the N-BEATS architecture with doubly residual stacking and interpretable trend/seasonality blocks.

Describe NHITS and how multi-rate hierarchical interpolation improves long-horizon forecasts.

Describe the Temporal Fusion Transformer's high-level architecture, including variable selection networks and gating.

Explain how vanilla transformers can underperform on long-horizon forecasting and how patching (PatchTST) addresses this.

Describe time-series foundation models (e.g., Chronos, TimesFM) and when zero-shot forecasting is viable.

Use lessons from the M5 and M6 forecasting competitions to choose between classical, GBM, and neural methods for a given problem.

Motivate probabilistic forecasts for inventory, pricing, and service-level decisions in e-commerce.

Estimate prediction intervals for classical models (ETS, ARIMA) using their analytical formulas and assumptions.

Generate prediction intervals for any point forecaster via the residual bootstrap method.

Simulate future sample paths from a probabilistic model via Monte Carlo to derive prediction intervals.

Implement quantile regression and explain the pinball (quantile) loss function.

Model the full predictive distribution with parametric output heads (e.g., Gaussian, Negative Binomial).

Apply conformal prediction to construct distribution-free prediction intervals for any forecaster.

Evaluate probabilistic forecasts using CRPS, pinball loss, coverage, and calibration plots.

Define hierarchical and grouped time series structures (e.g., country > category > brand > article).

Apply bottom-up, top-down, and middle-out reconciliation and discuss their bias-variance trade-offs.

Explain the optimal reconciliation (MinT) method and the derivation of its weights via Generalized Least Squares.

Implement MinT reconciliation on a hierarchical dataset using a standard forecasting library.

Reconcile probabilistic forecasts across a hierarchy while preserving quantile consistency.

Define cross-temporal hierarchies and explain why temporal consistency (e.g., daily summing to weekly) matters for planning.

Identify practical challenges of hierarchical forecasting at large SKU/article scale, such as computational cost and matrix sparsity.

Characterize intermittent, lumpy, and slow-moving demand patterns common to fashion catalogues.

Apply Croston's method and its variants (e.g., SBA, TSB) for intermittent demand.

Model intermittent demand with Poisson, Negative Binomial, and zero-inflated count models to handle excess zeros.

Distinguish observed sales from true (uncensored) demand and apply a basic uncensoring technique to correct for stock-outs.

Forecast new products with no sales history using attribute-based regression and nearest-neighbor borrowing from similar articles.

Represent product life cycle stages (launch, growth, maturity, end-of-life) as features in fashion demand models.

Explain the potential outcomes framework and distinguish prediction, correlation, and causation in a demand-forecasting context.

Define confounding and explain how it invalidates naive comparisons of forecasting models on observational data.

Apply regression adjustment to control for low-dimensional confounders (e.g., price, promotion flags).

Design and analyze a standard A/B test to evaluate the rollout of a new forecasting model on both accuracy and business metrics.

Apply difference-in-differences to estimate the effect of an intervention (e.g., a marketing campaign) on demand.

Contrast Average vs. Heterogeneous Treatment Effects and motivate uplift modeling for demand drivers.

Implement S-learner and T-learner uplift estimators and discuss their respective limitations.

Train thousands of forecasting models in parallel using PySpark's `applyInPandas` and pandas UDFs.

Define training/serving skew and illustrate how it manifests with time-dependent features like rolling aggregates.

Apply feature store patterns with point-in-time correctness for time-series features to avoid training/serving skew.

Orchestrate forecasting pipelines on a platform like Databricks with Delta Lake versioning.

Track experiments and register forecasting models with MLflow, including custom evaluation metrics.

Design a monitoring system for production forecasts to detect data drift, concept drift, and segment-level error degradation.

Compare and contrast retraining strategies: scheduled, performance-triggered, and incremental/online updates.

Identify fashion-specific demand drivers: micro-trends, weather, marketing, seasonality, and returns.

Model the price elasticity of demand and the causal uplift of promotions to inform pricing and marketing strategies.

Incorporate image and text embeddings of articles as features in demand models to improve new product forecasting.

Model the product return rate and combine it with gross sales forecasts to derive net demand and revenue forecasts.

Analyze a real-world forecasting system from a company tech blog (e.g., Zalando, Uber) in terms of its architecture and trade-offs.

Translate a business problem (e.g., reducing stockouts) into a precise forecasting task with a clear target, horizon, granularity, and justified evaluation metric.

Structure and present a project proposal for a new forecasting system, from business case to model selection and productionization.

Communicate forecast uncertainty and method trade-offs to non-technical stakeholders (e.g., PMs, buyers) using clear language and visualizations.