Forecasting

Foundational text establishing ARIMA and the Box-Jenkins approach to model identification.

Establishes ETS state space framework enabling model selection, prediction intervals, and likelihood calculation.

Landmark 3,003-series competition establishing that simpler methods often outperform complex ones.

100,000 series showing hybrid statistical-ML methods dominate; pure ML underperformed.

Foundational decomposition method separating seasonal, trend, and remainder components using local regression.

Extension of STL handling multiple seasonal periods; essential for complex seasonality like hourly data.

Definitive treatment of proper scoring rules; introduces CRPS, energy score, interval score.

Introduced quantile regression for estimating conditional quantiles; foundation for distributional forecasting.

GEFCom2014 establishing practical standards for probabilistic forecast evaluation.

Distribution-free prediction intervals with guaranteed coverage for time series; handles non-exchangeability.

Normalizing flows for flexible multivariate distributions in forecasting; captures complex dependencies.

Seminal paper showing combined forecasts yield lower MSE than individuals; foundation for ensemble methods.

MinT optimal reconciliation method minimizing forecast variance; outperforms bottom-up and top-down.

Authoritative review explaining why simple averages often beat optimal weights (estimation error, instability).

Comprehensive review covering cross-sectional, temporal, and cross-temporal reconciliation.

Unified framework for simultaneous cross-sectional and temporal reconciliation; optimal point forecasts.

Amazon's global probabilistic forecasting model; pioneered cross-learning for scale-diverse series.

Meta's additive model with trend, seasonality, holidays; designed for analyst-in-the-loop forecasting.

First pure DL model to beat M4 winner; interpretable version decomposes into trend and seasonality.

Google's attention architecture handling static covariates, known future inputs, observed past-only features.

42,840 Walmart series; LightGBM and DL dominated; 22% improvement over best benchmark.

Combines state space models with deep learning; enables interpretable components with neural flexibility.

AAAI Best Paper; ProbSparse attention achieving O(L log L) complexity for long sequences.

Novel auto-correlation mechanism replacing attention; built-in series decomposition improves interpretability.

Fourier-enhanced attention capturing global patterns in frequency domain with linear complexity.

Patching time series like images; channel-independence achieving SOTA on long-term benchmarks.

Google's MLP-Mixer adaptation showing simple MLPs can match or beat transformers on forecasting.

IBM's TSMixer variant with cross-variate mixing; strong M5 performance with fewer parameters.

Amazon's tokenized time series foundation model; zero-shot forecasting matching fine-tuned models.

First commercial time series foundation model; API-based zero-shot forecasting for practitioners.

Open-source decoder-only foundation model for probabilistic forecasting; strong zero-shot transfer.

CMU's open foundation model supporting forecasting, classification, anomaly detection, and imputation.

Google's 200M parameter foundation model trained on 100B time points; strong zero-shot performance.

Salesforce's multi-patch model handling variable frequencies and prediction lengths in single model.

Repurposing frozen LLMs for forecasting via prompt reprogramming; no time series pre-training needed.

Foundational method separating demand size from demand occurrence; basis for spare parts forecasting.

SBA method correcting Croston's bias; widely adopted in supply chain software.

TSB method explicitly modeling demand probability; improved coverage for slow-moving items.

Amazon's open-source toolkit; includes intermittent demand models and extensive benchmarking.

Elegant Bayesian framework for online change point detection; foundational for streaming applications.

Microsoft's production system using spectral residual and CNN; handles millions of time series.

Amazon's streaming anomaly detection algorithm; efficient updates and interpretable anomaly scores.

VAE-based approach for multivariate anomaly detection; captures temporal dependencies and normal patterns.

Google's CausalImpact for estimating causal effects from observational time series; widely used for attribution.

Foundational synthetic control paper; 'most important innovation in evaluation literature in 15 years'.

Comprehensive methodological review covering feasibility, inference, and best practices.

Google's BSTS framework for nowcasting with spike-and-slab regression for variable selection.

Modern DiD handling staggered treatment timing and heterogeneous effects; doubly-robust estimator.

Comprehensive synthesis of modern DiD developments; essential guide to recent methodological advances.

Uber's Bayesian forecasting framework combining ETS with global models; production-ready with uncertainty.

Meta's production system enabling analysts to create reliable forecasts with domain knowledge.

Large-scale demand forecasting system handling billions of SKUs; hierarchical and cross-learning approaches.

Practical lessons from production ML including forecasting; emphasizes offline-online gaps and iteration.

The Diebold-Mariano test for forecast comparison; essential for model selection in production.