Sign in
T
supervised-learningAcademicOpenSourceVERIFIED

Time-series forecasting

by N/A – general methodological area, not a single vendor

Research Paper

Time-series forecasting is a family of statistical and machine-learning techniques used to predict future values of a variable based on its historical, time-ordered data. It matters because many real-world processes—such as demand, prices, sensor readings, and traffic—are inherently temporal, and accurate forecasts enable better planning, optimization, and risk management across industries.

Key Features

  • •Models temporal dependence and seasonality (e.g., daily, weekly, yearly patterns)
  • •Supports a range of methods from classical statistics (ARIMA, ETS) to modern ML and deep learning (gradient boosting, RNNs, Transformers)
  • •Handles multivariate inputs, exogenous variables, and hierarchical time series
  • •Can incorporate uncertainty estimates and prediction intervals, not just point forecasts
  • •Often includes automatic model selection, hyperparameter tuning, and feature engineering for lags and calendar effects
  • •Scales from single series to large collections of related time series (e.g., thousands of SKUs)
  • •Integrates with MLOps workflows for retraining, monitoring, and deployment as APIs or batch jobs

Use Cases

  • •Demand forecasting for retail, e-commerce, and supply chain planning
  • •Energy load and renewable generation forecasting for utilities and grid operators
  • •Financial market and risk forecasting (prices, volatility, credit risk)
  • •Predictive maintenance using sensor and telemetry time series from industrial equipment
  • •Forecasting web traffic, user activity, and capacity needs for digital services
  • •Healthcare and epidemiological forecasting (patient volumes, disease incidence)
  • •Transportation and mobility forecasting (traffic flows, ride-hailing demand, public transit usage)

Adoption

Market Stage
Early Majority

Used By

AmazonWalmartUberGoogleMicrosoftMetaJP MorganSiemensGE

Performance Benchmarks

M4 forecasting competition
Multiple methods evaluated; top methods combined statistical and ML approaches
2018-05
M5 forecasting competition
Best-performing methods used gradient boosting and hierarchical reconciliation for retail demand
2020-06
M6 forecasting competition
Focuses on financial time-series forecasting with probabilistic evaluation metrics
2022-07

Alternatives

Classical statistical forecasting (ARIMA/ETS-only workflows)

Forecasting methods

Relies purely on traditional time-series models without modern ML or deep learning extensions.

Well-understood theory and interpretabilityStrong performance on many univariate, low-noise series

Rule-based or heuristic forecasting (e.g., moving averages, seasonal naïve)

Forecasting methods

Uses simple rules or extrapolations rather than fitted statistical/ML models.

Very simple to implement and explainLow computational cost

Optimization-based planning without explicit forecasting

Planning and optimization

Uses scenario analysis or robust optimization directly on historical data instead of explicit forecasts.

Can be robust to forecast errors in some formulationsFocuses directly on decision quality

Industries

Retail and e-commerceSupply chain and logisticsEnergy and utilitiesFinancial servicesManufacturing and industrial IoTTechnology and internet servicesHealthcare and life sciencesTransportation and mobilityTelecommunications