A practical, verified guide to 20 forecasting models spanning classical statistics, machine learning, deep learning, and specialized techniques — each paired with a hand-checked tutorial or reference.
Time-tested statistical methods that model trend, level, and seasonality directly. Fast to fit, easy to interpret, and strong baselines for most forecasting problems.
Smooths short-term fluctuations by averaging past observations over a fixed window. Best for simple, low-noise series without a strong trend or seasonality.
Read tutorial →Weights recent observations more heavily than older ones using exponentially decaying weights. Suited to data with no clear trend or seasonal pattern.
Read tutorial →Extends simple exponential smoothing with a trend component, allowing it to capture data that is steadily increasing or decreasing over time.
Read tutorial →Adds a seasonal component on top of level and trend, making it well suited to data with repeating seasonal cycles, such as monthly sales.
Read tutorial →Combines autoregression, differencing, and moving averages to model non-seasonal time series that have trend but no repeating seasonal pattern.
Read tutorial →Seasonal ARIMA extends the ARIMA framework with seasonal autoregressive and moving-average terms, handling series with both trend and recurring seasonality.
Read tutorial →Automates the selection of ARIMA/SARIMA orders through stepwise search and information criteria, removing the need for manual parameter tuning.
Read tutorial →An additive model that decomposes a series into trend, seasonality, and holiday effects. Robust to missing data and outliers, and popular for business forecasting.
Read tutorial →Tree-based ensemble methods that treat forecasting as a supervised regression problem over lagged and engineered features, capturing nonlinear patterns and exogenous drivers.
Applies gradient-boosted decision trees to time series using lag and window feature engineering. Handles nonlinear relationships and multiple exogenous variables well.
Read tutorial →An ensemble of decision trees that averages predictions to reduce variance. Robust to noise and overfitting when temporal structure is captured through engineered features.
Read tutorial →A fast, memory-efficient gradient boosting framework built for large-scale time series with many features, offering quicker training than traditional boosting methods.
Read tutorial →Neural network architectures that learn temporal representations directly from raw sequences, excelling on large datasets with complex, long-range dependencies.
Long Short-Term Memory recurrent neural networks use gated memory cells to capture long-range temporal dependencies, widely used for complex sequential data.
Read tutorial →Gated Recurrent Units offer a simplified, faster alternative to LSTM with fewer parameters, often achieving comparable performance on many sequence tasks.
Read tutorial →An attention-based deep learning architecture that combines interpretability with high accuracy across multiple forecast horizons and covariates.
Read tutorial →A pure deep learning architecture built on backward and forward residual links, achieving strong accuracy without requiring hand-crafted features.
Read tutorial →Amazon's autoregressive recurrent network produces probabilistic forecasts by learning shared patterns across many related time series simultaneously.
Read tutorial →Specialized methods built for particular data shapes: intermittent demand, multivariate interdependence, and complex or multiple seasonal cycles.
Specialized for intermittent demand series with many zero values, separately modeling the size and interval of nonzero demand occurrences.
Read tutorial →Decomposes a series into 'theta lines' that adjust the curvature of the trend. Simple to compute yet consistently competitive in forecasting competitions.
Read tutorial →Models multiple interdependent time series jointly, capturing linear relationships and feedback effects among the variables.
Read tutorial →Handles complex seasonality — multiple, non-integer, or high-frequency cycles — using trigonometric terms, Box-Cox transforms, ARMA errors, and trend/seasonal components.
Read tutorial →