Time-series data is everywhere, tracking everything from carbon emissions and traffic flows to disease outbreaks and financial market fluctuations. To forecast the future from the past, models must untangle a complex web of long-term trends, random shocks, and cyclical patterns. The Autoregressive Integrated Moving Average (ARIMA) framework, along with its seasonal extension (SARIMA), provides a structured mathematical approach to decompose these temporal dynamics into predictable components.
Achieving Stationarity through Differencing
Most classical statistical models operate on the assumption of stationarity—the requirement that the data's mean, variance, and autocorrelation remain constant over time. Real-world data, however, is rarely stationary; it climbs, drops, or drifts due to long-term growth or systemic shifts. As noted by Wang et al., accounting for this non-stationarity is essential for projecting long-term trajectories accurately.
To resolve this, the Integrated (I) component of ARIMA applies differencing. Differencing subtracts the value of the current observation from the previous one ($Y_t - Y_{t-1}$). If a time series has a constant upward trend, taking the first difference strips away this growth, transforming a drifting signal into a stable, flat baseline. Teoh et al. highlight that pairing formal stationarity testing with information criteria is a critical step to ensure the model selects the most robust fit for longitudinal trends.
Capturing History and Shocks
Once differencing has stabilized the series, the model addresses the remaining signal using two distinct, complementary mechanisms: Autoregression (AR) and Moving Average (MA).
Autoregression (AR)
The AR component assumes that the future depends linearly on the past. For instance, if traffic volume or energy consumption was high yesterday, it is statistically likely to remain elevated today. Aman et al. and Tao et al. demonstrate how capturing these temporal dependencies allows models to handle sequential data effectively, whether predicting building energy demand or urban traffic flow.
Moving Average (MA)
The MA component accounts for sudden, unpredictable shocks—such as extreme weather events, policy changes, or sudden market shifts—by regressing the future value on past forecast errors. Instead of looking at the raw values of the past, it looks at how wrong the previous predictions were. If an unexpected event causes a sudden spike, the MA component helps the model smooth out the noise and adjust its subsequent predictions.
By combining these components, ARIMA builds a comprehensive picture of sequential data patterns. This combined approach has proven effective across diverse domains, including nursing resource allocation (Du et al.), carbon emission projections (Sun et al.), ecosystem dynamics (Bo et al.), and macroeconomic planning (Li et al.).
Modeling Periodic Cycles
Many real-world phenomena do not just drift or fluctuate randomly; they repeat. Solar irradiance changes with the seasons, exchange rates follow annual market cycles, and helpline volumes fluctuate predictably by the day or week. Standard ARIMA treats these periodic fluctuations as noise, but Seasonal ARIMA (SARIMA) explicitly incorporates them by adding seasonal parameters that mirror the core AR, I, and MA components at fixed intervals.
By identifying specific seasonal configurations, the model isolates predictable temporal patterns from stochastic noise (Wang et al.). For instance, de Boer et al. use seasonal models to capture recurring daily and weekly patterns in mental health helpline demand. On a larger scale, Minghui Liao applies a SARIMA model with an annual period to capture recurring yearly fluctuations in financial exchange rates, enhancing predictive power by accounting for cyclical trends that non-seasonal frameworks ignore.
This capacity to adjust for periodic fluctuations is equally vital for environmental and public health forecasting:
- Environmental Cycles: Bao et al. utilize SARIMA to forecast solar irradiance by modeling long-term seasonal patterns alongside short-term dependencies. Similarly, Huang et al. integrate quarterly variations into their framework to capture periodic fluctuations in environmental radioactivity levels influenced by agricultural cycles.
- Epidemiology and Commerce: In public health, Zhang et al. demonstrate how accounting for seasonality improves near real-time disease surveillance. In global trade, Liao et al. project long-term trajectories while adjusting for periodic market cycles. For commercial applications, Bin Liu leverages SARIMA's ability to capture recurring patterns over time to optimize replenishment quantities for perishable goods in dynamic supermarket environments.
Establishing Baselines and Counterfactuals
Beyond pure forecasting, the strength of the (S)ARIMA framework lies in its ability to establish a robust statistical baseline of what "normal" behavior looks like based on historical patterns. This baseline can then be used for anomaly detection or counterfactual analysis.
Per Sieverts Nielsen leverages this by combining a periodic auto-regression model with real-time streaming data to perform online anomaly detection in smart grids, flagging consumption deviations from historical baselines. Conversely, de Villiers et al. use ARIMA models to project what healthcare metrics would have been in the absence of a specific disruption, allowing researchers to isolate and quantify the collateral damage of the COVID-19 pandemic on global tuberculosis care.
Key Foundational Work
Prediction Models for Dynamic Demand Response (2015) — contributed 1 concept elements to this direction.
Regression-based Online Anomaly Detection for Smart Grid Data. (2016) — contributed 1 concept elements to this direction.