Introduction to Time Series Forecasting: Concepts, Applications, and Evolution

When a commuter opens a navigation application on a weekday morning and receives a predicted travel time of 27 minutes to their office, or when a grid operator adjusts power generation schedules based on a forecast of peak electricity demand for the upcoming afternoon, they are interacting with outputs from time series forecasting systems. Time series forecasting, at its core, is the task of predicting future values of a sequence of observations collected over time, leveraging the inherent temporal structures, patterns, and dependencies embedded in historical data. Unlike cross-sectional predictive modeling, which treats observations as independent and identically distributed (i.i.d.) samples, time series forecasting explicitly accounts for the temporal ordering of data, recognizing that past values often contain critical information about future outcomes. This chapter introduces the foundational concepts, cross-domain applications, historical evolution, and structural roadmap of this book, providing a holistic entry point for readers to navigate the field of temporal forecasting.

Core Definition and Distinctive Characteristics of Time Series Forecasting

A time series is formally defined as a sequence of observations $\{y_t\}_{t=1}^T$ collected at regular or irregular time intervals, where $y_t$ denotes the value of the target variable at time $t$, and $T$ is the total number of historical observations. The primary objective of time series forecasting is to learn a mapping function that maps historical observations (and optional exogenous features) to future values of the target variable. Formally, let $Y_t = \{y_1, y_2, ..., y_t\}$ denote the historical target values up to time $t$, and let $\mathbf{X}_t = \{\mathbf{x}_1, \mathbf{x}_2, ..., \mathbf{x}_t\}$ denote a set of $K$ exogenous features at each time step, where $\mathbf{x}_t \in \mathbb{R}^K$ may include variables such as temperature, holiday flags, or macroeconomic indicators that are believed to influence the target variable. For a forecasting horizon $H$, the goal is to predict the future values $\{y_{t+1}, y_{t+2}, ..., y_{t+H}\}$ as: $$\hat{y}_{t+1:t+H} = f(Y_t, \mathbf{X}_t; \theta)$$ where $f(\cdot)$ is the forecasting model, and $\theta$ denotes its learnable parameters.

Time series forecasting is distinct from other predictive modeling tasks in three key aspects: 1. Temporal dependency: Observations in a time series are not independent, as the value at time $t$ is often correlated with values at previous time steps $t-1, t-2, ...$. This violates the i.i.d. assumption that underpins most traditional machine learning models, requiring specialized training and validation strategies that respect temporal ordering. 2. Structural patterns: Most real-world time series exhibit systematic patterns such as trend (long-term upward or downward movement), seasonality (regular periodic fluctuations, e.g., daily, weekly, or annual cycles), and autocorrelation (correlation between a value and its own past values), which must be explicitly modeled to achieve accurate forecasts. 3. Temporal alignment: The output of a forecasting task is aligned with future time steps, meaning predictions are evaluated based on their accuracy at specific future time points, rather than on arbitrary held-out samples.

The core objectives of time series forecasting vary across use cases, but can be categorized into three primary output types:

• **Point forecasting**: The most common form, which outputs a single predicted value for each future time step. For example, a point forecast for tomorrow’s electricity load might be 42.7 GWh.
• **Interval forecasting**: Outputs a range of plausible values for each future time step, accompanied by a confidence level. For example, a 95% prediction interval for tomorrow’s electricity load might be [40.2 GWh, 45.1 GWh], indicating that we are 95% confident the true value will fall within this range.
• **Probabilistic forecasting**: Outputs a full probability distribution over possible future values, rather than a single point or interval. This is particularly valuable for risk-sensitive applications such as financial risk management or disaster preparedness, where understanding the likelihood of extreme outcomes is critical.

Similarly, forecasting tasks are often categorized by their horizon length, as the difficulty and required modeling approach vary significantly across different time scales:

• **Short-term forecasting**: Horizon $H$ ranges from 1 to several time steps (e.g., 1 hour to 1 day for traffic flow or electricity load forecasting). Short-term forecasts typically prioritize high precision and rely heavily on recent historical patterns.
• **Medium-term forecasting**: Horizon $H$ ranges from several days to several weeks (e.g., 1 week to 1 month for retail sales or renewable energy output forecasting). Medium-term forecasts need to capture weekly or monthly seasonal patterns, as well as short-term trends.
• **Long-term forecasting**: Horizon $H$ ranges from several months to years (e.g., 1 year to 5 years for macroeconomic forecasting or urban planning). Long-term forecasts prioritize capturing macro-level trends and structural changes, rather than short-term fluctuations.

Cross-Domain Application Scenarios

Time series forecasting is a foundational capability across nearly all industries, with high-impact applications that drive economic value, improve public welfare, and reduce operational risk. Below we survey four representative domains with concrete real-world use cases and quantitative impact data.

Financial Asset Pricing and Risk Management

Financial markets generate massive volumes of time-stamped data, including asset prices, trading volumes, order book data, and macroeconomic indicators, making forecasting a core capability for financial institutions. Key applications include:

• **Asset price and volatility forecasting**: Predicting future stock prices, exchange rates, or commodity prices is critical for investment decision-making. For example, the Generalized Autoregressive Conditional Heteroskedasticity (GARCH) model, a classical statistical method for volatility forecasting, has been the industry standard for risk management for over 40 years. A 2022 study by the Bank for International Settlements found that 78% of global systemically important banks use GARCH-based models for calculating risk capital requirements.
• **High-frequency trading**: In electronic trading, forecasting models predict order book dynamics and price movements at microsecond timescales. For example, Virtu Financial, one of the world’s largest high-frequency trading firms, reported in its 2023 annual

When a commuter opens a navigation application on a weekday morning and receives a predicted travel time of 27 minutes to their office, or when a grid operator adjusts power generation schedules based on a forecast of peak electricity demand for the upcoming afternoon, they are interacting with outputs from time series forecasting systems. Time series forecasting, at its core, is the task of predicting future values of a sequence of observations collected over time, leveraging the inherent temporal structures, patterns, and dependencies embedded in historical data. Unlike cross-sectional predictive modeling, which treats observations as independent and identically distributed (i.i.d.) samples, time series forecasting explicitly accounts for the temporal ordering of data, recognizing that past values often contain critical information about future outcomes. This chapter introduces the foundational concepts, cross-domain applications, historical evolution, and structural roadmap of this book, providing a holistic entry point for readers to navigate the field of temporal forecasting.

Core Definition and Distinctive Characteristics of Time Series Forecasting

A time series is formally defined as a sequence of observations $\{y_t\}_{t=1}^T$ collected at regular or