Fact-checked by Grok 2 weeks ago

Time series

A time series is a sequence of data points collected or recorded at successive points in time, often at regular intervals, representing the evolution of a variable over time. These data exhibit temporal dependence, where the value at one time point is typically correlated with values at previous or subsequent points, distinguishing them from independent observations in . Time series data commonly decompose into key components: a trend capturing long-term increase or decrease, seasonality reflecting regular periodic fluctuations (e.g., daily or yearly cycles), cyclicality indicating irregular longer-term waves, and an irregular or residual component accounting for random noise. This additive or multiplicative structure helps in understanding underlying patterns, with methods like moving averages used to isolate the trend-cycle from seasonal effects. For instance, monthly retail sales might show a yearly seasonal peak during holidays atop an upward economic trend. Analysis of time series focuses on modeling these dependencies to forecast future values or infer past behaviors, employing techniques such as functions to detect patterns and stationarity tests to ensure model applicability. Common models include (ARIMA) for univariate series, which capture short-term dependencies, and for trend and seasonal forecasting. More advanced approaches incorporate , like neural networks, for complex multivariate series. Time series arise across disciplines, including (e.g., GDP ), (stock prices), (temperature records), and (disease incidence rates), enabling predictions that inform , , and . Foundational texts, such as those by Brockwell and Davis, emphasize rigorous statistical frameworks for handling non-stationarity and serial correlation in real-world applications.

Fundamentals

Definition and Characteristics

A time series is typically defined as an ordered sequence of data points representing the values of a variable measured at successive points in time, often at equally spaced intervals. These data points are typically denoted mathematically as \{X_t\}_{t=1}^n, where t indexes the time points from 1 to n, and each X_t captures the observation at time t. This structure distinguishes time series from cross-sectional data, as the ordering by time introduces inherent dependencies that must be accounted for in analysis. Key characteristics of time series data include temporal dependence, primarily through , which measures the between an observation and its lagged values, reflecting how past data influence the present. Additional features often encompass a trend, indicating a long-term increase or decrease in the data; , characterized by recurring patterns at fixed periods such as daily or annual cycles; cyclicity, involving irregular fluctuations longer than seasonal periods but without fixed timing; and , representing random, unpredictable variations. These properties can affect the stationarity of the series, a concept explored further in statistical testing. The study of time series emerged in the early as a subfield of , with pioneering contributions from Udny , who developed autoregressive methods in the 1920s, building on earlier efforts to model periodic phenomena. Its roots lie in the analysis of economic indicators and meteorological records, such as cycles and business fluctuations, where sequential data were essential for understanding dynamic patterns. Representative examples of time series include daily stock prices, which exhibit trends and influenced by market events; hourly readings from weather stations, showing diurnal and seasonal cycles; and quarterly (GDP) measurements, capturing long-term amid cyclical recessions.

Types of Time Series Data

Time series data are distinguished from cross-sectional data primarily by their sequential nature, where observations are ordered over time to capture dependencies such as , whereas collect simultaneous measurements across entities at a fixed point without temporal structure. Time series can be categorized by dimensionality into univariate and multivariate types. Univariate time series consist of observations from a single variable tracked over time, such as the daily closing price of an individual , allowing focus on its internal patterns like trends or . In contrast, multivariate time series involve multiple interrelated variables observed simultaneously, exemplified by economic indicators including GDP, rates, and , which are often modeled jointly to account for interdependencies, as in frameworks. Another key distinction lies in the timing of observations: versus continuous. Discrete time series feature data points at fixed, equally spaced intervals, such as daily readings or quarterly financial reports, facilitating straightforward indexing by integers. Continuous time series, however, use real-valued time indices without fixed spacing, commonly arising from ongoing monitoring, like data from physical experiments or environmental recordings, where the underlying process evolves smoothly over . , or cross-sectional time series, extend univariate and multivariate frameworks by combining multiple entities observed over repeated time periods, such as GDP per capita across 126 countries from 2000 to 2021, enabling analysis of both temporal dynamics and cross-entity variations. These datasets often incorporate fixed effects to control for unobserved, time-invariant entity-specific factors like cultural influences or policies, modeled as individual intercepts (α_i) correlated with predictors, or random effects assuming such factors are random and uncorrelated with explanatory variables, allowing inclusion of time-invariant covariates. Irregularly spaced time series arise when observations occur at uneven or missing intervals, posing challenges for standard models due to gaps in the time index. This type is prevalent in event-driven contexts, such as recordings of earthquakes or floods, where occurrences are unpredictable and non-periodic, necessitating or specialized processes like continuous autoregressive models to handle the irregularity. Hierarchical time series feature a nested aggregation structure, where finer-grained series sum to coarser levels, ensuring across the . A common example is retail sales data, where daily transactions for specific products (e.g., road versus mountain bikes) aggregate to monthly totals by category and then to national figures, supporting reconciled forecasting that maintains additive consistency.

Preprocessing and Exploration

Data Acquisition and Cleaning

Time series data acquisition involves collecting sequential observations over time from diverse sources, ensuring temporal alignment and consistency for subsequent analysis. Common sources include sensors in devices, which capture continuous measurements such as environmental variables or machine performance metrics. , such as time series databases with indexing, store historical records from operational systems, facilitating efficient querying of time-stamped events. APIs from financial feeds or public repositories, like NASA's portal, provide structured streams of economic indicators or satellite . For scenarios lacking real data, simulated generation methods employ generative models to produce synthetic sequences mimicking real-world patterns, aiding in testing and validation. Once acquired, cleaning addresses imperfections inherent in time series , such as gaps and anomalies that could models. values, often arising from failures or transmission errors, are handled through imputation techniques tailored to temporal dependencies. estimates absent points by drawing straight lines between neighboring observations, preserving trends in continuous series like readings. Forward fill propagates the last valid value forward, suitable for stable processes where persistence is expected. Backward fill applies the subsequent value retrospectively; both methods minimize assumptions but risk introducing in volatile . Outlier detection identifies anomalous points that deviate significantly from expected patterns, potentially due to measurement errors or . The z-score computes deviations from the mean in standard deviation units, flagging values exceeding thresholds like ±3 as outliers, which is effective for normally distributed residuals in time series. The (IQR) approach marks points outside 1.5 times the IQR from quartiles as outliers, offering robustness against non-normal distributions common in financial or sensor data. Resampling adjusts the frequency of time series to align with analysis needs, aggregating high-frequency data into coarser intervals. For instance, converting hourly stock prices to daily summaries uses mean aggregation to capture average behavior or sum for cumulative volumes like sales totals. This process ensures uniform spacing, reducing noise while retaining essential dynamics. Normalization transforms the data to enhance stationarity, stabilizing mean and variance for modeling. Differencing subtracts consecutive observations to remove trends, such as first-order differencing for linear growth in economic indicators. Logarithmic transforms compress scale for multiplicative processes, like exponential growth in populations, mitigating heteroscedasticity without altering the sequential nature. Recent advances include machine learning-based methods, such as deep adaptive input normalization, for handling complex non-stationarities in neural network applications. Final quality checks verify structural integrity before analysis. Ensuring a monotonic time index confirms observations are strictly increasing, preventing ordering errors from merged sources like panel data across entities. Handling duplicates involves detecting and resolving repeated timestamps, often by averaging values or retaining the most recent, to avoid artificial inflation in aggregated statistics.

Exploratory Data Analysis

Exploratory data analysis (EDA) in time series involves initial examinations of the data to reveal underlying structures, such as trends, dependencies, and irregularities, prior to formal modeling. This process typically assumes that basic data cleaning, such as handling missing values or outliers from acquisition, has been completed, allowing focus on pattern discovery through descriptive measures and visualizations. EDA helps practitioners understand the data's temporal dynamics, informing subsequent steps like . Summary statistics provide a quantitative foundation for EDA by capturing , , and serial dependence in time series data. The summarizes the average level of the series, while the variance quantifies its variability over time. A key measure of dependence is the lag-1 autocorrelation coefficient, defined as \rho_1 = \frac{\Cov(X_t, X_{t-1})}{\Var(X_t)}, which indicates the linear between consecutive observations and highlights short-term persistence in the data. Visual plotting forms the core of EDA, enabling intuitive detection of temporal patterns. Time series line plots display observations against time, revealing overall trajectories and potential irregularities at a glance. plots extend this by graphing the autocorrelation coefficients at various lags, visualizing the decay of dependence and aiding in the identification of periodic or persistent structures in the data. To identify specific patterns, moving averages smooth the series to isolate trends by averaging values over a fixed , such as a simple k-period average that reduces while preserving long-term direction. For seasonality, seasonal subseries plots aggregate data by (e.g., months) across multiple periods, plotting each subseries separately to highlight recurring cycles and variations in or over time. These techniques allow for the detection of non-stationarities without assuming a full model. Anomaly spotting during EDA relies on visual inspection of line plots to flag deviations like sudden spikes, which represent point outliers, or structural breaks, indicating shifts in the series' level or variance. Such inspections are essential for preliminary quality checks, as unaddressed anomalies can distort pattern recognition. The Box-Jenkins approach integrates these EDA elements into its identification stage, an iterative process where summary statistics and ACF plots guide the provisional selection of ARIMA model orders (p, d, q) by examining evidence of non-stationarity, autoregression, and moving average effects, without yet estimating parameters. This stage emphasizes empirical diagnostics to ensure the model class aligns with observed data features.

Statistical Foundations

Stationarity and Testing

In time series analysis, stationarity refers to the property where the statistical characteristics of the series remain constant over time. Strict stationarity, also known as strong stationarity, requires that the joint of any collection of observations is invariant under time shifts, meaning the probability structure does not depend on the specific time period considered. Weak stationarity, or covariance stationarity, is a less stringent condition that applies when the process has a constant mean, finite and constant variance, and that depends solely on the time lag between observations rather than their absolute positions in time. This weaker form is sufficient for many practical modeling purposes, as it focuses on the first two moments of the . Stationarity is a foundational assumption for numerous time series models, including autoregressive (ARMA) processes, which require the series to be weakly to ensure the validity of parameter estimation and . Without stationarity, analyses can produce misleading results, such as spurious regressions, where non-stationary series exhibit apparent correlations and high R-squared values despite lacking any true economic relationship, as demonstrated in early simulation studies. Achieving or verifying stationarity thus prevents invalid conclusions and enables reliable forecasting and hypothesis testing in econometric and statistical applications. To assess stationarity, several statistical tests are employed, with the Augmented Dickey-Fuller (ADF) test being one of the most widely used. The ADF test, an extension of the original Dickey-Fuller procedure, evaluates the of a —indicating non-stationarity—by fitting an augmented with lagged difference terms to control for serial correlation in the errors: \Delta y_t = \alpha + \beta t + \gamma y_{t-1} + \sum_{i=1}^{p} \delta_i \Delta y_{t-i} + \epsilon_t The test statistic for \gamma is derived from the residuals of this and compared to asymptotic critical values, rejecting the null if the series appears stationary. Complementing the , the Kwiatkowski-Phillips-Schmidt-Shin ( reverses the hypotheses, with the null assuming stationarity (or trend stationarity) and the alternative positing a ; it computes a statistic based on cumulative residuals from a deterministic trend , providing robustness against certain size distortions in tests. If a series is found to be non-stationary, transformations are applied to induce stationarity. First-order differencing removes linear trends by computing the successive differences: \Delta X_t = X_t - X_{t-1} For series with seasonal patterns, seasonal differencing targets periodicity by subtracting observations at the seasonal interval s, such as \Delta_s X_t = X_t - X_{t-s} for monthly data where s=12; higher-order or combined differencing may be needed for more complex non-stationarities. To address heteroscedasticity or variance instability, the Box-Cox transformation family is often used, defined as: X_t^{(\lambda)} = \begin{cases} \frac{X_t^\lambda - 1}{\lambda} & \lambda \neq 0 \\ \log X_t & \lambda = 0 \end{cases} where \lambda is estimated via maximum likelihood to stabilize variance while preserving the series' positivity. In multivariate settings, non-stationary series may still exhibit meaningful relationships through cointegration, where each individual series has a unit root but a linear combination forms a stationary process, capturing long-run equilibria. The Engle-Granger two-step method tests for cointegration by first regressing one series on the others to obtain residuals, then applying an ADF test to those residuals for a unit root; rejection indicates cointegration, allowing for error-correction models to model both short- and long-run dynamics.

Decomposition and Trend Analysis

Time series decomposition involves separating a observed series into its underlying components—typically trend, seasonal, and irregular—to reveal the structural patterns influencing the data. This process aids in understanding long-term movements, periodic fluctuations, and random variations, facilitating more informed analysis and modeling. The choice of decomposition model depends on the nature of the seasonal variations: additive models assume constant absolute changes across levels, while multiplicative models account for proportional changes that grow with the series magnitude. In an additive decomposition, the observed value Y_t at time t is expressed as the sum of the trend component T_t, the seasonal component S_t, and the irregular (or ) component I_t:
Y_t = T_t + S_t + I_t.
This approach is suitable when the variance of the series remains constant over time, such as in series with stable seasonal swings regardless of the trend level. Conversely, the multiplicative models the interaction as a product:
Y_t = T_t \times S_t \times I_t,
which applies to series where seasonal effects amplify with increasing trend, leading to heteroscedastic variance. These formulations, introduced in foundational time series work, allow for the isolation of components through iterative techniques.
Classical decomposition employs s to estimate the trend-cycle component, providing a simple yet effective method for non-robust fitting. For a series with seasonal period m, the trend is often approximated using a centered of order $2m+1, which smooths out seasonal irregularities by averaging symmetric windows around each point. The seasonal component is then derived by averaging the detrended residuals over multiple cycles, and the irregular component is obtained by subtracting (additive) or dividing (multiplicative) the estimated trend and seasonal from the original series. This method, formalized in the X-11 program, assumes a stable seasonal pattern and is computationally efficient for preliminary analysis. For more robust decomposition, especially with outliers or nonlinear trends, the Seasonal-Trend decomposition using (STL) procedure applies locally weighted regression () smoothers iteratively to extract components. STL begins with an initial fit for the trend over a specified window, followed by seasonal smoothing of the residuals and a final adjustment for the trend-cycle, repeating until convergence. Unlike classical methods, STL handles varying seasonal amplitudes and is adaptable to multiple seasonal periods, making it suitable for complex series like daily data. The original STL algorithm, developed for robust fitting, uses inner-loop iterations to refine seasonal estimates and outer loops for trend updates, ensuring stability in the presence of . Trend estimation within decomposition focuses on capturing the long-term direction, often via parametric or nonparametric methods. A basic approach uses , modeling the series as Y_t = \beta_0 + \beta_1 t + \epsilon_t, where t is the time index and \epsilon_t represents deviations; the fitted line provides the trend slope and intercept. For curvilinear patterns, extends this to higher degrees, such as Y_t = \beta_0 + \beta_1 t + \beta_2 t^2 + \epsilon_t, fitting flexible curves while avoiding through degree selection via criteria like AIC. The Hodrick-Prescott offers a nonparametric alternative, minimizing a that balances fit to the and of the trend:
\min_{T} \sum (Y_t - T_t)^2 + \lambda \sum (\Delta^2 T_t)^2,
where \lambda penalizes second differences in the trend; for quarterly , \lambda = 1600 is standard, effectively separating smooth growth from cyclical deviations in economic series.
Seasonal adjustment refines the decomposition by removing periodic effects to isolate the trend and irregular components, particularly valuable for economic indicators. The method, an evolution of classical techniques, integrates modeling for regression-based adjustments and the SEATS (Signal Extraction in ARIMA Time Series) algorithm for decomposition, handling trading-day and holiday effects. It applies moving averages enhanced with regressors to estimate seasonal factors, then adjusts the series multiplicatively or additively, producing diagnostics for model stability. Widely used by statistical agencies for official data like GDP or figures, X-13ARIMA-SEATS ensures adjusted series reflect underlying economic trends without seasonal distortion. Analysis of the irregular component examines the residuals after extracting trend and seasonal elements, verifying they approximate to validate the . Diagnostics include the Ljung-Box test, which assesses in residuals up to h via the statistic
Q = n(n+2) \sum_{k=1}^h \frac{\hat{\rho}_k^2}{n-k},
distributed as \chi^2_h under the null of no serial ; significant p-values indicate unmodeled structure. Additional checks involve plotting residual ACFs for lingering patterns and tests like Jarque-Bera. If residuals exhibit properties—zero mean, constant variance, and uncorrelated errors—the successfully captures the systematic components, supporting subsequent applications like .

Modeling Techniques

Linear Time Series Models

Linear time series models represent a class of parametric approaches that assume the observed process can be expressed as a of past values and errors, suitable for capturing dependencies in or transformed non-stationary data. These models form the foundation of classical time series analysis, emphasizing linear relationships to describe structures. They are particularly effective for univariate series where the goal is to model temporal dependencies through autoregressive, , or combined components. The of order p, denoted AR(p), posits that the current value X_t depends linearly on its p previous values plus a error \varepsilon_t: X_t = \phi_1 X_{t-1} + \phi_2 X_{t-2} + \cdots + \phi_p X_{t-p} + \varepsilon_t, where \{\varepsilon_t\} is a sequence of independent and identically distributed random variables with mean zero and variance \sigma^2. For stationarity, the roots of the $1 - \phi_1 z - \phi_2 z^2 - \cdots - \phi_p z^p = 0 must lie outside the unit circle; for the simple AR(1) case, this reduces to |\phi_1| < 1. This model originated in early work on periodicities in disturbed series, such as sunspot data. The of order q, MA(q), expresses X_t as a of and errors: X_t = \varepsilon_t + \theta_1 \varepsilon_{t-1} + \theta_2 \varepsilon_{t-2} + \cdots + \theta_q \varepsilon_{t-q}, which is always for finite q but requires invertibility for efficient estimation, meaning the roots of $1 + \theta_1 z + \theta_2 z^2 + \cdots + \theta_q z^q = 0 must lie outside the unit circle. The combined , ARMA(p, q), integrates both: X_t = \phi_1 X_{t-1} + \cdots + \phi_p X_{t-p} + \varepsilon_t + \theta_1 \varepsilon_{t-1} + \cdots + \theta_q \varepsilon_{t-q}, applicable to stationary processes with both autoregressive and moving average dependencies. These structures were formalized as tools for time series modeling in foundational statistical literature. To handle non-stationary series, the autoregressive integrated moving average model, ARIMA(p, d, q), extends ARMA by incorporating d differences of the series to achieve stationarity: \nabla^d X_t = (1 - B)^d X_t, where B is the backshift operator, followed by an ARMA(p, q) on the differenced series. Model identification follows the Box-Jenkins methodology, using autocorrelation function (ACF) plots to suggest MA order (decay after lag q) and partial autocorrelation function (PACF) plots for AR order (cutoff after lag p). This approach revolutionized time series forecasting by providing a systematic framework for model selection. For series with seasonality, the seasonal ARIMA, or SARIMA(p, d, q)(P, D, Q, s), augments with seasonal autoregressive (P), differencing (D), and (Q) terms at period s: \Phi_P(B^s) \phi_p(B) (1 - B^s)^D (1 - B)^d X_t = \Theta_Q(B^s) \theta_q(B) \varepsilon_t, where uppercase operators denote seasonal components. This extension captures periodic patterns in economic and environmental . Parameter estimation in these models typically employs (MLE), which maximizes the likelihood of observing the data under Gaussian errors, or conditional for initial fits. For AR models specifically, the Yule-Walker equations relate sample autocorrelations to coefficients via a solved iteratively, offering a moment-based alternative to MLE. These methods ensure consistent and asymptotically efficient estimates under standard assumptions. Variations addressing time-varying volatility include autoregressive conditional heteroskedasticity (ARCH) and generalized ARCH (GARCH) models, which extend linear frameworks by modeling as a function of past squared errors. The GARCH(p, q) process specifies \sigma_t^2 = \omega + \sum_{i=1}^p \alpha_i \varepsilon_{t-i}^2 + \sum_{j=1}^q \beta_j \sigma_{t-j}^2, capturing in financial time series.

Nonlinear and State-Space Models

Nonlinear time series models extend classical linear approaches, such as , by accommodating asymmetries, regime shifts, and complex dependencies that linear models cannot capture effectively. These models are particularly useful for phenomena like economic cycles or financial , where dynamics change based on thresholds or latent states. Threshold autoregressive (TAR) models introduce nonlinearity through regime-switching mechanisms, where the autoregressive parameters depend on whether the series exceeds a specified value. In a self-exciting TAR (SETAR) model, for example, the process switches regimes based on lagged values of the series itself, enabling the modeling of cyclical patterns in . The general form divides the process into multiple regimes, each governed by its own autoregressive equation, with estimation typically involving within regimes after threshold selection via grid search or information criteria. Time-varying autoregressive (TVAR) models further generalize this by allowing coefficients to evolve over time, capturing non-stationarities like structural breaks or gradual shifts. For instance, the autoregressive parameter \phi_t may follow a process, \phi_{t+1} = \phi_t + \nu_t, where \nu_t is , enabling to changing environments in macroeconomic series. Estimation often employs kernel smoothing or splines to flexibly model the time variation, with Bayesian methods providing posterior on the evolving parameters. State-space models represent the observed time series as a of unobserved states, offering a flexible framework for incorporating latent dynamics and measurement error. The observation equation is X_t = Z_t \alpha_t + \epsilon_t, where \epsilon_t \sim N(0, H_t), and the state equation is \alpha_{t+1} = T_t \alpha_t + \eta_t, with \eta_t \sim N(0, Q_t). The recursively estimates the states through prediction and updating steps, providing filtered and smoothed estimates for and in linear Gaussian settings. Among nonlinear extensions, bilinear models introduce interactions between past observations and shocks, generalizing linear autoregressions while remaining relatively tractable. The model takes the form X_t = \sum_{i=1}^p \phi_i X_{t-i} + \sum_{j=1}^q \sum_{k=1}^r \beta_{jk} X_{t-j} \epsilon_{t-k} + \epsilon_t, capturing quadratic dependencies useful for short-term nonlinearities in economic time series. More recently, (Neural ODEs) model continuous-time dynamics via \frac{dh(t)}{dt} = f(h(t), t; \theta), where f is a , offering interpretable representations for irregularly sampled time series. For modeling conditional heteroskedasticity, generalized autoregressive conditional heteroskedasticity (GARCH) models extend by incorporating lagged conditional variances into the variance equation. The standard GARCH(1,1) specification is \sigma_t^2 = \alpha_0 + \alpha_1 \epsilon_{t-1}^2 + \beta_1 \sigma_{t-1}^2, which parsimoniously captures in financial returns with parameters often satisfying \alpha_1 + \beta_1 \approx 1 for persistence. Bayesian approaches enhance state-space models by enabling full posterior inference, particularly for non-Gaussian states via (MCMC) methods. In these frameworks, MCMC samples from the joint posterior of states and parameters, using techniques like the forward-filtering backward-sampling algorithm to handle latent variables efficiently in nonlinear or switching regimes. This allows incorporation of prior information and , improving robustness in applications like macroeconomic forecasting.

Forecasting and Prediction

Forecasting Methods

Forecasting methods in time series analysis aim to extrapolate future values based on historical patterns observed in the data. These techniques integrate fitted models to generate predictions, often distinguishing between point estimates for specific future points and interval estimates that quantify uncertainty. Classical approaches rely on statistical models like , while more recent developments incorporate elements to handle complex structures such as and trends. Point forecasting produces a single predicted value for future observations, typically expressed as the conditional expectation \hat{Y}_{t+h|t} = E[Y_{t+h} \mid Y_1, \dots, Y_t], where h denotes the forecast horizon. In the Box-Jenkins methodology, ARIMA models are fitted to stationary data after differencing, enabling multi-step ahead predictions by iteratively applying the model's autoregressive and moving average components. This approach, introduced in the seminal work by Box and Jenkins, systematically identifies model orders through autocorrelation analysis and refines forecasts via maximum likelihood estimation. Interval forecasting extends point estimates by constructing prediction intervals that capture the range within which future values are likely to fall, often at a specified level like 95%. These intervals can be derived assuming asymptotic of forecast errors, where the interval is \hat{Y}_{t+h|t} \pm z_{\alpha/2} \sqrt{\hat{\sigma}^2_h}, with z_{\alpha/2} from the standard and \hat{\sigma}^2_h the estimated variance of the h-step error. Alternatively, -based methods, such as residuals from the fitted model, generate empirical distributions to form non-parametric intervals, providing robustness when assumptions fail. Empirical studies confirm that such intervals maintain coverage probabilities close to nominal levels for nearly non-stationary series. Exponential smoothing methods offer a simple yet effective way to forecast by weighted averaging of past observations, with weights decaying exponentially. The Holt-Winters extension, or triple exponential smoothing, accounts for level, trend, and seasonality using three parameters: \alpha for the level smoothing, \beta for the trend, and \gamma for the seasonal component. In the additive form, the one-step forecast is \hat{Y}_{t+1|t} = \ell_t + b_t + s_{t+1-m}, where \ell_t is the level, b_t the trend, and s_{t+1-m} the seasonal factor for period m. Originally proposed by Winters in , this method excels in capturing periodic patterns without requiring differencing, making it suitable for short-term predictions in seasonal data. Hybrid methods combine statistical decompositions with to improve flexibility, particularly for data with external regressors. Facebook's model employs an additive structure y(t) = g(t) + s(t) + h(t) + \epsilon_t, where g(t) fits piecewise linear or logistic trends, s(t) uses for , and h(t) incorporates holiday effects, while allowing additional covariates. Developed by Taylor and Letham in 2018, automates parameter tuning via for , enabling robust forecasts at scale for business time series with irregularities like outliers. In recent advancements as of 2025, transformer-based models like Informer address challenges in long-sequence forecasting by mitigating the of standard self-attention. Informer introduces ProbSparse self-attention to focus on dominant query-key pairs, reducing time and to O(L \log L), and a generative decoder for parallel long-output prediction. Proposed by Zhou et al. in 2021, it outperforms prior baselines on benchmarks like load datasets, achieving approximately 60% lower for 720-step horizons by better capturing long-range dependencies. More contemporary approaches include decoder-only foundation models, such as TimesFM developed by in 2024, and (LLM)-empowered methods like TimeCMA from 2025, which leverage pre-trained models for scalable multivariate forecasting. Forecast accuracy generally decays with increasing horizon due to error accumulation in recursive predictions, where multi-step errors compound from one-step approximations. Short-term forecasts (e.g., 1-7 steps) maintain high by leveraging recent , while long-term ones (e.g., beyond 24 steps) suffer greater uncertainty, often requiring into trend and seasonal components for stabilization. This horizon-dependent degradation underscores the need for tailored to prediction length, as validated in multi-horizon evaluations.

Model Evaluation and Selection

Model evaluation in time series analysis involves assessing how well a fitted model captures the underlying patterns in the data while ensuring generalizability to unseen observations. In-sample fit measures, such as the (AIC) and (BIC), balance model likelihood against complexity to select parsimonious models. The AIC is defined as \text{AIC} = -2 \log L + 2k, where L is the maximized likelihood and k is the number of parameters, providing an estimate of relative model quality by penalizing excessive parameters to approximate out-of-sample prediction error. The BIC, given by \text{BIC} = -2 \log L + k \log n with n as the sample size, imposes a stronger penalty on complexity, favoring simpler models especially in large samples and deriving from Bayesian principles. Out-of-sample validation is essential in time series to prevent lookahead bias, where future information contaminates training. Hold-out sets divide data chronologically, training on earlier periods and testing on later ones to mimic real scenarios. Time-series cross-validation adapts traditional methods by using expanding or rolling windows, ensuring no future data leaks into training folds and providing robust estimates of predictive performance across multiple horizons. Residual diagnostics verify model adequacy by checking if errors resemble . The Ljung-Box test assesses in residuals, with the statistic Q = n(n+2) \sum_{k=1}^h \frac{\rho_k^2}{n-k}, where n is the sample size, h the number of lags, and \rho_k the sample at k; under the null of no serial , Q follows a with h , aiding detection of misspecified dynamics. Forecasting accuracy relies on error metrics like (MAE), Root Mean Squared Error (RMSE), and (MAPE), which quantify deviations between predictions and actual values. MAE averages absolute errors for interpretability in original units, RMSE penalizes larger errors via squaring to emphasize outliers, and MAPE expresses errors as percentages for scale-independent comparisons, though it can be unstable near zero. The Diebold-Mariano test compares predictive accuracy between models by testing if the difference in their loss functions is zero, using a robust to serial correlation in forecast errors. To avoid , where models capture noise rather than signal leading to poor generalization, information criteria like AIC and are preferred over measures like R-squared, as they explicitly penalize parameter proliferation. In financial time series, simulates strategy performance on historical data by iteratively training and validating out-of-sample, revealing overfitting through degraded future returns. Model selection procedures automate or guide choice among candidates. Stepwise methods iteratively add or remove parameters based on criteria like AIC, while automated tools such as auto.arima in R's forecast package employ tests, criteria minimization, and stepwise search to identify optimal orders efficiently.

Advanced Applications

Classification and Anomaly Detection

Time series involves techniques to assign labels to entire time series or subsequences based on their patterns, often requiring feature extraction to transform raw data into discriminative representations. Common approaches include extracting features, such as coefficients or transforms, which capture -domain characteristics, followed by application of classifiers like support vector machines (SVMs). For instance, in heartbeat tasks, power in specific bands serves as input to SVMs, achieving high accuracy by distinguishing rhythmic patterns from irregularities. Shapelets, small discriminative subsequences, enable motif-based by identifying segments that best separate classes; the seminal method involves searching for shapelets that maximize the gain criterion in decision trees. methods, such as convolutional neural networks (CNNs) and transformer-based models, enable end-to-end directly from raw time series, achieving state-of-the-art results on large-scale benchmarks as of 2025. Dynamic time warping (DTW) provides a robust for , particularly in nearest-neighbor frameworks, by aligning time series of varying lengths or speeds. The DTW distance d(i,j) between two series X = (x_1, \dots, x_n) and Y = (y_1, \dots, y_m) is computed as the minimum cost path in a warping , where the cost at each alignment point is |x_i - y_j|, subject to boundary and monotonicity constraints. d(i,j) = \min \left\{ d(i-1,j-1) + |x_i - y_j|, \, d(i-1,j) + |x_i - y_j|, \, d(i,j-1) + |x_i - y_j| \right\} This alignment allows k-NN classifiers to handle temporal distortions, outperforming Euclidean distance in tasks like gesture recognition. Anomaly detection in time series identifies unusual observations deviating from expected behavior, using both statistical and machine learning methods. The generalized extreme studentized deviate (ESD) test detects outliers in univariate series by iteratively computing studentized residuals and comparing them to critical values from the t-distribution, suitable for normally distributed data with potential multiple anomalies. For multivariate cases, isolation forests isolate anomalies by randomly partitioning the feature space, requiring fewer splits for outliers than normal points, and have been adapted for time series by treating lagged observations as features. These techniques find applications in fraud detection, where transaction time series anomalies signal unauthorized activities, such as sudden spikes in spending patterns. In sensor data, they enable fault diagnosis by flagging deviations in vibration or temperature readings indicative of equipment failure. Recent advances leverage for , including autoencoders that learn compressed representations of normal and flag high reconstruction errors as anomalies. LSTM networks extend this by modeling sequential dependencies, predicting future values on normal subsequences and detecting deviations in prediction errors, particularly effective for non-stationary series.

Clustering and Segmentation

Time series clustering involves grouping similar sequences based on their temporal patterns to uncover hidden structures in data, often using distance metrics tailored to sequential dependencies. Common metrics include the , which measures point-wise differences but is sensitive to shifts and scaling, and (DTW), which aligns sequences by allowing non-linear warping to accommodate variations in timing or speed. Algorithms such as k-means, adapted for time series by incorporating DTW or feature-based representations, partition data into clusters by minimizing intra-cluster variance, while builds a of merges or splits based on linkage criteria like single or complete linkage. These methods enable scalable analysis by transforming raw series into feature vectors, such as statistical summaries of trends and , reducing computational demands for large datasets. Subsequence clustering extracts segments via sliding windows to identify recurring patterns, known as motifs, particularly useful in domains like for discovering repeated biological signals. However, direct clustering of overlapping subsequences often yields meaningless results, as clusters degenerate into arbitrary patterns due to trivial matches from minor shifts, violating natural data constraints. To address this, motif discovery focuses on non-trivial matches, defining a k-motif as the subsequence with the highest count of similar instances separated by at least a minimum distance, then applying k-means or to these significant segments for meaningful grouping. Time series segmentation divides sequences into homogeneous segments by detecting change points where underlying properties, such as mean or variance, shift, revealing regime changes. The Pruned Exact Linear Time (PELT) uses dynamic programming with pruning to exactly minimize a penalized , achieving linear O(n) computational cost while estimating both the number and locations of change points more accurately than quadratic alternatives. segmentation iteratively identifies the most significant change point in a segment and recurses on sub-segments, with extensions like Wild Binary Segmentation enhancing detection in high-dimensional or noisy data by monitoring statistics over random intervals. In finance, clustering segments markets by grouping asset return series with similar volatility patterns, aiding enhanced index tracking and portfolio diversification. For wearable devices, it recognizes human activities by clustering accelerometer time series from smartwatches or smartphones, using methods like Fuzzy C-means to categorize motions such as walking or running without labeled data. Evaluation of time series clusters employs adapted internal metrics like the silhouette score, which measures how well a series fits its cluster relative to others using domain-specific distances such as DTW, with scores above 0.5 indicating strong cohesion and separation. Internal validation relies solely on the data's intrinsic structure, such as silhouette or gap statistics, while external validation compares clusters against ground-truth labels, like known activity types, using indices like adjusted Rand index to assess agreement.

Visualization Methods

Basic Time Series Plots

Basic time series plots provide essential visualizations for univariate data, enabling the identification of trends, , distributions, and dependencies over time. These plots are foundational in , helping practitioners detect patterns that inform subsequent modeling steps. Line plots, histograms, functions (ACF), partial autocorrelation functions (PACF), lag plots, and seasonal plots each offer unique insights into the structure of a time series. Line plots, also known as time plots, the observed values X_t against time t, typically connecting points with straight lines to highlight . This reveals long-term trends, such as gradual increases or decreases, and cyclic patterns, including short-term fluctuations or annual cycles. For instance, in monthly passenger data for Ansett Airlines from 1987 to 1992, a line plot discloses an overall upward trend interrupted by seasonal dips during holidays and anomalies like zero passengers in 1989 due to an industrial dispute. Similarly, sales data for antidiabetic drugs exhibit a clear upward trend with pronounced seasonal peaks in , attributable to renewals. These plots are the starting point for any time series examination, as they directly display temporal evolution. Histograms and density plots illustrate the marginal distribution of the time series values, often computed for the entire series or subsets at fixed lags to assess stationarity or changes in variability. A histogram bins the values X_t and counts frequencies, while a density plot overlays a smoothed kernel estimate to approximate the probability density function, providing a continuous view of the distribution shape—such as unimodal, skewed, or forms. In time series contexts, these are particularly useful for lagged versions, like plotting the distribution of X_{t-k} for a specific k, to evaluate if marginal properties remain consistent across lags, indicating weak dependence or potential non-stationarity. For residuals in fitted models, histograms confirm approximate normality, with overlaid density curves aiding in detecting deviations like heavy tails. These visualizations prioritize understanding the overall spread and central tendency rather than temporal order. ACF plots consist of bar charts displaying the autocorrelation coefficients \rho_k at various lags k, where \rho_k = \text{Corr}(X_t, X_{t-k}), typically with bars extending beyond a 95% confidence interval (blue shaded area) indicating significant linear dependence. These plots decay gradually for non-stationary series with trends, showing slow attenuation, or exhibit spikes at seasonal lags for periodic patterns, aiding in detecting cycles or the need for differencing. In the Box-Jenkins methodology, ACF plots help identify the moving average (MA) order q, as significant correlations cut off after lag q in an MA(q) process. For example, in quarterly beer production data, ACF spikes at multiples of 4 reveal quarterly seasonality. PACF plots, similarly rendered as bar charts of partial autocorrelations \phi_{kk}, measure the correlation between X_t and X_{t-k} after removing intermediate lags' effects, cutting off after lag p for an autoregressive (AR(p)) process. In ARIMA modeling, PACF identifies the AR order p by noting where partial correlations become insignificant. Both are crucial for model order selection in ARIMA frameworks. Lag plots are scatterplots of X_t against X_{t-k} for a fixed k, revealing serial dependence through patterns like linear trends or clusters. Strong positive correlations appear as upward-sloping lines, while negative ones show downward slopes; random scatter suggests . For quarterly beer production, lag-4 and lag-8 plots display positive relationships due to annual , whereas lag-2 and lag-6 show negative ones from alternating peaks and troughs. These plots complement ACF by visualizing non-linear dependencies and are effective for detecting cycles at specific lags without assuming linearity. Multiple lag plots, arrayed in a , provide a comprehensive view of structure across k. Seasonal plots overlay subseries from multiple periods, such as plotting values against months with lines or points for each year, or using boxplots grouped by to summarize medians, quartiles, and outliers. This isolates seasonal effects, making patterns clearer than in full line plots; for example, monthly antidiabetic drug sales show consistent January peaks across years via overlaid lines, while boxplots highlight variability in by half-hour across days. Subseries plots connect observations within each over time, revealing trend changes in . These visualizations are ideal for confirming and quantifying periodic components referenced in time series characteristics.

Multivariate and Interactive Visualizations

Multivariate time series visualizations extend univariate plotting techniques to handle multiple interrelated series or dimensions simultaneously, enabling the identification of patterns, correlations, and dependencies across variables over time. These methods address the challenges of high-dimensional by employing matrix-based representations, superimposed elements, or faceted views, which facilitate comparative analysis without overwhelming the viewer. For instance, in financial datasets involving multiple asset prices, such visualizations reveal co-movements that inform and portfolio strategies. Heatmaps are particularly effective for depicting in multivariate time series, where rows and columns represent different series or time lags, and cell colors encode strengths. This approach highlights both linear and nonlinear dependencies, such as multifractal patterns in economic indicators like sales and GDP, by using detrended cross-correlation analysis (DCCA) within sliding windows to generate color-coded matrices that uncover cyclical behaviors spanning 3-4 years. Clustered heatmaps further enhance this by grouping similar series via , displaying average trends as band-shaped rectangles while allowing interactive expansion to detailed line graphs for precise value inspection, as demonstrated in data analyses where this method achieved an 83.3% correct response rate in feature recognition tasks compared to standalone heatmaps or lines. Parallel coordinates plots provide a radial or linear arrangement of vertical axes, each representing a of the multivariate time series, with polylines connecting values across axes to visualize high-dimensional slices or trajectories. This technique is suited for exploring temporal evolutions in multi-attribute data, such as player movements in , where enhancements like density distributions (e.g., plots along axes) and brushing interactions reveal clusters and correlations in features like speed and acceleration over match time series. Integrating time-series subplots directly into parallel coordinates axes allows for the detection of temporal trends within high-dimensional contexts, such as incident data spanning multiple variables over years. Overlapping charts superimpose multiple time series lines on a shared axis, often using (alpha blending) to mitigate clutter and reveal aggregate . The DenseLines method, for example, computes a normalized density heatmap from millions of series—such as NYSE prices or usage logs—where color indicates overlap , enabling scalable trend detection without tracing individual paths. In predictive applications, overlaying actual and forecasted values (e.g., for categories) with distinct line styles and distinguishes performance deviations, supporting constructions for monitoring limits in time-varying processes like temperature records. Separated charts, or small multiples, arrange identical subplot frames side-by-side to display subsets of multivariate data, such as different series or time periods, maintaining consistent scales for direct comparison. This approach excels in contexts, like NDVI vegetation indices across sites, where each small chart isolates a variable's temporal profile to highlight divergences without superposition artifacts. Originating from principles in graphical studies, small multiples improve tasks in multiple time series by leveraging spatial repetition, as evidenced in evaluations showing superior slope estimation accuracy over integrated single charts. Interactive tools amplify these visualizations through user-driven exploration, incorporating zoom, pan, and animation in libraries like and for dynamic multivariate time series analysis. enables custom, data-driven manipulations of elements to create zoomable or animated heatmaps, ideal for bespoke high-dimensional explorations. supports declarative, web-based interactions such as hovering for cross-section details in overlaid stock series or faceted forecasts, with recent advances in 2024-2025 emphasizing real-time updates and guidance systems for preprocessing tasks like anomaly brushing in . These tools, as reviewed in systematic studies, enhance decision-making in domains from to by integrating static methods with responsive interfaces.

Notation and Implementation

Mathematical Notation and Measures

In time series analysis, a univariate time series is commonly denoted as \{X_t\}_{t=1}^n, where X_t represents the observation at time t, and the sequence spans n time points. This notation assumes the data are ordered chronologically, with t indexing equidistant intervals such as daily or monthly observations. For multivariate series, extensions like \mathbf{X}_t = (X_{1,t}, \dots, X_{p,t})^\top are used, but the univariate form serves as the foundational representation. White noise is a fundamental component in time series models, denoted as \{\varepsilon_t\} where \varepsilon_t \sim \mathrm{WN}(0, \sigma^2), indicating a sequence of uncorrelated random variables with zero mean and constant variance \sigma^2. This implies \mathbb{E}[\varepsilon_t] = 0, \mathrm{Var}(\varepsilon_t) = \sigma^2, and \mathrm{Cov}(\varepsilon_t, \varepsilon_s) = 0 for all t \neq s, often assuming independence for stronger conditions. White noise represents the unpredictable residual component after accounting for systematic patterns. The , denoted B or L, facilitates compact expression of temporal dependencies, where B X_t = X_{t-1}, and higher powers shift further back, i.e., B^k X_t = X_{t-k}. Lag polynomials, such as \Phi(B) = 1 - \phi_1 B - \dots - \phi_p B^p for an autoregressive component, multiply these operators to model linear combinations of past values. This notation simplifies the representation of difference equations in stationary processes. Key assumptions underpin asymptotic theory in time series, including ergodicity and mixing conditions. Ergodicity ensures that time averages converge to ensemble expectations, allowing sample moments to estimate population parameters consistently. Mixing strengthens this by quantifying dependence decay over time lags, with strong mixing (α-mixing) defined via \alpha(k) = \sup_{m} \sup_{A,B} |\mathbb{P}(A \cap B) - \mathbb{P}(A)\mathbb{P}(B)| approaching zero as k \to \infty, enabling central limit theorems for dependent data. Second-order stationarity, or weak stationarity, requires the mean \mathbb{E}[X_t] = \mu to be constant across t, the variance \mathrm{Var}(X_t) = \gamma_0 to be finite and time-invariant, and the \mathrm{Cov}(X_t, X_{t+k}) = \gamma_k to depend only on the k. These conditions ensure the function \gamma_k fully characterizes second-moment properties, facilitating analysis of linear processes without trends or heteroskedasticity. The provides a frequency-domain measure of periodicity, defined for a series as f(\omega) = \frac{1}{2\pi} \sum_{k=-\infty}^{\infty} \gamma_k e^{-i \omega k}, where \omega \in [-\pi, \pi] is the and \gamma_k is the at k. This integrates to the variance \int_{-\pi}^{\pi} f(\omega) \, d\omega = \gamma_0, decomposing power across frequencies for . The H quantifies long-memory behavior in time series, where $0.5 < H < 1 indicates positive persistence, meaning large values are likely followed by large values, contrasting with short-memory processes where H = 0.5. Estimated via rescaled range analysis, H relates to the decay rate \rho_k \sim k^{2H-2} for large k, aiding identification of in financial or hydrological data. Distance measures assess similarity between series, with the defined as d(X, Y) = \sqrt{\sum_{t=1}^n (X_t - Y_t)^2}, requiring aligned lengths and assuming rigid temporal correspondence, suitable for synchronized data. For misaligned series, (DTW) extends this by finding an optimal elastic alignment to minimize the cumulative squared differences, capturing shape similarities despite timing variations.

Software Tools and Libraries

Several open-source and proprietary software tools and libraries facilitate time series analysis, ranging from classical statistical modeling to deep learning-based forecasting and scalable processing for large datasets. These resources emphasize ease of use, integration with data pipelines, and support for modern workflows as of 2025. In the R programming language, the forecast package provides comprehensive methods for univariate time series forecasting, including automatic ARIMA modeling, exponential smoothing, and ETS models, enabling users to generate predictions and visualize uncertainty intervals efficiently. Complementing this, the tseries package offers tools for ARIMA estimation, unit root testing, and ARCH/GARCH modeling, supporting computational finance applications alongside general time series diagnostics. For handling tidy temporal data, the tsibble package introduces a data infrastructure that aligns with tidyverse principles, allowing seamless manipulation of time-indexed datasets, gap filling, and aggregation over irregular intervals. Python libraries dominate modern time series workflows due to their flexibility and ecosystem integration. The statsmodels library implements classical models such as , SARIMAX, and , providing robust statistical tools for , diagnostics, and testing in time series contexts. For automated , Facebook's Prophet library fits additive models to handle , holidays, and trends, making it suitable for business applications with minimal configuration. Addressing deep learning needs, the GluonTS toolkit from AWS supports probabilistic with models like DeepAR and Transformers, leveraging or MXNet backends for scalable training on multivariate time series. MATLAB's Econometrics Toolbox includes specialized functions for time series econometrics, notably the garch object for estimating and simulating Generalized Autoregressive Conditional Heteroskedasticity (GARCH) models to capture volatility clustering in financial data. GNU Octave provides an open-source econometrics package offering similar capabilities for model fitting and forecasting. For big data environments, Apache Spark's MLlib enables scalable time series processing through distributed pipelines, supporting feature engineering, regression, and clustering on massive datasets, often combined with custom transformations for temporal dependencies. Cloud integrations like AWS Timestream provide serverless storage and querying for time series, featuring built-in gap-filling functions such as interpolation to handle missing observations in high-volume IoT or operational data streams. Specialized tools further enhance workflow automation and performance. offers a visual platform for building time series pipelines via drag-and-drop nodes, incorporating components for , , and without extensive coding. In , the TimeSeries.jl package delivers high-performance structures and operations for time-indexed arrays, optimizing computations for large-scale simulations and statistical .

References

  1. [1]
    Time Series and Spectral Analysis
    Feb 19, 2018 · A time series is a sequence of observations recorded at a succession of time intervals. In general, time series are characterized by dependence.
  2. [2]
    What are Time Series Made of? | D-Lab
    Dec 10, 2024 · The most common decomposition has four components: A Trend component, which represents the overall direction of the series, a Seasonal ...
  3. [3]
    [PDF] Chapter 19 Time Series Analysis
    A time series is data ordered by time, typically at regular intervals. It includes trend, seasons, cycles, and an irregular component.
  4. [4]
    [PDF] Time Series Introduction - UCLA Statistics & Data Science
    Overview. A time-series is a set of observations on the same object over time. Classically, these observations are assumed to be at regular (i.e. evenly ...
  5. [5]
    [PDF] 1 Short Introduction to Time Series
    Mar 21, 2023 · A time series is a sequence of stochastic variables observed over time. Most of what we do in economic modeling involves time series.
  6. [6]
    [PDF] Introduction to Time Series Forecasting - Data-X
    Time series are used to forecast some measured quantity at points in the future. Forecasting time series applications naturally arise in many business and ...
  7. [7]
    [PDF] Classical and Contemporary Approaches to Big Time Series ...
    Jun 30, 2019 · This tutorial aims to bring together classical forecasting techniques, time series data mining techniques, and neu- ral forecasting methods ...
  8. [8]
    [PDF] Lecture 1: Characteristics and Examples of Time Series Data
    – Economics: stock prices or stock returns over time. – Social science: birth rates or school acceptance rates over time. – Epidemiology: Covid-19 cases or ...
  9. [9]
    [PDF] Peter J. Brockwell Richard A. Davis Third Edition - CINEC Repository
    May 2, 2020 · This book is aimed at the reader who wishes to gain a working knowledge of time series and forecasting methods as applied in economics, ...
  10. [10]
    6.4.1. Definitions, Applications and Techniques
    Definition, Definition of Time Series: An ordered sequence of values of a variable at equally spaced time intervals. Time series occur frequently when ...Missing: statistics | Show results with:statistics
  11. [11]
    [PDF] TIME SERIES MODELLING, INFERENCE AND FORECASTING
    1.1. The expression time series data, or time series, usually refers to a set of observations collected sequentially in time. These observations could have ...
  12. [12]
    [PDF] Time Series Basics
    7.1 The Big Picture. A time series is a sequence of observed data values that are collected over time.
  13. [13]
    [PDF] Time Series: Autoregressive models AR, MA, ARMA, ARIMA
    Oct 23, 2018 · Components of a Time Series. • In general, a time series is affected by four components, i.e. trend, seasonal,cyclical and irregular components.
  14. [14]
    A brief history of time series analysis - Stockholm University
    May 13, 2022 · U Yule and J. Walker in the 1920s and 1930s. During this time the moving average was introduced to remove periodic fluctuations in the time ...
  15. [15]
    [PDF] 25 Years of Time Series Forecasting - Rob J Hyndman
    Jan 6, 2006 · Abstract: We review the past 25 years of research into time series forecasting. In this silver jubilee issue, we naturally highlight results ...
  16. [16]
    Understanding Time Series: Analyzing Data Trends Over Time
    By examining financial data like stock prices, earnings, or GDP, investors and analysts can identify influential factors over time. Time series analysis enables ...
  17. [17]
    Time Series Analysis Introduction - Statistics By Jim
    Time series analysis tracks characteristics of a process at regular time intervals. Use it to understand how a metric changes over time and make forecasts.
  18. [18]
    Panel data (time-series cross-section)
    Panel data (time-series cross-section)¶. Cross sectional data means that we have data from many units, at one point in time. Time series data means that we have ...
  19. [19]
    [PDF] Lecture 1 ELE 301: Signals and Systems - Princeton University
    A continuous-time signal has values for all points in time in some (possibly infinite) interval. A discrete time signal has values for only discrete points in ...
  20. [20]
    [PDF] Comparison of continuous and discrete-time data-based modeling ...
    In this paper, we compare the standard, continuous-time approach with a fully discrete-time approach, in a special case where the observations are known in ...
  21. [21]
    [PDF] Panel Data Analysis Fixed and Random Effects using Stata
    Setting data as panel. Once the data is in long form, we need to set it as panel so we can use Stata's panel data xt commands and the time series operators.
  22. [22]
    [PDF] MODELLING IRREGULARLY SPACED TIME SERIES UNDER ...
    as earthquakes, floods, or volcanic eruptions typically occur at irregular time intervals. In observational astronomy, for example, measurements of properties.
  23. [23]
    Chapter 10 Forecasting hierarchical or grouped time series - OTexts
    Hierarchical time series are nested categories, like product types. Grouped time series combine product and geographic hierarchies. Forecasts must add up ...
  24. [24]
    End-to-End Architecture for Real-Time IoT Analytics and Predictive ...
    May 7, 2025 · This paper presents a novel comprehensive architecture that enables end-to-end processing of IoT data streams, from acquisition to actionable insights.
  25. [25]
    [PDF] New Ways to Store and Access Data
    Time series databases enable a fundamental step in the central storage and analysis of many types of machine data. As such, they lie at the heart of the ...
  26. [26]
    NASA Open APIs
    Welcome to the NASA API portal. The objective of this site is to make NASA data, including imagery, eminently accessible to application developers. This catalog ...NASA Open Data Portal · SBDB Close Approach Data API · Open Source Software
  27. [27]
    TSGM: A Flexible Framework for Generative Modeling of Synthetic ...
    Jul 9, 2024 · In this work, we introduce Time Series Generative Modeling (TSGM), an open-source framework for the generative modeling and evaluation of synthetic time series ...
  28. [28]
    Benchmarking Missing Data Imputation Methods for Time Series ...
    Aug 29, 2025 · For loop, linear interpolation again had significantly lower bias than LOCF across all percentages for NMAR and MAR (all p < 0.005), but no ...Missing: cleaning | Show results with:cleaning
  29. [29]
    Outlier Detection in Time-Series Receive Signal Strength ... - MDPI
    2.1. Z-Score Method. This method calculates the difference between each RSS observation and the mean of the time-series RSS observations. Then, the result ...
  30. [30]
    3.2 - Identifying Outliers: IQR Method | STAT 200
    We can use the IQR method of identifying outliers to set up a “fence” outside of Q1 and Q3. Any values that fall outside of this fence are considered outliers.
  31. [31]
    pandas.DataFrame.resample — pandas 2.3.3 documentation
    Resample time-series data. Convenience method for frequency conversion and resampling of time series. The object must have a datetime-like index ...
  32. [32]
    8.1 Stationarity and differencing | Forecasting - OTexts
    Transformations such as logarithms can help to stabilise the variance of a time series. Differencing can help stabilise the mean of a time series by ...
  33. [33]
    Survey:Time-series data preprocessing: A survey and an empirical ...
    Time-series data preprocessing transforms raw data into quality input for AI, including techniques like dimensionality reduction, normalization, and outlier ...
  34. [34]
    Improving Time Series Data Quality: Identifying Outliers and ... - MDPI
    This study developed and validated a dual-phase framework to improve data quality using a 60-month gas and weather dataset from Jubail Industrial City, Saudi ...Missing: monotonic | Show results with:monotonic
  35. [35]
    2.8 Autocorrelation | Forecasting: Principles and Practice (3rd ed)
    Autocorrelation measures the linear relationship between lagged values of a time series. There are several autocorrelation coefficients, corresponding to each ...
  36. [36]
    Time Series Analysis | Wiley Series in Probability and Statistics
    Author(s):. George E. P. Box, Gwilym M. Jenkins, Gregory C. Reinsel ; First published:12 June 2008 ; Print ISBN:9780470272848 | ; |Online ISBN:9781118619193 ...
  37. [37]
    6.4.4.6. Box-Jenkins Model Identification
    At the model identification stage, our goal is to detect seasonality, if it exists, and to identify the order for the seasonal autoregressive and seasonal ...
  38. [38]
    6.2 Moving averages | Forecasting: Principles and Practice (2nd ed)
    A moving average estimates the trend-cycle by averaging values within k periods of time, smoothing the data and removing randomness.
  39. [39]
    6.4.4.3.1. Seasonal Subseries Plot
    Purpose, Seasonal subseries plots (Cleveland 1993) are a tool for detecting seasonality in a time series. This plot is only useful if the period of the ...
  40. [40]
    Postwar U.S. Business Cycles: An Empirical Investigation - jstor
    ... Hodrick-Prescott filter and studied its properties in the intervening eighteen years since its original development. ROBERT J. HODRICK is Nomura Professor ...
  41. [41]
    X-13ARIMA-SEATS Seasonal Adjustment Program
    Jul 10, 2025 · X-13ARIMA-SEATS is seasonal adjustment software by the Census Bureau, using ARIMA models and SEATS software for seasonal adjustments.
  42. [42]
    VII. On a method of investigating periodicities disturbed series, with ...
    On a method of investigating periodicities disturbed series, with special reference to Wolfer's sunspot numbers. George Udny Yule.
  43. [43]
    [PDF] Conditions for Stationarity and Invertibility James L. Powell ...
    Invertibility of Moving Average Processes. If an MA(q) process yt. = μ + εt + ... More generally, invertibility of an MA(q) process is the flip side of ...
  44. [44]
    On periodicity in series of related terms - Journals
    Published:03 June 1931https://doi.org/10.1098/rspa.1931.0069. Abstract. An important extension of our ideas regarding periodicity was made in 1927 when Yule ...
  45. [45]
    Generalized autoregressive conditional heteroskedasticity
    A natural generalization of the ARCH (Autoregressive Conditional Heteroskedastic) process introduced in Engle (1982) to allow for past conditional variances.
  46. [46]
    [PDF] Threshold models in time series analysis „ 30 years on
    This paper is a selective review of the development of the threshold model in time series analysis over the past. 30 years or so.Missing: seminal | Show results with:seminal
  47. [47]
    Threshold models in time series analysis—30 years on
    Aug 6, 2025 · This paper is a selective review of the development of the threshold model in time series analysis over the past 30 years or so.
  48. [48]
    [PDF] On-a-Threshold-Model.pdf - ResearchGate
    After a brief summary of some pertinent properties of the threshold models introduced by Tong (1978), we describe a statistical method for their ...
  49. [49]
    [PDF] Threshold Autoregression, Limit Cycles and Cyclical Data
    Jul 1, 1980 · This paper extends the classical two-regime threshold autoregressive model by introducing hysteresis to its regime-switching structure ...
  50. [50]
    A Tutorial on Estimating Time-Varying Vector Autoregressive Models
    Nov 14, 2017 · In this tutorial paper, we introduce methods to estimate time-varying VAR models based on splines and kernel-smoothing with/without regularization.
  51. [51]
    [PDF] A Tutorial on Estimating Time-Varying Vector Autoregressive Models
    In this tutorial paper, we introduce methods to estimate time-varying VAR models based on splines and kernel-smoothing with/without regularization. We use ...
  52. [52]
    Time Series Analysis by State Space Methods | Oxford Academic
    Dec 17, 2013 · This book presents a comprehensive treatment of the state space approach to time series analysis. The distinguishing feature of state space time ...
  53. [53]
    On the Theory of Bilinear Time Series Models - Oxford Academic
    SUMMARY. The theory of bilinear time series models is considered in this paper. The sufficient conditions for asymptotic stationarity of the bilinear time.Missing: nonlinear | Show results with:nonlinear
  54. [54]
    Ch. 21. On the theory of discrete and continuous bilinear time series ...
    Nonlinear models often produce better forecasts than linear models. One of the models which received considerable attention is the bilinear model. In this ...
  55. [55]
    [PDF] GENERALIZED AUTOREGRESSIVE CONDITIONAL ...
    In the ARCH(q) process the conditional variance is specified as a linear function of past sample variances only, whereas the. GARCH(p, q) process allows lagged ...
  56. [56]
    [PDF] Bayesian State Space Models in Macroeconometrics - Joshua Chan
    This paper outlines re- cent developments in state space modelling applied to macroeconomics using Bayesian methods. We outline the directions of recent ...
  57. [57]
    [PDF] Fully Bayesian Analysis of Switching Gaussian State Space Models
    In the present paper we study switching state space models from a. Bayesian point of view. We discuss various MCMC methods for Bayesian estimation, among them ...Missing: seminal | Show results with:seminal<|control11|><|separator|>
  58. [58]
    [PDF] Box-Jenkins modelling - Rob J Hyndman
    May 25, 2001 · The Box-Jenkins approach to modelling ARIMA processes was described in a highly in- fluential book by statisticians George Box and Gwilym ...Missing: SARIMA | Show results with:SARIMA
  59. [59]
    Box, G. and Jenkins, G. (1970) Time Series Analysis Forecasting ...
    Data was analyzed using Exponential Smoothing (ES) and Autoregressive Integrated Moving Average (ARIMA) models. The results showed that ARIMA (2, 1, 2) was the ...
  60. [60]
    Empirical prediction intervals revisited - ScienceDirect.com
    This paper re-examines the properties of the empirical prediction interval. Specifically, we provide conditions for its asymptotic validity, evaluate its small ...
  61. [61]
    [PDF] PDF - Empirical prediction intervals revisited (WP 2/2013)
    Jul 31, 2013 · Both simulation and empirical stud- ies indicate that empirical prediction intervals are particularly robust for time series that are nearly non ...
  62. [62]
    The Holt-Winters Forecasting Procedure - jstor
    This generalizes simple exponential smoothing so as to cope with trend and seasonal variation. A technical description of the method is given in the appendix.
  63. [63]
    7.3 Holt-Winters' seasonal method | Forecasting - OTexts
    The smoothing parameters and initial estimates for the components have been estimated by minimising RMSE (α=0.306 α = 0.306 , β∗=0.0003 β ∗ = 0.0003 , γ=0.426 γ ...
  64. [64]
    [PDF] Forecasting at Scale
    Prophet, our model from Section 2, is the only method that accounts for the clear dip around the holidays and also the recent change in trend.
  65. [65]
    Informer: Beyond Efficient Transformer for Long Sequence Time ...
    Dec 14, 2020 · Recent studies have shown the potential of Transformer to increase the prediction capacity.
  66. [66]
  67. [67]
    Forecasting with Time Series Models | Intro to Time Series Class Notes
    Short-term forecasts (days, weeks) are typically more accurate than long-term forecasts (months, years). Forecast intervals provide a range of plausible ...
  68. [68]
    Multi-Horizon Forecast Comparison - Taylor & Francis Online
    Jul 16, 2019 · The variance of loss differences typically increases with horizon, and as such adding an additional horizon may actually decrease power.
  69. [69]
    Estimating the Dimension of a Model - Project Euclid
    Abstract. The problem of selecting one of a number of models of different dimensions is treated by finding its Bayes solution, and evaluating the leading terms ...
  70. [70]
    3.4 Evaluating forecast accuracy | Forecasting: Principles ... - OTexts
    The accuracy of forecasts can only be determined by considering how well a model performs on new data that were not used when fitting the model.
  71. [71]
    On a measure of lack of fit in time series models - Oxford Academic
    The paper considers the Box & Pierce test for lack of fit in autoregressive-moving average models and proposes a modification for improved approximation.Missing: original | Show results with:original
  72. [72]
    Comparing Predictive Accuracy - Taylor & Francis Online
    Abstract. We propose and evaluate explicit tests of the null hypothesis of no difference in the accuracy of two competing forecasts.
  73. [73]
    Backtesting - Definition, Example, How it Works
    Backtesting involves applying a strategy or predictive model to historical data to determine its accuracy. It can be used to test and compare the viability.
  74. [74]
    Automatic Time Series Forecasting: The forecast Package for R
    We describe two automatic forecasting algorithms that have been implemented in the forecast package for R. The first is based on innovations state space models ...
  75. [75]
    [PDF] Heartbeat Time Series Classification With Support Vector Machines
    This study uses support vector machines (SVMs) to classify heartbeat time series, using statistical methods and signal analysis to extract features. SVMs are ...
  76. [76]
    [PDF] Time Series Shapelets: A New Primitive for Data Mining
    ABSTRACT. Classification of time series has been attracting great interest over the past decade. Recent empirical evidence has strongly suggested.Missing: seminal | Show results with:seminal
  77. [77]
    [PDF] Dynamic programming algorithm optimization for spoken word ...
    Abstract-This paper reports on an optimum dynamic programming. (DP) based time-normalization algorithm for spoken word recognition. First, a general ...
  78. [78]
    [PDF] Making Time-series Classification More Accurate Using learned ...
    Aug 14, 2003 · Abstract. It has long been known that Dynamic Time Warping. (DTW) is superior to Euclidean distance for classification.Missing: seminal | Show results with:seminal
  79. [79]
    1.3.5.17.3. Generalized Extreme Studentized Deviate Test for Outliers
    The generalized (extreme Studentized deviate) ESD test (Rosner 1983) is used to detect one or more outliers in a univariate data set that follows an ...Missing: time | Show results with:time
  80. [80]
    Deep Learning in Financial Fraud Detection - ScienceDirect.com
    Aug 20, 2025 · This study presents a systematic review of 108 peer-reviewed publications (2019–2024) on the application of deep learning (DL) to financial ...
  81. [81]
    Anomaly Detection Using Autoencoders with Nonlinear ...
    This paper proposes to use autoencoders with nonlinear dimensionality reduction in the anomaly detection task.
  82. [82]
    Long Short Term Memory Networks for Anomaly Detection in Time ...
    This paper uses stacked LSTM networks for anomaly/fault detection in time series and demonstrates the power of this approach on ECG, space shuttle, ...
  83. [83]
    [PDF] Dynamic Time Warping Algorithm Review - CSDL
    The Sakoe-Chiba band runs along the main diagonal, i = j, and con- straints the range of warping while specified by the fixed width R ∈ N. The. 11. Page 12 ...
  84. [84]
    [PDF] A Scalable Method for Time Series Clustering - Rob J Hyndman
    In this paper we propose a scalable method for time series clustering that replaces the time series point values with some global measures of the.Missing: seminal | Show results with:seminal
  85. [85]
    Clustering of time-series subsequences is meaningless: implications ...
    Clustering of time-series subsequences is meaningless. More concretely, clusters extracted from these time series are forced to obey a certain constraint.
  86. [86]
    [PDF] Clustering of Time Series Subsequences is Meaningless
    In Section 5 we show that the many algorithms that use. STS clustering as a subroutine produce results indistinguishable from random clusters. Since the main ...
  87. [87]
    Optimal Detection of Changepoints With a Linear Computational Cost
    This article introduces a new method for detecting changepoints with linear computational cost, which is faster than existing methods. It also improves ...
  88. [88]
    Clustering Time-Series Data Generated by Smart Devices for ...
    Aug 4, 2025 · This paper presents a comparative study of three clustering algorithms such as K-means, hierarchical agglomerative clustering and Fuzzy C-means ...
  89. [89]
    Evaluating multivariate time-series clustering using simulated ...
    Dec 15, 2023 · Silhouette coefficient and stability are two widely used alternative measures to evaluate clustering. Their results are shown in Fig. 4, Fig. 5, ...
  90. [90]
    Validation of cluster analysis results on validation data: A systematic ...
    Dec 23, 2021 · External validation makes use of additional (external) information that was not used for clustering. For example, when clustering a cancer gene ...Abstract · INTRODUCTION · DIFFERENT PERSPECTIVES... · A SYSTEMATIC...
  91. [91]
  92. [92]
    5.4 Residual diagnostics | Forecasting: Principles and Practice (3rd ...
    A related (and more accurate) test is the Ljung-Box test ... Thus, we can conclude that the residuals are not distinguishable from a white noise series.
  93. [93]
    [PDF] 454-2013: The Box-Jenkins Methodology for Time Series Models
    Before using differencing to transform seasonal non-stationary time series values into stationary time series values, we need to check if the data over time ...
  94. [94]
    Chapter 9 ARIMA models | Forecasting: Principles and Practice (3rd ed)
    - **PACF Plots**: Used to identify the order of an AR (autoregressive) model by showing partial autocorrelations (φ_{kk}) at different lags.
  95. [95]
    2.7 Lag plots | Forecasting: Principles and Practice (3rd ed) - OTexts
    Lag plots show lagged values of a time series on the horizontal axis, plotting yt against yt-k. Lags 4 and 8 show positive relationships, while 2 and 6 show ...Missing: hyndman | Show results with:hyndman<|separator|>
  96. [96]
  97. [97]
    [PDF] MATH 545 - Intro to Time Series Analysis - Léo Belzile
    Using the standard notation FX1,...,Xn (x1,...,xn) = P(X1, ≤ x1,X2 ≤ x2,...,Xn ≤ xn), this function fully specifies the joint probability model for the.
  98. [98]
    [PDF] s Part 1: White Noise and Moving Average Model - NYU Stern
    White Noise t t −1 t −1 t. A stationary time series ε is said to be white noise if t t s. Corr(ε ,ε ) = 0 for all t ≠s . e. Thus, ε is a sequence of ...
  99. [99]
    [PDF] Chapter 3. Stationarity, white noise, and some basic time series ...
    Use white noise to construct some basic time series models. Page 2. Definition: Weak stationarity and strict stationarity. A time series model which is ...
  100. [100]
    [PDF] Stationarity, Lag Operator, ARMA, and Covariance Structure
    Definition 6. Lag operator Denoted L. Lyt = yt−1. + apyt−p 3) Lag polynomials can be multiplied.
  101. [101]
    Basic Properties of Strong Mixing Conditions. A Survey and Some ...
    This paper surveys basic properties of strong mixing conditions, updating a 1986 paper, and includes new developments and open problems.
  102. [102]
    [PDF] Lecture 13 Time Series: Stationarity, AR(p) & MA(q)
    2nd order (weak) stationarity is weaker. Weak stationarity only considers means & covariances (easier to verify in practice). • Moments describe a distribution.
  103. [103]
    [PDF] A comment on measuring the Hurst exponent of financial time series
    Nov 1, 2004 · We obtain a more precise statistical significance test for the Hurst exponent and apply it to real financial data sets. Our empirical analysis ...
  104. [104]
    [PDF] Everything you know about Dynamic Time Warping is Wrong
    The Dynamic Time Warping (DTW) distance measure is a technique that has long been known in speech recognition community. It allows a non-linear mapping of one ...<|separator|>
  105. [105]
    Forecasting Functions for Time Series and Linear Models
    The R package forecast provides methods and tools for displaying and analysing univariate time series forecasts including exponential smoothing via state space ...
  106. [106]
    CRAN: Package tseries
    Sep 23, 2024 · tseries: Time Series Analysis and Computational Finance ; Depends: R (≥ 3.4.0) ; Imports: graphics, stats, utils, quadprog, zoo, quantmod (≥ 0.4-9) ...
  107. [107]
    CRAN: Package tsibble
    Jan 30, 2025 · The 'tsibble' provides tools to easily manipulate and analyse temporal data, such as filling in time gaps and aggregating over calendar periods.
  108. [108]
    Time Series analysis tsa - statsmodels 0.14.4
    statsmodels.tsa contains model classes and functions that are useful for time series analysis. Basic models include univariate autoregressive models (AR), ...Descriptive Statistics And... · Estimation · Arma Process
  109. [109]
    facebook/prophet: Tool for producing high quality forecasts ... - GitHub
    Sep 29, 2024 · Prophet is a procedure for forecasting time series data based on an additive model where non-linear trends are fit with yearly, weekly, and daily seasonality.Forecasting Time-Series Data ...Issues 449Pull requests 5ActionsActivity
  110. [110]
    GluonTS documentation
    GluonTS is a Python package for probabilistic time series modeling, focusing on deep learning based models, based on PyTorch and MXNet.Gluonts.time_feature... · Gluonts.time_feature.holiday... · Gluonts.dataset...
  111. [111]
    garch - GARCH conditional variance time series model - MATLAB
    The garch function returns a garch object specifying the functional form of a GARCH(P,Q) model, and stores its parameter values.
  112. [112]
    Econometrics Toolbox Documentation - MathWorks
    Econometrics Toolbox™ provides functions and interactive workflows for analyzing and modeling time series data. It offers a wide range of visualizations and ...
  113. [113]
    Machine Learning Library (MLlib) Guide - Apache Spark
    MLlib is Spark’s machine learning library, aiming to make practical machine learning scalable and easy, providing tools like ML algorithms and pipelines.
  114. [114]
    Time series functions - Amazon Timestream - AWS Documentation
    Amazon Timestream for LiveAnalytics supports timeseries functions, such as derivatives, integrals, and correlations, as well as others.
  115. [115]
    Time-Series Database – Amazon Timestream – AWS
    Timestream offers fully managed, purpose-built time-series database engines for workloads from low-latency queries to large-scale data ingestion.Timestream pricing · What is Timestream for InfluxDB? · Amazon Timestream FAQs
  116. [116]
    Time Series Analysis with Components - KNIME
    Mar 9, 2020 · The time series components are available on the KNIME Hub. Drag and drop them into your workflow editor and start building your KNIME workflows ...
  117. [117]
    TimeSeries Overview - JuliaStats
    Documentation for TimeSeries.jl ... The TimeSeries package provides convenient methods for working with time series data in Julia.