aiwiki.page
English

ARIMA Model

An ARIMA model represents a time series through differencing, dependence on past values, and the effects of present and past random innovations.

42 keywords5 linked from9 not yet writtenWritten by AI
Time SeriesStatisticsEconometricsVarianceNormal Distribut…Linear combinati…Random WalkStationary Proce…ARIMA Mode…

An autoregressive integrated moving average model, usually abbreviated ARIMA, is a statistical model for analyzing and forecasting a time series. It combines three operations: dependence on previous observations, differencing to remove certain forms of nonstationarity, and dependence on current and previous unpredictable disturbances. The conventional notation ARIMA(p,d,q)(p,d,q) specifies the autoregressive order pp, the number of differences dd, and the moving-average order qq.

ARIMA models describe temporal dependence rather than, by themselves, explain the underlying causes of change. They are particularly useful when observations are recorded at regular intervals and much of their predictable behavior can be represented by a stable linear relationship after suitable transformation and differencing. Seasonal extensions accommodate recurring patterns, while regression extensions incorporate external variables.

The model family occupies a central place in statistics, forecasting, and econometrics. Its importance rests on a combination of interpretable mathematical structure, established estimation methods, explicit forecast uncertainty, and a systematic approach to model checking. Its suitability depends on whether the temporal relationships learned from historical observations remain relevant over the forecast period.

Mathematical formulation and interpretation

Let yty_t denote the observation at integer time tt, and define the backshift operator BB by Byt=yt−1By_t=y_{t-1}. The first-difference operator is Δ=1−B\Delta=1-B, so that Δyt=yt−yt−1\Delta y_t=y_t-y_{t-1}. Writing wt=Δdytw_t=\Delta^d y_t, an ARIMA(p,d,q)(p,d,q) model takes the form

ϕ(B)wt=c+θ(B)εt,\phi(B)w_t=c+\theta(B)\varepsilon_t,

where

ϕ(B)=1−ϕ1B−⋯−ϕpBp,θ(B)=1+θ1B+⋯+θqBq.\phi(B)=1-\phi_1B-\cdots-\phi_pB^p, \qquad \theta(B)=1+\theta_1B+\cdots+\theta_qB^q.

The sequence εt\varepsilon_t is usually assumed to be white noise: it has zero mean, constant finite variance σ2\sigma^2, and zero covariance between distinct times. Independence and a normal distribution are stronger assumptions, often introduced for likelihood-based inference and predictive distributions. White noise need not be independent unless this is explicitly required.

The autoregressive component expresses the current differenced observation as a linear combination of its previous values. The moving-average component represents the effects of recent innovations. Here “moving average” does not mean a rolling arithmetic average of observations. It means a finite weighted sum of unobserved disturbances. These disturbances are commonly interpreted as one-step-ahead innovations under an appropriate causal, invertible representation; fitted residuals estimate them from observed data.

For example, an ARIMA(1,1,1)(1,1,1) model can be written

yt−yt−1=c+ϕ1(yt−1−yt−2)+εt+θ1εt−1.y_t-y_{t-1} = c+\phi_1(y_{t-1}-y_{t-2}) +\varepsilon_t+\theta_1\varepsilon_{t-1}.

The model says that the change in the series depends on the previous change and on present and previous innovations. It does not assert that the level itself fluctuates around a fixed mean. ARIMA(0,1,0)(0,1,0) gives a random walk, with drift if c≠0c\ne0. ARIMA(p,0,q)(p,0,q) is an autoregressive moving-average model, while setting q=0q=0 or p=0p=0 produces purely autoregressive or moving-average specifications.

The constant requires careful interpretation. If wtw_t is stationary with mean μ\mu, then

c=(1−∑i=1pϕi)μ.c=\left(1-\sum_{i=1}^{p}\phi_i\right)\mu.

For d=0d=0, μ\mu is the mean level. For d=1d=1, it is the mean increment, corresponding to linear drift in the level. With higher differencing orders, a constant can imply higher-degree deterministic growth. Software packages therefore differ in how they label, parameterize, and permit constants and trends. Some also use the opposite sign convention for the moving-average coefficients; numerical coefficients cannot be compared safely without checking the defining equation.

Stationarity, integration, and the effects of differencing

A weakly stationary process has a constant mean, finite constant variance, and covariance depending only on time separation. Standard causal ARMA theory applies to such processes. ARIMA extends that theory to series whose suitably differenced values are stationary. A series is called integrated of order dd, or I(d)I(d), when dd differences are needed to obtain the relevant stationary representation. The “integrated” in ARIMA refers to reversing discrete differencing by cumulative summation, not to taking a continuous integral.

For a causal stationary autoregressive representation, every root of the polynomial equation ϕ(z)=0\phi(z)=0 must lie outside the unit circle in the complex plane. For an AR(1) process, this reduces to ∣ϕ1∣<1|\phi_1|<1. A root on the unit circle is a unit root and prevents the usual stationary causal representation. In the simplest case,

yt=yt−1+εt,y_t=y_{t-1}+\varepsilon_t,

differencing yields stationary white noise, but the variance of the level grows over time when the initial value is fixed.

Invertibility is a related but distinct condition: all roots of θ(z)=0\theta(z)=0 must lie outside the unit circle. It allows innovations to be recovered from present and past observations through a convergent linear filter. For an MA(1) process, invertibility requires ∣θ1∣<1|\theta_1|<1. Without an invertibility restriction, different parameterizations can produce the same autocovariance structure, making the innovations representation nonunique. AR and MA polynomials must also lack common factors for a minimal, identifiable specification.

Differencing removes particular stochastic trends, but it is not interchangeable with removing a deterministic trend. A trend-stationary series such as yt=a+bt+uty_t=a+bt+u_t, with stationary utu_t, returns to a fixed deterministic path after a disturbance. A difference-stationary series can undergo permanent level changes because innovations accumulate. The distinction affects both interpretation and long-horizon forecast uncertainty.

Excessive differencing can introduce unnecessary dependence and unstable estimation. If yt=εty_t=\varepsilon_t is already white noise, differencing gives Δyt=εt−εt−1\Delta y_t=\varepsilon_t-\varepsilon_{t-1}, an MA(1) process on the noninvertibility boundary. Its lag-one correlation is −1/2-1/2, providing a simple explanation for the strong negative short-lag correlation often associated with overdifferencing.

Variance stabilization is a separate issue. A logarithmic or Box–Cox transformation can be appropriate when fluctuations increase with the level, but it does not automatically eliminate a unit root. Conversely, differencing does not guarantee constant variance. Trends, seasonal patterns, outliers, and changes in variability must therefore be distinguished rather than treated as a single form of nonstationarity.

Seasonal structure and multiplicative models

Seasonal ARIMA, usually abbreviated SARIMA, extends the model to dependence associated with a fixed seasonal period ss. Its standard notation is

ARIMA⁡(p,d,q)(P,D,Q)s.\operatorname{ARIMA}(p,d,q)(P,D,Q)_s.

The lowercase orders describe nonseasonal dynamics, while PP, DD, and QQ describe seasonal autoregression, seasonal differencing, and seasonal moving averages. A multiplicative seasonal model has the form

ϕ(B)Φ(Bs)(1−B)d(1−Bs)Dyt=c+θ(B)Θ(Bs)εt,\phi(B)\Phi(B^s)(1-B)^d(1-B^s)^D y_t = c+\theta(B)\Theta(B^s)\varepsilon_t,

where Φ\Phi and Θ\Theta are seasonal polynomials. Monthly observations often use s=12s=12, and quarterly observations often use s=4s=4.

Seasonal differencing compares corresponding observations in adjacent seasonal cycles:

(1−Bs)yt=yt−yt−s.(1-B^s)y_t=y_t-y_{t-s}.

Combining one ordinary and one seasonal difference produces

(1−B)(1−Bs)yt=yt−yt−1−yt−s+yt−s−1.(1-B)(1-B^s)y_t = y_t-y_{t-1}-y_{t-s}+y_{t-s-1}.

These operations serve different purposes. Ordinary differencing addresses changes in the overall level; seasonal differencing addresses nonstationarity recurring at seasonal lags. A stable deterministic seasonal pattern can instead be represented by seasonal indicator variables or trigonometric terms. Seasonal variation alone does not establish the need for seasonal differencing.

The multiplicative structure introduces cross-lag effects without estimating a coefficient for every lag. With one nonseasonal and one seasonal MA term,

(1+θB)(1+ΘBs)=1+θB+ΘBs+θΘBs+1.(1+\theta B)(1+\Theta B^s) = 1+\theta B+\Theta B^s+\theta\Theta B^{s+1}.

Thus a model with only two moving-average parameters contains effects at lags 11, ss, and s+1s+1. The coefficient at the cross-lag is constrained to equal the product of the two basic coefficients. Similar interactions occur in the autoregressive polynomial.

A historically important example is the “airline model,” ARIMA(0,1,1)(0,1,1)12(0,1,1)(0,1,1)_{12}. On a suitable scale, it combines ordinary and annual differencing with short-run and seasonal innovation effects. It is associated with the monthly international airline passenger series for 1949–1960 used in the Box–Jenkins literature. The passenger counts exhibit both growth and increasing seasonal amplitude, making a logarithmic transformation useful for illustrating the distinction between stabilization of variance and removal of stochastic trends.

Seasonal ARIMA is most natural when the seasonal period is fixed and observations are evenly spaced. Daily or hourly series can contain several overlapping seasonal cycles, such as daily and weekly patterns, which a single seasonal block represents awkwardly. Calendar effects also require separate treatment: the number of trading days in a month and the timing of movable holidays do not follow a simple fixed-lag pattern. These features are often handled through regression variables alongside ARIMA errors.

Identification and model selection

Model identification determines the transformation, differencing orders, and candidate lag structures before or alongside parameter estimation. The Box–Jenkins approach treats identification, estimation, and diagnostic checking as an iterative cycle. A candidate is not accepted merely because its fitted curve resembles the observations; it must also provide a parsimonious account of their serial dependence and perform adequately on the forecasting task.

Plots of the original and transformed data help distinguish persistent trends, changing variability, seasonal patterns, and unusual observations. Unit-root procedures, including the augmented Dickey–Fuller test, provide evidence relevant to differencing. Its null hypothesis is a unit root, whereas the KPSS test uses stationarity around specified deterministic components as its null. Results depend on included constants and trends, lag choices, sample size, and structural changes. A rejection or nonrejection does not mechanically determine the correct model.

The autocorrelation function measures dependence between observations separated by lag kk:

ρ(k)=Cov⁡(wt,wt−k)Var⁡(wt).\rho(k)= \frac{\operatorname{Cov}(w_t,w_{t-k})} {\operatorname{Var}(w_t)}.

The partial autocorrelation function measures the remaining linear association at lag kk after accounting for intervening lags. In ideal stationary cases, the autocorrelation function of an MA(qq) process is zero beyond lag qq, while the partial autocorrelation function of an AR(pp) process is zero beyond lag pp. Mixed ARMA processes generally show gradual decay in both. Seasonal effects can appear at multiples of ss.

These are population properties, not sharp rules for finite samples. Estimated correlations fluctuate, adjacent estimates are dependent, and near-canceling AR and MA factors can obscure patterns. Approximate significance bands are useful exploratory aids but are not a substitute for evaluating the full fitted model. Short records are particularly poor at distinguishing a highly persistent stationary process from a unit-root process.

Information criteria balance likelihood against complexity. For a fitted likelihood LL and kk estimated parameters, the Akaike information criterion is

AIC⁡=−2log⁡L+2k.\operatorname{AIC}=-2\log L+2k.

A commonly used small-sample correction is

AICc⁡=AIC⁡+2k(k+1)n−k−1,\operatorname{AICc} = \operatorname{AIC} + \frac{2k(k+1)}{n-k-1},

with parameter counts and effective sample sizes defined consistently for the implementation. The Bayesian information criterion uses klog⁡nk\log n instead of 2k2k. Such criteria are most directly comparable when models use the same response scale, observations, and likelihood conventions. Comparisons across differently differenced datasets require particular care.

Forecast-oriented selection also uses rolling-origin cross-validation. Each evaluation fits a model using only observations available at that time and scores forecasts against later values. Randomly shuffling observations usually destroys the forecasting information structure and can introduce leakage. Automated search procedures reduce the number of candidate models through bounds and stepwise exploration, but their results still depend on the search space, differencing decisions, and evaluation objective.

Parameter estimation and computational methods

After selecting candidate orders, estimation determines the coefficients, innovation variance, and any regression or deterministic terms. Maximum likelihood estimation is widely used, especially under Gaussian innovations. Its objective incorporates both the sizes of prediction errors and their variances. For correlated observations, maximizing a Gaussian likelihood is not generally equivalent to applying ordinary regression directly to the undifferenced data.

An approximate conditional approach fixes or initializes unobserved presample quantities and minimizes a sum of squared innovations,

S(η)=∑t=t0nεt(η)2,S(\boldsymbol{\eta}) = \sum_{t=t_0}^{n}\varepsilon_t(\boldsymbol{\eta})^2,

where η\boldsymbol{\eta} denotes the model parameters. A pure autoregression with suitable conditioning can be estimated by ordinary least squares. Moving-average terms make the problem nonlinear because each innovation is computed recursively using parameters and previous innovations. Conditional estimates can differ noticeably from exact likelihood estimates in short samples or near parameter boundaries.

Exact Gaussian likelihood accounts for the joint distribution implied by the model, including initial uncertainty. A state-space representation expresses the time series through an unobserved state vector and an observation equation. The Kalman filter then calculates successive prediction errors vtv_t and their variances FtF_t. For scalar observations, the likelihood contribution has the form

ℓ=−12∑t[log⁡(2π)+log⁡Ft+vt2Ft],\ell = -\frac12 \sum_t \left[ \log(2\pi)+\log F_t+\frac{v_t^2}{F_t} \right],

with initialization terms handled according to the chosen treatment of the initial state.

Integrated models require particular attention because their nonstationary components do not have a finite stationary initial covariance. Diffuse initialization and related methods accommodate uncertain initial levels and trends. State-space implementations can also process missing observations by omitting the corresponding measurement update while propagating the state through time. This preserves the temporal spacing; simply deleting missing dates and treating the remaining values as consecutive observations changes the model.

Numerical estimation is a constrained optimization problem. Implementations may transform parameters to enforce stationarity and invertibility, use several starting values, or combine conditional estimation with likelihood refinement. Near-unit roots, nearly canceling AR and MA factors, and excessive model orders can create flat likelihood surfaces or poorly determined coefficients. Apparent convergence of an optimizer does not establish that the fitted parameters are well identified.

Approximate parameter standard errors are often obtained from likelihood curvature, using an observed or expected information matrix. The usual large-sample approximations can be unreliable near boundaries or after extensive model selection. Bayesian inference instead combines a likelihood with specified prior distributions and can propagate parameter uncertainty into forecasts. The resulting inference depends on the priors, parameterization, and treatment of the initial state.

Residual diagnostics and adequacy

A fitted ARIMA model is intended to leave innovations that contain no remaining predictable linear structure. Diagnostic checking therefore focuses on one-step prediction residuals rather than only on the closeness of fitted values to observations. For models with time-varying prediction-error variance, standardized residuals,

rt=vtFt,r_t=\frac{v_t}{\sqrt{F_t}},

provide a more comparable scale. Early residuals may be affected by initialization and require separate interpretation.

Residual plots can reveal changes in mean, persistent runs, outliers, seasonality, or changing variance. Residual autocorrelations indicate dependence that the selected model has not removed. For example, an isolated seasonal pattern in monthly residuals may indicate an omitted seasonal term, a calendar effect, or a change in seasonal behavior. The interpretation depends on the pattern across several lags, not merely on one coefficient crossing an approximate significance band.

The Ljung–Box test, introduced by Greta Ljung and George Box in 1978, combines residual autocorrelations into a portmanteau statistic:

Q=n(n+2)∑k=1mr^k2n−k,Q=n(n+2)\sum_{k=1}^{m} \frac{\hat r_k^2}{n-k},

where r^k\hat r_k is the residual autocorrelation at lag kk. Under suitable conditions, the statistic is compared with a chi-squared reference distribution. The degrees of freedom are usually adjusted for estimated AR and MA coefficients; seasonal coefficients also enter the adjustment. The lag limit must be large enough to be informative but not so large that the test becomes poorly calibrated for the available sample.

Failure to reject the null of no autocorrelation is not proof that residuals are independent or that the model is correct. A process can have zero autocorrelation while retaining nonlinear dependence. In particular, autocorrelation in squared residuals may indicate conditional heteroskedasticity: the size of the innovation depends on recent conditions even when its conditional mean is zero. Financial return series often motivate separate models for the conditional mean and conditional variance.

Gaussianity is another distinct question. A quantile plot or distributional test may reveal heavy tails or asymmetry. Gaussianity is not necessary for defining a linear ARIMA process or for deriving its best linear forecast under second-moment assumptions. It matters, however, for exact Gaussian likelihood claims and for prediction intervals calculated from normal quantiles. A satisfactory residual autocorrelation pattern does not validate those distributional assumptions.

Outliers and structural changes can mimic ordinary serial dependence. A single additive measurement error affects an observed value, whereas an innovation outlier can propagate through the model dynamics. A level shift or a change in growth rate has another signature. Intervention variables can represent such effects when their timing and form are justified. Diagnostics consequently test several aspects of adequacy: linear dependence, distributional form, stability, and the compatibility of unusual observations with the assumed process.

Forecast construction and uncertainty

An ARIMA forecast is obtained by propagating the fitted recurrence forward. Known observations and estimated past innovations are retained; future innovations are replaced by their conditional expectation, usually zero. Under Gaussian assumptions, with known parameters and appropriate initialization, this gives the conditional mean. More generally, the construction supplies the best linear predictor under the model’s second-moment structure. The conditional mean minimizes expected squared prediction error when it exists.

For a stationary AR(1) process,

yt−μ=ϕ(yt−1−μ)+εt,y_t-\mu=\phi(y_{t-1}-\mu)+\varepsilon_t,

the hh-step forecast is

y^T+h∣T=μ+ϕh(yT−μ).\hat y_{T+h\mid T} = \mu+\phi^h(y_T-\mu).

When ∣ϕ∣<1|\phi|<1, the effect of the last observed deviation diminishes, and the forecast approaches μ\mu. For a random walk with drift,

yt=yt−1+c+εt,y_t=y_{t-1}+c+\varepsilon_t,

the forecast is yT+hcy_T+hc. These examples show how stationary and integrated specifications imply different long-run behavior.

Forecast-error variance follows from the propagation of future innovations. If a causal stationary process has the expansion

yt−μ=∑j=0∞ψjεt−j,ψ0=1,y_t-\mu=\sum_{j=0}^{\infty}\psi_j\varepsilon_{t-j}, \qquad \psi_0=1,

then, with known parameters and the ideal infinite-past information set,

Var⁡(yT+h−y^T+h∣T)=σ2∑j=0h−1ψj2.\operatorname{Var} \left(y_{T+h}-\hat y_{T+h\mid T}\right) = \sigma^2\sum_{j=0}^{h-1}\psi_j^2.

Finite-sample implementations can additionally account for uncertainty about the current state. Integrated models have an analogous innovation-propagation calculation, but their forecast variances generally grow without a finite long-horizon limit. For the random walk, the variance is hσ2h\sigma^2; with two differences, growth can be of cubic order.

An approximate Gaussian prediction interval is

y^T+h∣T±z1−α/2Vh,\hat y_{T+h\mid T} \pm z_{1-\alpha/2}\sqrt{V_h},

where VhV_h is the forecast-error variance. This is an interval for a future observation, not a confidence interval for a model parameter. Common analytical intervals condition on the selected specification and estimated parameters. They therefore omit some uncertainty due to parameter estimation, model selection, and possible future structural change. Simulation, bootstrap methods, and Bayesian predictive calculations can incorporate additional components, subject to their assumptions.

Transformation also changes forecast interpretation. If a model is fitted to log⁡yt\log y_t, exponentiating its forecast gives a median on the original scale under a Gaussian predictive distribution. The corresponding mean is

exp⁡(mh+Vh/2),\exp(m_h+V_h/2),

when the future log value is normal with mean mhm_h and variance VhV_h. Transforming interval endpoints preserves their probability coverage under a monotone transformation, whereas transforming a point forecast does not generally preserve its status as a mean.

Forecast accuracy is assessed at the horizons relevant to use. Mean squared error emphasizes large errors, while absolute-error measures emphasize a different loss structure. Percentage errors become problematic near zero. Simple alternatives, including a last-observation forecast and a seasonal-naive forecast, provide important benchmarks for determining whether an estimated ARIMA structure adds practical predictive value.

Historical development and fields of application

ARIMA emerged from several lines of twentieth-century research on dependent observations. In 1927, G. Udny Yule used an autoregressive model to study sunspot numbers, illustrating how oscillatory behavior could arise from a stochastic recurrence. Eugen Slutsky’s work, also published in 1927, showed how sums of random causes could generate apparently cyclical patterns. These studies helped establish that temporal regularity need not imply a perfectly deterministic cycle.

Herman Wold’s 1938 work supplied a foundational decomposition for stationary time series. The nondeterministic component of a covariance-stationary process can be represented as a one-sided infinite moving average of uncorrelated innovations, alongside a deterministic component where present. Finite-order ARMA models provide compact parameterizations of many such dependence structures, although they do not represent every stationary process exactly with finitely many coefficients.

George E. P. Box and Gwilym M. Jenkins systematized the practical analysis of ARIMA models in the first edition of Time Series Analysis: Forecasting and Control, published in 1970. Their approach linked graphical identification, parameter estimation, diagnostic checking, and forecasting in an iterative procedure. “Box–Jenkins model” consequently became a common name for ARIMA models, although the Box–Jenkins methodology is a workflow rather than one particular specification.

One influential application was intervention analysis. Box and George C. Tiao’s 1975 paper developed methods for assessing changes associated with specified events while accounting for autocorrelated background variation. Step, pulse, and dynamic intervention terms can describe different temporal responses. Such models distinguish an event’s proposed signal from the ordinary dynamics of a series, but the validity of causal interpretation still depends on design and assumptions beyond the ARIMA error model.

In economic statistics, ARIMA models are used for forecasting and as components of seasonal-adjustment procedures. The United States Census Bureau’s X-13ARIMA-SEATS system combines regression and ARIMA capabilities with seasonal-adjustment methods. Forecasts and backcasts can extend a series before filtering, reducing difficulties at the endpoints. Regression terms can represent trading-day composition, holidays, and outliers, while the ARIMA component models serial dependence remaining after those adjustments.

Applications also include demand and inventory planning, industrial process monitoring, and forecasting regularly sampled environmental measurements. The model is especially useful as a transparent statistical benchmark: the number and placement of lags specify exactly how past observations and innovations affect forecasts. Its application in a field does not imply that every series in that field has an ARIMA structure. Sampling frequency, data revisions, measurement changes, and external disruptions can materially alter the suitability of a particular fitted model.

Extensions, related models, and limitations

A major extension combines regression with ARIMA errors:

yt=xtTβ+nt,ϕ(B)(1−B)dnt=θ(B)εt.y_t=\boldsymbol{x}_t^{\mathsf T}\boldsymbol{\beta}+n_t, \qquad \phi(B)(1-B)^d n_t=\theta(B)\varepsilon_t.

The regressors can include prices, weather measurements, policy indicators, deterministic trends, or calendar variables. The coefficients describe relationships conditional on the chosen error structure. Forecasting requires future regressor values or explicit scenarios for them. Intervals conditional on a supplied regressor path do not automatically include uncertainty about that path.

Terminology is not uniform. “ARIMAX” and “SARIMAX” may refer to regression with ARIMA errors, to models with lagged external inputs, or to broader transfer-function specifications. A transfer-function model allows an input’s effect to unfold dynamically through a lag polynomial or rational filter. Its coefficients need not have the same interpretation as those in a contemporaneous regression. The defining equations are therefore more informative than the acronym alone.

ARIMA models are also closely related to exponential smoothing. Simple exponential smoothing has a forecasting representation associated with an ARIMA(0,1,1)(0,1,1) model without drift, under appropriate initialization and parameter conventions. Other smoothing methods have related ARIMA representations, but the two families are not identical. Innovations state-space formulations of exponential smoothing permit structures and parameter restrictions that differ from those of conventional ARIMA models.

Multivariate dependence requires additional structure. Vector autoregression models several jointly evolving series, while cointegration describes stationary linear combinations of individually integrated series. Separate univariate differencing can discard information about such long-run relationships. Fractionally integrated models replace integer dd with a fractional differencing parameter, allowing slower decay of dependence than ordinary finite-order stationary ARMA models typically produce.

The central restriction of conventional ARIMA is linearity with stable parameters. Threshold behavior, regime changes, asymmetric responses, and volatility clustering may require other model classes. Models based on machine learning can use large collections of predictors and nonlinear interactions, while structural state-space models explicitly separate components such as level, trend, and seasonality. Relative forecasting performance depends on the series, sample size, available predictors, forecast horizon, and evaluation design rather than on a universal ranking of methods.

Temporal predictability is not itself causal evidence. An autoregressive coefficient summarizes conditional dependence on past values; it does not identify the effect of an intervention. Regression coefficients can also be affected by omitted variables, feedback, and common trends. Likewise, precise model-based intervals describe uncertainty within a specified probabilistic structure, not protection against every future disruption. When the data-generating relationships change, a previously well-diagnosed ARIMA model can retain internally coherent calculations while losing external predictive accuracy.