Stochastic Time Series and State-Space Models

Introduction and Context

A stochastic time series is a sequence of observations indexed by time whose future values are uncertain. Examples include daily sales, hourly electricity demand, asset returns, customer arrivals, inflation, and machine-sensor measurements. Unlike a cross-sectional dataset, a time series contains temporal dependence: observations close in time may be statistically related.

Stochastic time-series models provide a language for separating predictable structure from random innovation. They support forecasting, simulation, anomaly detection, risk measurement, inventory planning, and operational decision-making. This chapter develops the principal tools used in business analytics: stationarity, autocorrelation, ARIMA models, volatility models, state-space representations, hidden Markov models, Kalman filtering, regime switching, probabilistic forecasting, and forecast uncertainty.

The central modeling challenge is that the data-generating process is only partially observable. A measured series may combine trend, seasonality, changing variance, temporary shocks, and unobserved business conditions. State-space and regime-switching models are especially useful because they represent these hidden components explicitly.

Learning Objectives

By the end of this chapter, you should be able to:

1. Distinguish strict and weak stationarity and explain why stationarity matters.
2. Interpret autocorrelation and partial autocorrelation functions.
3. Specify and diagnose AR, MA, ARMA, and ARIMA models.
4. Explain conditional heteroskedasticity and estimate volatility models.
5. Represent dynamic systems in state-space form.
6. Explain the logic of Kalman filtering and smoothing.
7. Use hidden Markov and regime-switching models for discrete latent states.
8. Produce and interpret probabilistic forecasts and forecast intervals.
9. Evaluate model assumptions, uncertainty, governance risks, and practical limitations.

1. Foundations of Stochastic Time Series

Let Yₜ denote an observation at time t. A stochastic process is a collection {Yₜ: t ∈ T} of random variables. A realization is the particular sequence observed in the data. The information available through time t is commonly represented by the filtration ℱₜ, which includes current and past observations and any known explanatory variables.

A useful decomposition is

where mₜ is a trend or level component, sₜ is a seasonal component, and εₜ is an irregular component. This additive representation is appropriate when seasonal fluctuations have approximately constant absolute size. A multiplicative decomposition, Yₜ = mₜsₜεₜ, is often more suitable when seasonal variation grows with the level; logarithms can convert it approximately into an additive form.

A shock or innovation is new information arriving at time t. In many models, innovations satisfy E(εₜ | ℱₜ₋₁) = 0. This means that, conditional on past information, the shock has zero expected value. It does not mean that shocks are independent, normally distributed, or constant in variance.

Stationarity

Strict stationarity means that the joint distribution of any collection of observations is unchanged by a time shift. In practice, analysts frequently use weak, or covariance, stationarity. A process is weakly stationary if:

• E(Yₜ) = μ, a constant independent of t;
• Var(Yₜ) = γ(0), finite and constant;
• Cov(Yₜ, Yₜ₋ₖ) = γ(k), which depends only on the lag k and not on calendar time.

The autocorrelation at lag k is

where γ(k) is the lag-k autocovariance. The formula applies when the variance is finite and nonzero. Autocorrelation is dimensionless, whereas covariance retains the squared measurement units of Y.

Stationarity matters because many estimation and forecasting results rely on a stable relationship between past and future observations. A trending series can produce high correlations even when its changes are unrelated. This is the problem of spurious regression.

A unit root is a common source of nonstationarity. For the first-order autoregression

with white-noise innovations εₜ, the process is covariance stationary when |φ| < 1. If φ = 1, the model becomes a random walk with drift, and shocks have permanent effects. If |φ| > 1, the process is explosive and generally unsuitable as a stable description of ordinary business data.

First differencing is defined as ∆Yₜ = Yₜ − Yₜ₋₁. A series may become stationary after d differences; such a series is said to be integrated of order d, written I(d). Differencing should be used carefully: excessive differencing can amplify noise and create unnecessary dependence.

Autocorrelation and Partial Autocorrelation

The sample autocorrelation at lag k estimates the relationship between observations separated by k periods. The autocorrelation function (ACF) is the sequence of autocorrelations across lags. The partial autocorrelation at lag k measures the association between Yₜ and Yₜ₋ₖ after controlling for observations at intermediate lags.

Typical identification patterns are useful but not definitive:

Sampling variability means that apparent spikes should not be interpreted mechanically. A model must also pass residual diagnostics and perform well out of sample.

2. ARIMA Modeling

AR and MA Models

An autoregressive model of order p, AR(p), expresses the current value as a linear function of its past values and an innovation:

Here c is an intercept, φᵢ are autoregressive coefficients, and εₜ is an innovation with conditional mean zero. Stationarity requires the roots of the autoregressive polynomial to lie outside the unit circle. For AR(1), this reduces to |φ₁| < 1.

A moving-average model of order q, MA(q), uses current and lagged innovations:

The parameter μ is the mean under this convention. An MA(q) is covariance stationary for finite coefficients, but invertibility is normally imposed so that the unobserved innovations can be represented uniquely from past observations. Invertibility requires the roots of the moving-average polynomial to lie outside the unit circle.

An ARMA(p,q) combines both structures. It is appropriate for stationary series whose dependence is not adequately represented by a pure AR or pure MA model.

ARIMA and Seasonal Models

An ARIMA(p,d,q) model applies an ARMA(p,q) process to the d-th difference of a nonstationary series. Using the backshift operator B, where BYₜ = Yₜ₋₁, the compact form is

where φ(B) is the autoregressive polynomial and θ(B) is the moving-average polynomial. Sign conventions for θ vary across software, so model documentation must be checked.

Seasonal ARIMA models add seasonal differencing and seasonal AR and MA terms. A common notation is SARIMA(p,d,q)(P,D,Q)ₛ, where s is the seasonal period, such as 12 for monthly data with annual seasonality or 7 for daily data with a weekly cycle. Seasonal differencing is ∆ₛYₜ = Yₜ − Yₜ₋ₛ.

Exogenous predictors can be included in ARIMAX models. Examples include promotions in demand forecasting, temperature in electricity demand, and interest rates in financial applications. Predictors must be known or forecast separately over the forecast horizon; otherwise the uncertainty of their future values is omitted.

Identification, Estimation, and Diagnosis

A disciplined workflow is:

1. Plot the series and inspect transformations, missing values, outliers, trend, and seasonality.
2. Consider a variance-stabilizing transformation, such as log(Yₜ) for positive data.
3. Difference only as needed to obtain a plausible stationary series.
4. Use ACF and PACF as preliminary identification tools.
5. Estimate several defensible candidate models using maximum likelihood or conditional least squares.
6. Compare information criteria such as AIC or BIC, while recognizing that lower in-sample criteria do not guarantee better forecasts.
7. Examine residuals and perform rolling-origin forecast evaluation.

For a fitted model, residuals should have approximately zero mean, no meaningful autocorrelation, and an appropriate variance structure. The Ljung–Box statistic tests the joint null hypothesis that a set of residual autocorrelations is zero. A significant result indicates remaining serial dependence, although the test can be sensitive in large samples. Residual normality is less important for point forecasts than for accurate interval forecasts and likelihood-based inference.

Worked Example: AR(1) Forecast

Suppose a stationary demand-deviation series follows Yₜ = 0.4Yₜ₋₁ + εₜ, with current observation Yₜ = 10 and E(εₜ₊₁ | ℱₜ) = 0. The one-step forecast is

The two-step forecast is

More generally, for an AR(1) with zero mean, Ŷₜ₊h|ₜ = φʰYₜ. Since |0.4| < 1, the forecast converges to zero as the horizon increases. If the model had a nonzero mean μ, the forecast would converge to μ instead.

3. Volatility and Conditional Heteroskedasticity

Many business and financial series have changing variance. Large changes tend to cluster with large changes, and quiet periods tend to follow quiet periods. This phenomenon is called volatility clustering. A series may have uncorrelated returns but strongly autocorrelated squared returns.

A basic ARCH(q) model specifies the conditional variance as

where σₜ² = Var(εₜ | ℱₜ₋₁), ω > 0, and αᵢ ≥ 0. A GARCH(1,1) model is

For positivity under the standard specification, ω > 0 and α, β ≥ 0. A sufficient condition for covariance-stationary variance is α + β < 1. The closer α + β is to one, the more persistent volatility is. Under a zero-mean innovation process, the unconditional variance is ω ÷ (1 − α − β), provided α + β < 1.

The conditional mean and variance should be modeled separately but jointly estimated when appropriate. Innovations may be Gaussian, Student-t, or another distribution. Heavy-tailed distributions can improve risk estimates, but their use must be validated rather than selected solely because they fit historical extremes.

Extensions include EGARCH, which models log variance and can represent asymmetric responses to positive and negative shocks, and GJR-GARCH, which adds an indicator for the sign of past innovations. Volatility models are useful for portfolio risk, option pricing, credit exposure, capacity planning, and margin requirements. They can fail during structural breaks, market closures, illiquidity, or unprecedented events.

4. State-Space Representations

A state-space model separates an unobserved state from the observed measurement. In a linear Gaussian model,

where xₜ is the latent state vector, yₜ is the observation vector, uₜ contains known inputs, Fₜ is the transition matrix, Hₜ is the observation matrix, and wₜ and vₜ are process and measurement errors. Usually wₜ ∼ N(0,Qₜ), vₜ ∼ N(0,Rₜ), and the errors are mutually independent and independent over time. Qₜ describes uncertainty in state evolution; Rₜ describes measurement noise.

State-space models can represent local level, local linear trend, seasonal effects, time-varying regression coefficients, and dynamic factor models. For example, a local-level model is

The latent level changes gradually, while observations fluctuate around it. The ratio of process noise to measurement noise determines how quickly the estimated level responds to new data.

A major advantage is that state-space models accommodate missing observations, irregularly observed components, time-varying parameters, and online updating. They also provide a coherent framework for likelihood evaluation and forecasting.

5. Kalman Filtering and Smoothing

The Kalman filter recursively estimates the state using information available through the current time. Let aₜ|ₜ₋₁ be the predicted state and Pₜ|ₜ₋₁ its prediction-error covariance. The prediction step is

The forecast error, or innovation, is

with innovation covariance

The Kalman gain is

and the update equations are

These equations assume a correctly specified linear model and finite, nonsingular covariance matrices. The gain is large when the observation is reliable relative to the prior state estimate, and small when measurement noise is high. A smoother uses observations after time t as well as before time t to estimate a past state; smoothing is therefore retrospective, whereas filtering is suitable for real-time decisions.

If errors are non-Gaussian, the Kalman recursion still provides a linear minimum mean-square estimator under suitable conditions, but exact conditional distributions are no longer generally Gaussian. Nonlinear systems may require extended or unscented Kalman filters, particle filters, or simulation-based methods.

6. Hidden Markov and Regime-Switching Models

A hidden Markov model (HMM) contains a discrete latent state Sₜ taking values in {1,…,K}. The Markov property is

The transition probabilities are pᵢⱼ = P(Sₜ = j | Sₜ₋₁ = i), and each row of the transition matrix sums to one. Conditional on the state, the observation follows an emission distribution, such as Yₜ | Sₜ = j ∼ N(μⱼ, σⱼ²).

The forward algorithm recursively calculates the likelihood of observations and the probability of the current state. The Viterbi algorithm identifies the most probable state sequence. The Baum–Welch algorithm, an expectation–maximization procedure, estimates unknown transition and emission parameters.

Regime-switching time-series models allow parameters of an AR or regression process to depend on Sₜ. A business application might define regimes such as expansion and contraction, or low-, medium-, and high-volatility states. State probabilities can be filtered in real time and smoothed after the full sample is observed.

HMMs and state-space models are related but distinct. HMMs generally use discrete latent states, while classical Kalman models use continuous latent states and linear Gaussian dynamics. Hybrid models can combine discrete regimes with continuous state evolution.

Interpretation requires caution. Estimated regimes are statistical constructs, not automatically genuine economic phases. Label switching can occur: swapping the names of two states leaves the likelihood unchanged. Constraints, economic interpretation, and stability checks are needed.

7. Probabilistic Forecasting and Uncertainty

A point forecast summarizes a predictive distribution but does not describe its risk. A probabilistic forecast estimates

the distribution of the future observation at horizon h conditional on information available at t.

Forecast uncertainty has several sources:

• Innovation uncertainty: future shocks are unknown.
• Parameter uncertainty: estimated coefficients are not known exactly.
• State uncertainty: latent levels or regimes are imperfectly estimated.
• Exogenous-input uncertainty: future predictor values may be unknown.
• Model uncertainty: several plausible data-generating structures may exist.

For a correctly specified AR(1) model with zero mean and innovation variance σ², the h-step forecast-error variance is

When |φ| < 1, this approaches σ² ÷ (1 − φ²). This formula treats parameters as known and assumes independent innovations with constant variance. Estimated-parameter uncertainty makes actual predictive uncertainty larger.

A central 95% forecast interval under a Gaussian predictive distribution is approximately

where Vₜ₊h is the forecast-error variance. The 1.96 multiplier applies to a normal approximation; it is not universally valid. For skewed, heavy-tailed, bounded, intermittent, or regime-switching outcomes, simulation or quantile-based methods are preferable.

Forecast evaluation should match the decision objective. Point forecasts can be assessed using MAE and RMSE:

For probabilistic forecasts, calibration and sharpness are essential. Calibration means that stated probabilities correspond to observed frequencies. Sharpness means that forecasts are concentrated when justified by the information. Proper scoring rules include the logarithmic score and the continuous ranked probability score. A forecast that is narrow but systematically misses actual values is not useful.

8. Illustrative Case: Retail Demand with Hidden Operating Regimes

The following is a fictional illustrative case.

Background

A retailer sells a seasonal consumer product through stores and an online channel. Daily demand is affected by promotions, holidays, stock availability, and changing customer traffic. Managers need forecasts for replenishment seven days ahead. Stockouts are costly, but excessive inventory creates markdowns.

Problem

A preliminary ARIMA model produced reasonable point forecasts but failed during promotional periods. Residual variance increased sharply during campaigns, and the model treated unusually high demand as isolated noise. Managers also suspected that demand behavior differed between ordinary weeks and promotional events.

Analysis

The analytics team first separated days with known stockouts and treated observed sales as censored demand where appropriate. A log transformation was considered because variance increased with the level. Weekly seasonality was modeled with seasonal terms and promotion indicators. Residual ACF showed little remaining serial dependence, but squared residuals displayed clustering, suggesting conditional heteroskedasticity.

The team compared three models:

1. Seasonal ARIMAX with promotion and holiday predictors.
2. A dynamic regression model with a local-level state and Kalman filtering.
3. A two-regime switching model in which the mean and variance differed between ordinary and promotional regimes.

The regime model estimated transition probabilities rather than assuming that every promotion day had identical behavior. The predictive distribution was generated by weighting forecasts from each possible future regime according to filtered transition probabilities. This produced wider intervals when a regime transition was plausible.

Approach and Solution

The final system used a dynamic regression with a latent level, observed promotional variables, and a separate volatility component. A regime-switching specification was retained as a challenger model. Forecasts were generated through rolling-origin evaluation, with each training window containing only information available at the forecast date.

The business decision used a quantile forecast rather than the mean alone. If the cost of a stockout is higher than the cost of excess inventory, the selected replenishment target should be above the median. In general, the optimal order quantity for a single-period inventory decision is associated with the critical fractile

where Cᵤ is the underage cost and Cₒ is the overage cost. This relationship assumes a one-period decision, known costs, and a predictive distribution for demand.

Outcome

The probabilistic system reduced stockout frequency during promotions while making uncertainty visible to planners. The largest improvement came not from a more complex algorithm alone, but from correcting stockout treatment, including known predictors, and validating forecasts under realistic rolling-time conditions.

Learning Points

The case illustrates that: data quality can dominate model selection; changing variance should be modeled rather than ignored; latent regimes are useful but require interpretation; forecast intervals must reflect operational uncertainty; and a forecast should be evaluated against the decision it supports.

9. Implementation, Governance, and Limitations

Model selection and validation

Time order must be preserved in validation. Randomly shuffling observations allows future information to enter the training set and can produce misleadingly optimistic results. Rolling-origin evaluation repeatedly trains on an earlier period and tests on a later period. The evaluation design should reflect the actual forecast horizon, update frequency, missing-data process, and decision costs.

Models should be benchmarked against simple alternatives such as the naïve forecast, seasonal naïve forecast, drift model, and recent-history averages. Complexity is justified only when it improves out-of-sample performance, reliability, or decision value.

Data and operational issues

Important practical issues include time-zone conversion, irregular timestamps, daylight-saving changes, revisions to economic data, delayed reporting, missingness, outliers, structural breaks, and leakage from variables unavailable at forecast time. Aggregation can conceal intraday dependence, while disaggregation can increase noise.

Ethics and governance

Forecasts can influence staffing, credit, pricing, inventory, and resource access. Governance should document data provenance, transformations, model versions, assumptions, validation windows, and responsible owners. Analysts should distinguish predictive association from causal effect. A model trained on historical allocations may reproduce discriminatory decisions. Protected characteristics and their proxies require careful review when forecasts affect individuals or communities.

Probabilistic forecasts should communicate uncertainty rather than presenting a single number as fact. Automated decisions need monitoring, escalation procedures, audit trails, and human review for high-impact cases. Privacy protection is required when time-stamped data can identify customers, employees, or households.

Contemporary developments

Modern forecasting systems increasingly combine statistical models with machine learning. Gradient-boosted trees and neural networks can model nonlinear covariate effects, while temporal fusion and transformer-style architectures can represent long-range dependencies. Nevertheless, classical stochastic models remain valuable because their assumptions, uncertainty mechanisms, and updating procedures are comparatively transparent.

Ensemble forecasting combines predictions from multiple models. Bayesian dynamic models incorporate parameter and model uncertainty through posterior distributions. Hierarchical forecasting reconciles predictions across products, regions, and total-company levels. Conformal prediction can provide distribution-free prediction sets under exchangeability or related assumptions, although time dependence requires specialized methods and careful coverage assessment.

Best-Practice Checklist

• Define the forecasting decision, horizon, and information set before modeling.
• Inspect plots, transformations, calendar effects, missingness, and outliers.
• Test stationarity using substantive reasoning and statistical diagnostics rather than one test alone.
• Use ACF and PACF to guide, not dictate, ARIMA identification.
• Check both residual autocorrelation and residual variance dynamics.
• Treat parameter, state, input, and model uncertainty explicitly where material.
• Validate chronologically using realistic rolling-origin designs.
• Compare with simple benchmarks.
• Report calibrated intervals or quantiles, not only point forecasts.
• Monitor drift, regime changes, forecast errors, and downstream decisions after deployment.
• Preserve documentation, reproducibility, privacy, and human accountability.

Chapter Summary

Stochastic time-series analysis models dependence, uncertainty, and changing behavior over time. Stationarity provides the stability needed for many classical results, while autocorrelation and partial autocorrelation assist model identification. AR, MA, ARMA, and ARIMA models describe serial dependence; seasonal and exogenous extensions incorporate calendar structure and known drivers. ARCH and GARCH models represent time-varying conditional variance.

State-space models separate observations from latent dynamic states and provide a flexible framework for missing data, time-varying levels, and online estimation. The Kalman filter updates latent-state estimates recursively under linear Gaussian assumptions, while smoothing uses later information retrospectively. Hidden Markov and regime-switching models represent discrete latent conditions and allow behavior to change across regimes.

A responsible forecasting system estimates predictive distributions rather than only point predictions. Forecast uncertainty arises from future shocks, unknown parameters, latent states, future inputs, and model ambiguity. Calibration, sharpness, rolling-origin validation, governance, and decision-based evaluation are as important as statistical fit. The most useful model is therefore not necessarily the most complex, but the one that is reliable, interpretable enough for its context, and aligned with the decisions it supports.

Key Terms

Stationarity: Stability of a process’s relevant distributional or moment properties over time.

Weak stationarity: Constant mean and variance with autocovariance depending only on lag.

Innovation: New information or shock at a time point, often with conditional mean zero.

Autocorrelation function: Autocorrelations between observations at different lags.

Partial autocorrelation: Correlation at a lag after controlling for intermediate lags.

Unit root: A root on the unit circle that commonly produces nonstationarity.

Differencing: Subtracting a previous observation to remove level or trend behavior.

ARIMA: A model combining autoregression, differencing, and moving-average terms.

Seasonal differencing: Subtracting an observation from one complete seasonal cycle earlier.

Conditional heteroskedasticity: Time-varying conditional variance.

GARCH: A model in which current variance depends on past squared innovations and past variance.

State-space model: A model separating latent state dynamics from an observation equation.

Kalman filter: A recursive estimator for states in linear dynamic systems.

Kalman smoother: A retrospective state estimator using observations from the full sample.

Hidden Markov model: A model with discrete latent states following a Markov chain.

Regime switching: A model in which parameters change according to an unobserved regime.

Predictive distribution: The probability distribution assigned to a future outcome.

Calibration: Agreement between forecast probabilities and observed frequencies.

Forecast-error variance: Variance of the difference between a future value and its forecast.

Rolling-origin evaluation: Time-ordered repeated training and testing for forecast assessment.

Review Questions

5 Short Answer Questions

1. What distinguishes weak stationarity from strict stationarity?
2. Why can differencing improve an ARIMA model, and why can excessive differencing be harmful?
3. What information do the ACF and PACF provide?
4. What is the purpose of the Kalman gain?
5. Name four sources of forecast uncertainty.

5 Long Answer Questions

1. Explain the identification, estimation, and diagnostic workflow for an ARIMA model.
2. Compare ARCH and GARCH models, including their assumptions and interpretation.
3. Explain the prediction and updating stages of the Kalman filter.
4. Compare hidden Markov models with continuous linear Gaussian state-space models.
5. Discuss why probabilistic forecasting is more informative than point forecasting for operational decisions.

3 Analytical / Application Questions

1. A monthly sales series has increasing variance, strong annual seasonality, and a slow-decaying ACF. Propose a transformation, differencing strategy, candidate model class, and diagnostic plan. Justify each choice.
2. A retailer’s forecast intervals are narrow but actual demand falls outside the nominal 95% interval 18% of the time. Diagnose the likely issue and recommend corrective actions.
3. A financial-return series has little autocorrelation in returns but strong autocorrelation in squared returns. Explain what this pattern implies, propose a model, and identify limitations during a structural market change.

Multiple Choice Questions with Answers

Question 1

Which condition is required for covariance stationarity of an AR(1) process with coefficient φ?

A. φ = 1

B. |φ| < 1

C. φ > 1

D. φ = 0 only

Correct Answer: B. |φ| < 1

Question 2

For an MA(q) process, which function typically cuts off after lag q?

A. The ACF

B. The unconditional mean

C. The variance

D. The sample size

Correct Answer: A. The ACF

Question 3

In a standard GARCH(1,1) model, what does α + β close to one generally indicate?

A. Very short-lived volatility

B. Volatility persistence

C. Zero conditional variance

D. Guaranteed normality

Correct Answer: B. Volatility persistence

Question 4

What is the primary role of the Kalman filter?

A. To estimate latent states recursively as observations arrive

B. To remove all measurement error permanently

C. To calculate only unconditional means

D. To identify seasonal periods without data

Correct Answer: A. To estimate latent states recursively as observations arrive

Question 5

In a hidden Markov model, what is hidden?

A. The observed time index

B. The discrete state generating the observation behavior

C. The sample size

D. The definition of variance

Correct Answer: B. The discrete state generating the observation behavior

Question 6

For a zero-mean AR(1) model with φ = 0.5 and current value Yₜ = 8, what is the two-step forecast?

A. 2

B. 4

C. 8

D. 16

Correct Answer: A. 2

Question 7

Which validation method is most appropriate for evaluating a time-series forecast?

A. Randomly shuffling all observations

B. Using only the in-sample R²

C. Rolling-origin time-ordered evaluation

D. Training on future observations and testing on the past

Correct Answer: C. Rolling-origin time-ordered evaluation

Question 8

What does forecast calibration mean?

A. Forecast intervals are always as narrow as possible

B. Forecast probabilities agree with observed frequencies

C. The model has no parameters

D. Every forecast equals the sample mean

Correct Answer: B. Forecast probabilities agree with observed frequencies

Question 9

In a local-level state-space model, what does measurement noise represent?

A. Random variation in the observed value around the latent level

B. The transition probability between regimes

C. The seasonal period

D. The ARIMA differencing order

Correct Answer: A. Random variation in the observed value around the latent level

Question 10

If a model’s nominal 95% prediction interval contains only 70% of realized outcomes, what is the most direct interpretation?

A. The intervals are under-dispersed or poorly calibrated

B. The model is perfectly calibrated

C. The forecast horizon is necessarily zero

D. The point forecast must equal the median

Correct Answer: A. The intervals are under-dispersed or poorly calibrated