STAT8003 Time Series Forecasting
STAT8003 Overview
- The University of Hong Kong
- First Semester, 2026-27
- Postgraduate statistics
- Two-hour written examination
- R or Python
What this course asks you to see
Time series forecasting begins with dependence. The observations are ordered, the interval between adjacent indices has meaning, and earlier values or innovations can carry information into the present.
- Assessed by A written examination, assignments, a modelling project and a class test.
- Core method Establish stationarity, derive dependence, diagnose residuals and compare held-out forecasts.
- Hardest distinction Keep population model signatures separate from noisy sample ACF and PACF evidence.
- Where to confirm Check Moodle for the examination date, permitted materials and detailed format.
What you get in the full STAT8003 Exam Bible
- 8 chapters, 37 pages — every examinable topic
- 25 original practice questions, each with a model answer in the Written examination paper format
- 8 worked examples, solved the way STAT8003 marks
- 12 computed figures, redrawn from the course’s own numbers
- Exam-morning plan for the Written examination paper (50%) — what to revise last, where marks are lost
Free on this page: assessment table, chapter map, one worked example, key terms, FAQ. The rest is in the full exam bible.
How STAT8003 is assessed
| Component | Weight | Format |
|---|---|---|
| Written examination paper | 50% | Two-hour written examination |
| Assignments | 15% | Course assignments |
| Project | 20% | Time-series modelling project |
| Class test | 15% | In-term class test |
The current outline publishes these four weights, which total 100%. It does not publish a component hurdle or the detailed examination format; check Moodle for operational details.
What other students are asking about STAT8003
- Where can I find lecture notes or study guides for STAT8003 at HKU?
- The STAT8003 study guide in AskSia Library covers Time Series Forecasting in 8 chapters, with the assessment table, a chapter-by-chapter topic map, key terms and a worked example, free to read on this page.
- Is there a final exam in STAT8003, and how much is it worth?
- Yes. In the offering this AskSia Library guide is built from, Written examination paper carries 50% of the STAT8003 mark; the full split is Written examination paper 50%, Assignments 15%, Project 20%, Class test 15%.
- What does STAT8003 cover on moments, covariance and random walks?
- In AskSia Library's STAT8003 guide, the Moments, Covariance and Random Walks chapter explains mean, variance, autocovariance, autocorrelation and innovation overlap.
- What does STAT8003 cover on stationarity, white noise and differencing?
- In AskSia Library's STAT8003 guide, the Stationarity, White Noise and Differencing chapter explains strict and weak stationarity, white noise and transformed random walks.
- What does STAT8003 cover on linear processes and moving averages?
- In AskSia Library's STAT8003 guide, the Linear Processes and Moving Averages chapter explains causality, innovation weights, MA moments and the ACF cutoff.
What STAT8003 covers
Eight chapters move from time-indexed data and covariance through stationarity, MA, AR and ARMA theory to model checking and held-out forecast comparison.
Time Series Data and Dependence
observed records, random processes, regular spacing and past informationStart with the indexing contract
This chapter separates an observed time-ordered record from the stochastic process that models it.
- Regularly spaced observations
- Realisation and random process
- Time-index notation
- Deterministic trend model
- Stochastic dependence model
- Innovation information set
Moments, Covariance and Random Walks
mean, variance, autocovariance, autocorrelation and innovation overlapDescribe level, spread and dependence
This chapter builds the mean, variance, autocovariance and autocorrelation functions used throughout the course.
- Mean function
- Variance function
- Autocovariance function
- Autocorrelation function
- Random-walk moments
- Innovation overlap
Stationarity, White Noise and Differencing
strict and weak stationarity, white noise and transformed random walksCompare distributional and second-order stability
This chapter compares strict stationarity with the weak definition used in later calculations. Strict stationarity preserves every finite joint distribution after shifting all time indices.
- Strict stationarity
- Weak stationarity
- Finite-moment condition
- White noise
- Non-stationary random walk
- First differencing
Linear Processes and Moving Averages
causality, innovation weights, MA moments and the ACF cutoffBuild the innovation representation
This chapter defines linearity, causality and the general innovation representation.
- Linear process
- Causal representation
- Psi weights
- MA model convention
- MA moments
- ACF cutoff
Autoregressive Models and Root Conditions
characteristic polynomials, unit-circle roots and stationary expansionsDo not confuse an equation with its stationary process
This chapter treats an autoregressive equation as a model that still needs a stationary solution.
- AR model equation
- Characteristic polynomial
- Unit-circle root test
- Stationary mean
- Innovation expansion
- Recursive psi weights
Yule-Walker Equations and AR Moments
initial correlations, ACF recursion and innovation varianceCentre the process and derive the recurrence
This chapter derives the autocorrelation recursion for a stationary autoregressive process.
- Centred AR process
- Autocorrelation recursion
- Symmetry of correlations
- Yule-Walker system
- Lag-zero innovation term
- AR variance
ARMA Models and Invertibility
mixed feedback, shock polynomials, coefficient recursions and innovation recoveryKeep the two memory channels visible
This chapter combines autoregressive feedback and moving-average shock memory in one polynomial model.
- ARMA equation
- AR and MA polynomials
- Stationarity roots
- Psi coefficient matching
- Late-lag ACF recurrence
- Invertibility
Model Checking and Forecast Comparison
stationarity, ACF and PACF candidates, residual tests, selection and holdout forecastsBuild an auditable modelling chain
This chapter follows the published project workflow from raw record to held-out comparison.
- Time plot and transformation
- ACF and PACF candidates
- Parameter estimation
- Ljung-Box diagnostics
- Over-parameterisation check
- AIC selection
This guide first separates the realised record from the stochastic process used to model it, then builds the mean, autocovariance and autocorrelation functions that describe a process through time.
From stationarity to model signatures
Stationarity is the main simplifying condition.
You learn why a random walk fails it, why white noise supplies the reference pattern, and how differencing can turn an accumulated process into stationary increments. General linear processes then organise current and past innovations.
Moving-average models have finite shock memory and a population ACF that cuts off, while autoregressive models feed back past process values and require characteristic roots outside the unit circle.
From roots to forecasts
Yule-Walker equations turn AR coefficients into autocorrelations and variance.
ARMA models combine feedback with a finite innovation filter, while invertibility asks whether the shocks can be recovered from observed values. The final chapter follows the project workflow: inspect and transform the series, use sample ACF and PACF evidence to specify candidates, estimate, test residuals at several lags, compare adequate models and evaluate forecasts against values withheld from fitting.
Worked example · free
Choose an adequate model before comparing forecasts
- 3Separate adequacy from relative information-criterion ranking.
- 3Use the residual evidence to identify the defensible candidate.
- 2State how the holdout is used after selection.
Key terms
- Time Series
- A sequence of random variables ordered through regular time intervals.
- Autocovariance
- Covariance between values of one process at two times, indexed by their lag under stationarity.
- Autocorrelation
- Autocovariance standardised by the relevant process standard deviations.
- Weak Stationarity
- A constant-mean process whose autocovariance depends only on lag.
- White Noise
- An innovation sequence with zero mean, constant positive variance and no nonzero-lag autocorrelation under the course definition.
- Random Walk
- A process whose current level adds a new innovation to its previous level.
- Causality
- Dependence of a current value only on information available now or earlier.
- Moving Average Process
- A process formed from a finite linear combination of current and past innovations.
- Autoregressive Process
- A process that expresses the current value through past process values and a new innovation.
- Yule-Walker Equations
- Equations linking autoregressive coefficients to the process autocorrelation function.
- Invertibility
- The ability to recover innovations from current and past observed process values.
- Ljung-Box Test
- A residual diagnostic used to assess whether autocorrelation remains across a chosen group of lags.
STAT8003 FAQ
How is the course assessed?
The current outline assigns 50% to a two-hour written examination, 15% to assignments, 20% to a project and 15% to a class test. These numeric rows total 100%. The available outline does not label any component as a hurdle, and the detailed examination sections and permitted materials should be confirmed on Moodle.
Why is stationarity so important in time series forecasting?
Stationarity lets one observed record inform a stable mean and lag-dependence structure. Without it, a covariance estimated early in the series may not describe later behaviour. The course uses differencing and model conditions to obtain stationary processes before applying ACF patterns, Yule-Walker equations and stationary ARMA representations.
What is the main difference between an MA and an AR model?
An MA model uses a finite set of current and past innovations, so its population ACF cuts off after its order. An AR model feeds back earlier process values, requires a root condition for stationarity and usually has an ACF that tails away according to a recurrence rather than ending exactly.
How should ACF and PACF plots be used?
Treat sample ACF and PACF patterns as evidence for candidate orders after stationarity has been addressed. They fluctuate because the record is finite, so an apparent cutoff is not exact. Candidate models still need estimation, residual checking and comparison rather than selection from a plot alone.
What does the project workflow require?
The brief asks for a series longer than one hundred observations, a held-out final block, a time plot, stationarity checks and transformations, ACF and PACF candidates, parameter estimation, Ljung-Box tests at several lags, over-parameterisation checks, model selection, normality review and forecast comparison with the withheld truth.
Does a lower AIC always identify the best forecasting model?
No. AIC ranks fitted candidates by relative fit with a complexity penalty, but it does not certify residual adequacy. A model with remaining autocorrelation should be repaired or rejected before the information criterion is used to choose among adequate candidates. Held-out forecast performance is assessed only after that selection.
What examination details are currently published?
The available current outline states that the written examination lasts two hours and is worth 50% of the course result. It does not publish the date, section structure, question types or permitted materials in the captured source, so those operational details should be checked on the course Moodle page.
Why does invertibility matter for an ARMA model?
Invertibility lets the innovation at time t be written from current and past observations. That reverse representation makes the information sequence recoverable from data and supports residual analysis. It is distinct from stationarity, which is controlled by the AR polynomial and stabilises the process representation across time.
Questions students ask in each STAT8003 chapter
- Why does regular spacing matter?
- Regular spacing makes one lag correspond to the same elapsed time across the record. Without it, a one-row shift could mean one day in one place and a week elsewhere, so standard lag covariance and correlation calculations would mix unlike intervals. Chapter 01 · Time Series Data and Dependence
- Why are random-walk levels dependent if the shocks are independent?
- Each later level contains all earlier shocks. Two levels therefore share an accumulated block of innovations, and that overlap creates covariance even though distinct innovations are mutually independent. Chapter 02 · Moments, Covariance and Random Walks
- Can a process be strictly stationary but not weakly stationary?
- Yes. Strict stationarity does not guarantee that the mean and variance exist. An independent Cauchy sequence preserves its joint laws under shifts but lacks the finite moments required by weak stationarity. Chapter 03 · Stationarity, White Noise and Differencing
- Why is every finite MA process stationary?
- An MA model is a finite linear combination of identically distributed innovations with fixed coefficients. Its mean and variance are constant, and its covariance is determined by which coefficients overlap at each lag. Chapter 04 · Linear Processes and Moving Averages
- What is the AR(1) stationarity condition?
- For Zt=theta0+phi Zt-1+at, the characteristic root is 1/phi. Requiring that root outside the unit circle is equivalent to requiring the absolute value of phi to be strictly below one. Chapter 05 · Autoregressive Models and Root Conditions
- Why is symmetry needed in the first AR(2) equation?
- At lag one, the recurrence includes rho minus one. Autocorrelation symmetry identifies this with rho one, turning the expression into an equation that can be solved for the unknown initial correlation. Chapter 06 · Yule-Walker Equations and AR Moments
- Which part of an ARMA model controls stationarity?
- The roots of the AR characteristic polynomial control whether a unique stationary solution exists. Adding MA terms changes early covariance and innovation weights but does not replace the AR root condition. Chapter 07 · ARMA Models and Invertibility
- Why should residual tests be run at several lags?
- A single lag choice examines only one aggregation of residual autocorrelations. The project specifically asks for several K values so adequacy is not made to depend on one convenient window of dependence. Chapter 08 · Model Checking and Forecast Comparison
How to study for the exam
Build revision around decisions rather than a flat formula list. First rehearse the stationarity definitions and prove each verdict from mean and covariance or from the relevant root condition. Next derive MA moments by aligning innovation coefficients and AR moments through Yule-Walker equations. Keep a sign-convention box beside the MA formulas because this course writes the moving-average coefficients with minus signs.
Then practise mixed ARMA coefficient matching until the finite MA contribution and continuing AR recursion are visually distinct. Finish every session with one modelling-chain question: identify a transformation, nominate ACF and PACF candidates, explain a residual test, restrict AIC comparison to adequate models and protect the held-out values until final forecast evaluation.
On the day before the examination, confirm the University's current operational details on Moodle and use the exam-morning pages only for the mathematical method, since the captured outline does not publish examination sections or permitted materials.
Your AI Statistics tutor for STAT8003
Stuck on a hard STAT8003 question? Sia is AskSia’s AI Statistics tutor — ask any STAT8003 Time Series Forecasting question and get a clear, step-by-step explanation grounded in how the course is actually taught and assessed. Read this whole study guide free, then take your hardest questions to Sia.