IS6400 Business Data Analytics
IS6400 Overview
- City University of Hong Kong
- Semester A, 2026/27
- Postgraduate
- Three credit units
- Two-hour final examination
What this course asks you to do
Business Data Analytics joins statistical and machine-learning methods to decisions in finance, marketing, accounting and other business settings.
- Assessment Continuous assessment contributes 60%, and the two-hour final examination contributes 40%. Any separate attendance or class-participation minimum is unpublished; confirm the live rule on the course Canvas page.
- Pass condition Both coursework and examination must be passed for an overall pass. Any numeric attendance threshold is unpublished. Confirm the live thresholds and participation rule on the course Canvas page.
- Core method Define the business decision and time boundary before selecting or tuning a model.
What you get in the full IS6400 Exam Bible
- 10 chapters, 34 pages — every examinable topic
- 23 original practice questions, each with a model answer in the Exam format
- 10 worked examples, solved the way IS6400 marks
- 11 computed figures, redrawn from the course’s own numbers
- Exam-morning plan for the Exam (40%) — 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 IS6400 is assessed
| Component | Weight | Format |
|---|---|---|
| Class Participation | 5% | Randomly selected in-class quizzes and participation |
| Individual Assignments | 25% | About eight assignments linked to lectures and tutorials |
| Group Project Briefing | 5% | Proposal briefing |
| Group Project Presentation | 15% | Presentation and Q&A |
| Group Project Report | 10% | Analytical report |
| Exam | 40% | Two-hour final covering readings, lectures, tutorials, cases and examples |
| Coursework | Pass hurdle: coursework must be passed for an overall pass; threshold unpublished |
The current Semester A syllabus publishes six weighted components totalling 100%. The CityU catalogue corroborates the 60% continuous-assessment and 40% examination split, publishes a two-hour duration, and states that both coursework and examination must be passed. The numeric pass thresholds, permitted examination materials and any separate attendance or class-participation minimum are unpublished. Confirm the live thresholds, participation rule and examination instructions on the course Canvas page.
Assessment structure
The six numeric rows total 100%. Both coursework and examination must be passed.
What other students are asking about IS6400
- Where can I find lecture notes or study guides for IS6400 at CityU?
- The IS6400 study guide in AskSia Library covers Business Data Analytics in 10 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 IS6400, and how much is it worth?
- Yes. In the offering this AskSia Library guide is built from, Exam carries 40% of the IS6400 mark; the full split is Class Participation 5%, Individual Assignments 25%, Group Project Briefing 5%, Group Project Presentation 15%, Group Project Report 10%, Exam 40%.
- What does IS6400 cover on multivariate regression for prediction?
- In AskSia Library's IS6400 guide, the Multivariate Regression for Prediction chapter explains conditional coefficients, residual direction and business interpretation.
- What does IS6400 cover on feature selection and principal components?
- In AskSia Library's IS6400 guide, the Feature Selection and Principal Components chapter explains training-only scaling, component scores and retained meaning.
- What does IS6400 cover on kmeans and density-based clustering?
- In AskSia Library's IS6400 guide, the KMeans and Density-Based Clustering chapter explains centroid splits, density neighbourhoods and unassigned noise.
Current IS6400 dates
| Date | Item | Control |
|---|---|---|
| 11 September 2026 | Week 1 individual assignment | Due 23:59 on the current course Canvas page. |
| 25 September 2026 | Week 3 individual assignment | Due 23:59 on the current course Canvas page. |
| End of Week 5 | Group Project Briefing | The project guidance places the proposal briefing at the end of Week 5. |
| Weeks 12 to 13 | Group Project Presentation | The project guidance places final presentations in Weeks 12 to 13. |
The current Semester A course Canvas pages and project guidance state these dates and teaching weeks. Confirm live submission settings on Canvas.
What IS6400 covers
Ten chapters follow the published sequence from the analytics workflow through regression, feature reduction, clustering, forecasting and deep learning to model assessment and generative AI.
Business Analytics Workflow
business understanding, data preparation, evaluation and deploymentBegin with the decision
The course places business understanding, data understanding and preparation before model building. That order prevents a team from optimising a score for a target that does not match an operational choice.
- Business understanding and an operational owner
- One row and one prediction horizon
- Data understanding before feature construction
- Preparation without future leakage
- Testing against a baseline
- Deployment, monitoring and backtracking
Multivariate Regression for Prediction
conditional coefficients, residual direction and business interpretationCoefficients are conditional changes
A multivariate regression prediction adds a baseline and feature contributions. A coefficient states the expected target change for one unit of its feature while the other listed features are held fixed.
- Row definition and feature availability
- Baseline plus feature contributions
- Coefficient units and conditional interpretation
- Observed minus predicted residuals
- Residual plots by time and segment
- Prediction versus causal claims
Feature Selection and Principal Components
training-only scaling, component scores and retained meaningSelection keeps names; PCA constructs directions
Feature selection retains a subset of original variables, which supports direct explanation and cheaper collection.
- Operational filtering before statistical reduction
- Selection versus feature construction
- Training-only standardisation
- Weighted component scores
- Loadings and interpretable directions
- Validation on the downstream decision
KMeans and Density-Based Clustering
centroid splits, density neighbourhoods and unassigned noiseDifferent algorithms define groups differently
Bisecting KMeans repeatedly divides a cluster into compact groups around centres. DBSCAN connects records through dense neighbourhoods and can leave isolated records unassigned as noise.
- Similarity as a business definition
- Bisecting compact centroid groups
- Density radius and neighbour count
- Core, border and noise records
- Scaling before distance calculation
- Stability and action-based profiling
Decision Trees and Ensemble Learning
impurity reduction, pruning, bagging and stable decisionsA tree partitions with questions
Each split seeks child nodes that are less mixed in the target than their parent. Gini impurity is zero for a pure node and highest for balanced binary classes.
- Recursive feature splits
- Gini impurity and weighted reduction
- Depth and minimum leaf controls
- Pruning against validation evidence
- Bagging and random feature subsets
- Boosting and threshold-specific evaluation
Holt Trend Forecasting
level and trend updates across the operational horizonTwo states describe a moving series
Holt's method maintains a local level and a local trend. Each new observation updates the level using the previous level advanced by the old trend, then updates trend using the new movement in level.
- Chronological data splits
- Local level state
- Local trend state
- Smoothing weights
- Forecast horizon
- Rolling backtests and signed bias
Exponential Smoothing and Time Series
decaying weights, responsiveness and chronological validationRecent evidence receives more weight
Simple exponential smoothing blends the latest observation with the previous level. Expanding the update reveals geometrically declining influence from older observations.
- Latest observation and prior level
- Weights that sum to one
- Geometrically decaying history
- Responsiveness versus noise
- Chronological backtesting
- Seasonal and regime residual patterns
Neural Networks and Recurrent Models
weighted activations, sequence state and leakage-safe splitsA network starts with weighted sums
Each neural unit multiplies inputs by weights, adds a bias and applies an activation. Layers repeat that calculation to represent nonlinear combinations.
- Weighted sums and bias
- ReLU and sigmoid activations
- Layer dimensions
- Training loss and validation
- Recurrent hidden state
- Entity and time leakage
Image Mining and Recognition
capture context, feature maps and error-cost thresholdsPixels become features and scores
Image models learn local filters, combine them into feature maps, reduce spatial detail and produce class scores.
- Pixel grids and local filters
- Feature maps and pooling
- Capture-context shortcuts
- Session-level data splits
- Confusion counts
- Threshold cost and review capacity
Model Assessment and Generative AI
held-out evidence, business costs and monitored generationSelection requires one fixed protocol
Compare candidates on the same row definition, target, horizon, split, preprocessing and metric. Validation supports tuning, while one final holdout estimates performance after choices are complete.
- Training, validation and final holdout
- Baseline comparisons
- Metrics under class imbalance
- Calibration and threshold cost
- Generative task completion
- Groundedness, review effort and monitoring
The published aims move through a complete chain: understand a critical analysis problem, manage appropriate statistical and machine-learning models, implement them with software such as Python and generative AI tools, then analyse and interpret the output for decision making. That final interpretation is not an optional paragraph.
It determines whether a technically correct model has produced information a business can act on.
How the teaching sequence fits together
The course begins with the analytics workflow and then moves through multivariate regression, feature selection and principal components, clustering, decision trees and ensembles, time-series forecasting, artificial and recurrent neural networks, image recognition, model assessment and generative AI.
The sequence increases model flexibility while repeatedly returning to the same questions: what one row represents, when each field is available, how future data is protected from training, which error matters, and how a score changes a decision.
How difficulty shows up in facts
The final carries 40% and lasts two hours.
The official catalogue says it covers assigned readings, lectures, tutorials, cases and examples, while the available official information does not publish sections, question types or permitted aids. The other 60% includes participation, individual assignments and a three-part group project. CityU also publishes a separate pass condition: both coursework and the examination must be passed.
The numeric thresholds are unpublished, and no separate attendance or class-participation minimum is published in the available information; confirm the live rules on the course Canvas page.
Revision therefore needs both whole-course method retrieval and repeated numeric practice, not a narrow bet on an invented paper format.
Where the marks are earned
The current information provides no populated question rubric or official worked solutions. This guide therefore does not claim a University mark scheme.
Its original examples make the reasoning visible: define the unit and horizon, substitute values with units, check signs and denominators, compare a baseline under the same split, interpret the error pattern and finish with an owned action or monitoring trigger. The same sequence strengthens the project briefing, presentation and report because it connects analytical evidence to a manager's decision.
Worked example · free
Choose a model by decision cost
- 2Construct the four confusion counts. The mark allocation shown here is not a University assessment scheme.
- 3Compute precision, recall and accuracy with named denominators.
- 3Connect the threshold to review capacity and missed-fraud cost.
Key terms
- Business Understanding
- The workflow stage that turns a broad objective into a defined decision, unit, target, horizon and success measure.
- Residual
- Observed value minus predicted value; its sign states whether a model under-predicted or over-predicted.
- Principal Component
- A weighted combination of features chosen to represent a direction of high variance after an appropriate transformation.
- Core Point
- In density clustering, a record with enough neighbours inside the chosen distance radius to anchor a cluster.
- Gini Impurity
- A measure of class mixture used to compare tree splits, equal to one minus the squared class proportions.
- Ensemble Learning
- Combining varied models by averaging, voting or sequential correction to improve stability or reduce systematic error.
- Local Level
- A time-series state representing the current baseline around which recent observations move.
- Smoothing Parameter
- A weight controlling how strongly a time-series update responds to new evidence rather than the previous state.
- Activation Function
- A nonlinear transformation applied after a neural unit's weighted sum and bias.
- Recurrent State
- Information carried from one sequence position to the next so a prediction can depend on earlier positions.
- Confusion Matrix
- Counts of correct and incorrect predictions by actual and predicted class, used to derive threshold-specific metrics.
- Model Calibration
- Agreement between predicted probabilities and observed event frequencies at corresponding score levels.
IS6400 FAQ
How is the course assessed?
The current Semester A syllabus assigns 5% to class participation, 25% to individual assignments, 5% to the project briefing, 15% to the project presentation, 10% to the project report and 40% to the final examination. The numeric rows total 100%, and the project components together contribute 30%.
How long is the final examination and what does it cover?
The CityU catalogue publishes a two-hour duration and says the final assesses competence in the taught subjects, covering assigned readings together with lectures, tutorials, cases and examples mentioned in class. The available official information does not publish sections or question types, so prepare whole-course methods and applications.
Is there a separate pass condition?
Yes. The CityU catalogue states that both coursework and examination must be passed to obtain an overall pass. The numeric threshold for either side and any separate attendance or class-participation minimum are unpublished. Confirm the operational thresholds, participation rule and any update on the course Canvas page.
Is the final open book or closed book?
The available official information does not resolve permitted materials. A Week 1 slide contains the unresolved wording open/closed book, while the catalogue and current syllabus do not choose either status. Treat permitted aids as a live examination instruction and confirm them on Canvas rather than inferring a rule.
What should the group project demonstrate?
The project should connect a real business problem to an available data set, key variables, suitable analytical methods and a potential contribution. The final presentation is framed for managers, and the report should explain the analytical procedure and interpret results in language a lay reader can understand.
What is the safest way to choose between models?
Freeze the business unit, target, horizon, split and metric; compare a simple baseline and candidate models on the same held-out periods; inspect error distribution and calibration; price the errors at realistic thresholds; then include maintenance, explanation and monitoring in the selection rather than choosing the highest aggregate score.
Are the practice marks an official marking scheme?
No. The available assignment and quiz pages contain blank Canvas rubric shells rather than populated criteria, and no past paper or official worked solution is present. Any marks attached to AskSia practice are study allocations that make a solution auditable, not a claim about the University's question-level marking.
How should I revise quantitative methods efficiently?
For each chapter, redraw the relationship, state the business decision, recompute one example and answer one diagnostic case. Check units, signs, denominators, time order and the baseline. Finish each answer with what changes operationally and what failure would trigger review after deployment.
Questions students ask in each IS6400 chapter
- Why should the business question come before the model?
- Because the question fixes the row, target, horizon, costly error and action. Without those choices a model can optimise a technically valid score that cannot be used, or can learn from information that arrives after the real decision. Chapter 01 · Business Analytics Workflow
- Why is a regression coefficient not automatically a causal effect?
- Other factors can influence both the predictor and target, and the business may choose the predictor based on expected outcomes. Regression conditions on listed variables but does not by itself create the comparison required for causality. Chapter 02 · Multivariate Regression for Prediction
- When is selection better than PCA?
- Selection is usually better when the decision must name an original field, data collection is costly, or governance requires a direct explanation. PCA is attractive when correlated measurements repeat signal and compression matters more than original names. Chapter 03 · Feature Selection and Principal Components
- Why can two reasonable clustering methods disagree?
- They encode different group shapes. Centroid methods prefer compact groups around representatives and force assignments; density methods follow connected neighbourhoods and can leave sparse observations unassigned. The disagreement reflects assumptions, not necessarily an error. Chapter 04 · KMeans and Density-Based Clustering
- Why can a decision tree overfit even when every split looks sensible?
- Repeated splits can isolate tiny training-specific groups and capture noise. Control depth and leaf size, prune using validation evidence, and inspect whether the first rules remain stable across folds or time windows. Chapter 05 · Decision Trees and Ensemble Learning
- Why is a random train-test split unsafe for forecasting?
- It allows later observations to influence a model judged on earlier ones and fails to mimic deployment. Chronological splits and rolling origins preserve the direction in which information becomes available. Chapter 06 · Holt Trend Forecasting
- Does a larger alpha always improve a forecast?
- No. It responds quickly to genuine change and also follows random spikes more closely. Choose alpha on forward validation at the required horizon and inspect which events drive the largest misses. Chapter 07 · Exponential Smoothing and Time Series
- Why use an activation after the weighted sum?
- Without nonlinear activations, stacked linear layers collapse into another linear transformation. Activations allow the network to represent relationships that cannot be captured by one linear combination. Chapter 08 · Neural Networks and Recurrent Models
- Why split images by session instead of randomly?
- Near-duplicate frames from one session can otherwise appear on both sides, inflating performance. A session, device, site or time split better tests whether the model transfers to new capture conditions. Chapter 09 · Image Mining and Recognition
- Why keep a final holdout separate from validation?
- Validation results influence choices, so repeated use gradually adapts the pipeline to that data. A final untouched set provides a less biased estimate after model and threshold decisions are finished. Chapter 10 · Model Assessment and Generative AI
How to study for the exam
Use one business case across the course so that every method competes under the same decision. Write the row, target, horizon and costly error at the top of a page. For regression, trees, forecasting and neural models, compute one prediction and one error by hand. For PCA and clustering, explain what information the representation preserves and loses.
For image recognition and generative AI, define the capture or evidence boundary and inspect errors by context. Revise in three passes: retrieval without notes, calculation with every unit shown, then interpretation in one business sentence.
Keep a separate sheet of direction checks: residual equals actual minus predicted, negative trend lowers later forecasts, smoothing weights total one, child impurity is size-weighted, precision divides by flagged cases and recall divides by real positive cases. In the final week, compare models rather than studying them in isolation.
Choose a baseline, hold the split and horizon fixed, price the two error types, state the operating threshold and name the monitoring trigger. Confirm the live examination instructions and pass thresholds on Canvas because the current official information does not publish those details.