ECON7880 Foundations in Big Data Analytics: Concepts and Techniques
The Complete Exam Bible · ECON7880 Overview
- Hong Kong Baptist University
- Semester 1, 2026-27
- Postgraduate taught course
- Economics and analytics
- Moodle course materials
What this course is asking you to connect
ECON7880 joins organisational analytics, Python implementation and statistical learning.
- Assessment shape Coursework, quizzes, a group presentation, a midterm test and a final examination all contribute to the published total.
- Hardest calculation Preserving denominators and weights while moving from node impurity, residuals or scores to a defensible model comparison.
- Hardest design choice Building a split and feature cutoff that imitate the point at which a real prediction will be made.
- How to prepare Recompute each worked result, then explain the error direction, validation boundary and operational decision in one short paragraph.
How ECON7880 is assessed
| Component | Weight | Format |
|---|---|---|
| Class Participation | 15% | Three in-class quizzes |
| Written Assignments | 30% | Three assignments |
| Group Project & Presentation | 15% | Fifteen minutes per group |
| Midterm Test | 20% | Three-hour test |
| Final Examination | 20% | Two-hour exam |
The five published weights total 100%. The available course files do not state a component hurdle, so this guide does not display one. The Moodle page controls live dates, submission settings and any updated examination instruction.
Assessment structure
Segment widths reproduce the official percentage weights in assessment-table order.
Current ECON7880 dates
| Date | Item | Control |
|---|---|---|
| 25 September 2026 | Assignment 1 and in-class quiz 1 | Shown on the captured Moodle course page. |
| 9 October 2026 | Assignment 2 | Shown on the captured Moodle course page. |
| 16 October 2026 | Midterm Test | Published on the captured midterm page. |
| 27 November 2026 | Group Project Presentation | Published in the project timeline and course outline. |
Dates are as published in the captured course outline and Moodle pages for this offering. Confirm exact deadlines and submission settings in the live LMS.
What ECON7880 covers
Eight chapters connect organisational analytics and reproducible Python work to decision trees, regression, margins, validation and decision-focused evaluation.
Analytics Questions and the CRISP-DM Cycle
business questions, prediction boundaries, CRISP-DM stages and deployment choicesStart with the decision
This chapter turns an organisational request into a population, prediction time, future target, feature cutoff and action.
- Business questions and decision owners
- Prediction time and target windows
- Feature cutoffs and leakage
- CRISP-DM stages
- Deployment and monitoring
- Prediction versus causal explanation
Python Notebooks and Reproducible Data Work
notebook state, explicit types, data splits, leakage control and repeatable pipelinesMake state visible
Jupyter supports an execute-explore workflow, but hidden execution order can make a notebook appear complete when it cannot run from the top.
- Kernel state and execution order
- Raw-file preservation
- Explicit data types
- Train-only preparation
- Random seeds and repeatability
- Assertions and data invariants
Decision Trees and Information Gain
class counts, entropy, weighted child impurity, information gain and pruningMeasure uncertainty with counts
A decision tree chooses questions that create purer child nodes.
- Parent class counts
- Binary entropy
- Weighted child entropy
- Information gain
- Leaf probabilities
- Pruning and minimum leaves
Linear Regression and Residual Diagnostics
coefficients, fitted values, residuals, squared error, R-squared and model formInterpret the fitted equation in units
Linear regression predicts a numeric target from an intercept and weighted predictors.
- Intercept and slope units
- Fitted values
- Observed-minus-predicted residuals
- RSS and MSE
- R-squared
- Interactions and residual patterns
Logistic Regression and Probability
linear scores, log-odds, sigmoid probabilities, thresholds and log lossTransform a score into probability
Logistic regression represents log-odds with a linear predictor and uses the sigmoid to keep estimated probabilities between zero and one.
- Linear classification scores
- Log-odds and odds
- Sigmoid probabilities
- Decision thresholds
- Odds multipliers
- Binary log loss
Support Vector Machines and Margins
separating hyperplanes, support vectors, hinge loss, margins and regularisationClassify by the side of a hyperplane
A linear support vector machine predicts from the sign of a fitted score and uses a margin to separate classes robustly.
- Separating hyperplanes
- Score sign
- Parallel margins
- Support vectors
- Hinge loss
- C and regularisation strength
Overfitting, Validation and Regularisation
cross-validation, grid search, untouched testing, coefficient penalties and stabilityChoose complexity on unseen folds
Training performance improves with flexibility and therefore cannot select model complexity honestly.
- Training and generalisation error
- K-fold cross-validation
- Mean and standard deviation
- Grid search
- Untouched test sets
- L1 and L2 penalties
Model Evaluation and Project Decisions
confusion-matrix metrics, threshold curves, project evidence and deployment recommendationsPreserve both error directions
A confusion matrix keeps true positives, false positives, false negatives and true negatives separate.
- Confusion-matrix counts
- Precision and recall
- Accuracy under imbalance
- ROC and precision-recall curves
- Capacity-based thresholds
- CRISP-DM project defence
The published objectives move in a deliberate sequence: place big data inside an organisation, identify suitable data, build statistical models, evaluate performance, interpret results and implement analytics on business problems. That means the course is not a catalogue of algorithms.
Every method must be connected to a decision, a data boundary and an error that someone can interpret.
From a business request to a valid target
The opening chapter uses the CRISP-DM cycle to turn a broad request into a population, prediction time, future target, feature cutoff and action. This framing prevents a technically polished model from answering the wrong question.
It also separates prediction from causal explanation and makes deployment part of the analysis rather than an afterthought.
Python as an analysis record
The notebook chapter treats code as a reproducible route from raw data to model input. It focuses on explicit types, saved transformations, restart-and-run checks, train-only preparation and leakage barriers.
These habits matter because a notebook may display plausible output even when hidden state or post-outcome information has contaminated the result.
The quantitative core
Decision trees reduce uncertainty through weighted child impurity and information gain. Linear regression minimises squared residuals and requires interpretation in units.
Logistic regression maps a linear score through log-odds into probability, after which a separate threshold produces an action. Support vector machines use score sign, margins and hinge loss to balance separation against robustness.
The worked examples print every value needed to recompute the result.
Generalisation before performance claims
The validation chapter distinguishes training, validation and untouched testing. It covers cross-validation, grid search, coefficient penalties and the opposite parameter direction used by C in logistic regression and support vector machines.
The evaluation chapter then converts a confusion matrix into precision, recall, workload and missed cases, because a headline accuracy number rarely describes the cost of deployment.
Assessment and project preparation
The official outline publishes five components that total one hundred per cent. It also states a three-hour midterm and a two-hour final examination.
The Moodle midterm page adds open-book conditions and Week 1 to Week 6 coverage, while the captured final section supplies no further format. The group project follows CRISP-DM and ends in a short presentation, so the guide rehearses a narrow evidence-led storyline rather than an inventory of charts.
How to use the guide
Read one chapter, hide the finished answer and work the printed numbers yourself.
Check legal ranges, signs and denominators. Then answer the deployment question: what changes, for whom, at which threshold, with which costly error? The integrated practice chapter joins those moves into an original case. The final review states only the examination facts available in the course material and does not guess unpublished sections or aids.
Worked example · free
Choose a valid split for repeated customer data
- 3Name the repeated-entity and time leakage risks.
- 3Choose a split that protects the future boundary.
- 2Connect the evaluation to an operating action.
Key terms
- CRISP-DM
- An iterative analytics process linking business understanding, data understanding, preparation, modelling, evaluation and deployment.
- Target Leakage
- Information entering predictors even though it would not be available at the moment a real prediction is made.
- Entropy
- A measure of class uncertainty that is zero for a pure node and highest for an even binary split.
- Information Gain
- The parent entropy minus weighted child entropy, used to compare candidate tree splits.
- Residual
- The observed numeric target minus its fitted value in a regression model.
- Mean Squared Error
- The average squared prediction error, which penalises larger misses more heavily.
- Log-odds
- The logarithm of probability odds, represented as a linear predictor in logistic regression.
- Decision Threshold
- The probability cutoff that converts an estimated probability into an operational class label.
- Support Vector
- An observation on or inside an SVM margin that directly influences the fitted separating boundary.
- Hinge Loss
- A margin-based loss equal to the positive part of one minus the actual class sign times the fitted score.
- Cross-validation
- A repeated resampling procedure that rotates held-out folds to estimate generalisation performance and stability.
- Regularisation
- A penalty on model complexity that trades some training fit for a simpler and potentially more stable model.
ECON7880 FAQ
What is the published assessment split?
Class Participation is worth fifteen per cent, Written Assignments thirty per cent, the Group Project and Presentation fifteen per cent, the Midterm Test twenty per cent and the Final Examination twenty per cent. These figures come from the course outline and total one hundred per cent. Check Moodle for live submission settings and any updated instructions.
What is stated about the midterm test?
The captured Moodle page schedules the midterm for 16 October 2026 from 13:30 to 16:30, describes it as open book and limits its scope to Weeks 1 to 6. Students are told to bring a laptop. Search engines are allowed, while communication and generative AI tools are prohibited.
What is known about the final examination?
The official outline gives the Final Examination a twenty per cent weight and a two-hour duration. The captured final section does not provide sections, question types, coverage or permitted materials. Confirm those operational details on the live course Moodle page rather than preparing around an invented format.
How should information gain be calculated?
Calculate parent entropy, then each child entropy, weight each child by its share of the parent and subtract the weighted total from the parent value. Do not average children equally unless their sizes are equal. The largest positive gain gives the largest immediate reduction in class uncertainty.
Why can a high model score still be unreliable?
A score can be inflated by target leakage, repeated entities across partitions, tuning on the test set or a split that ignores deployment time. It can also hide the expensive error behind class imbalance. Inspect the data boundary, validation design, confusion counts and operating threshold before treating one number as evidence of generalisation.
How should I prepare for the group project presentation?
Structure the presentation around the CRISP-DM decisions: business question, row and target, preparation boundary, candidate methods, validation design, error trade-off and deployment implication. The published presentation is brief and includes question time, so rehearse the justification for the split, baseline, threshold and main limitation.
What is the difference between probability and a class label?
A model such as logistic regression estimates a probability from its fitted score. A separate threshold converts that probability into a label. Changing the threshold changes which cases receive action and therefore changes precision, recall and workload, but it does not change the fitted probability for a case.
Which regularisation direction should I remember?
For a penalty multiplier such as ridge alpha, a larger value means stronger shrinkage. For the C convention used with logistic regression and support vector machines in the course notes, a larger C means weaker regularisation and less tolerance of loss. Always name the parameter before stating a direction.
Questions students ask in each ECON7880 chapter
- Why begin with a decision rather than a model?
- The decision fixes the population, target, timing, action and costly error. Without those choices, an impressive score may answer a different question from the one the organisation can act on. Chapter 01 · Analytics Questions and the CRISP-DM Cycle
- Why restart a notebook kernel?
- Restarting removes hidden variables and proves that the visible cell order contains every dependency. A notebook that only works in the author’s current memory state is not reproducible. Chapter 02 · Python Notebooks and Reproducible Data Work
- When is binary entropy highest?
- It is highest at one bit when the two classes are equally represented. It falls toward zero as one class dominates and equals zero at a pure node with only one class remaining. Chapter 03 · Decision Trees and Information Gain
- How should a regression slope be interpreted?
- It is the predicted change in the target for a one-unit predictor increase, holding other included predictors fixed. Give the units, identify the fitted specification and stay within the observed data range. Chapter 04 · Linear Regression and Residual Diagnostics
- Why not use linear regression for class probability?
- An unconstrained line can predict below zero or above one. Logistic regression models log-odds linearly and maps the result into the legal probability range through the sigmoid function. Chapter 05 · Logistic Regression and Probability
- What makes an observation a support vector?
- It lies on or inside a margin and therefore directly influences the fitted boundary. Distant correctly classified points usually do not move the separator because they already satisfy the margin condition comfortably. Chapter 06 · Support Vector Machines and Margins
- What does cross-validation standard deviation show?
- It shows how much performance changes across held-out folds. A model with a slightly higher mean but much larger spread may be less dependable on new data from the deployment setting. Chapter 07 · Overfitting, Validation and Regularisation
- Why can accuracy mislead on rare events?
- Predicting the common negative class can create high accuracy while finding few positive cases. Inspect confusion counts, precision, recall and workload at the chosen threshold. Chapter 08 · Model Evaluation and Project Decisions
How to study for the exam
Start each study session with one method and one blank sheet. Write the business question, data boundary, model objective, fitted output, diagnostic and deployment action. Then recompute a small example without code: entropy and weighted children for a tree, a fitted value and residual for linear regression, a score and probability for logistic regression, or a signed margin and hinge loss for an SVM.
Check the legal range and add units. Next, describe how the result would be validated without leakage and which error matters at the operating threshold. Finish by explaining the result aloud in four sentences: what was predicted, how the model was evaluated, what the main error costs, and what the organisation should do.
This routine prepares you for calculations, interpretation and project questions without assuming an unpublished final-exam structure. In the week before assessment, revisit the two weakest methods and redo their practice cards from memory before consulting the answers.
Your AI Economics tutor for ECON7880
Stuck on a hard ECON7880 question? Sia is AskSia’s AI Economics tutor — ask any ECON7880 Foundations in Big Data Analytics: Concepts and Techniques 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.
