Course Introduction and Semester Project

<p class="kicker">MST0052 · Lecture 1 · Fall 2026</p>

## Course Introduction<br>& the Semester Project

### Vegard H. Larsen — BI Norwegian Business School

---

**Build reliable predictive models** — and know when to trust them.

This semester: an end-to-end ML project on a dataset you choose.

---

## About me

- **Vegard H. Larsen** — Associate Professor, Data Science and Analytics, BI
- Economics × data science — ML, NLP, sentiment and uncertainty
- Previously Senior Researcher at **Norges Bank**

---

## About you

- Your **name**
- Your **background** — what did you study before?
- One thing you hope to **get out of** this course

---

## Quick poll

- **Programmed in Python before?** — yes / a little / no
- **Touched ML before?** — built models / read about it / not really
- **Have a dataset in mind?** — yes / vaguely / no

---

## What this course is about

---

## What is predictive modelling?

Using historical data to predict a target on new data.

![Prediction flow: regression and classification](../figures/prediction-flow.svg)

---

Most ML projects fail on **practice**, not algorithms.

This course teaches the practice — not just the menu of algorithms.

---

## Predictive vs explanatory

<div class="cols">
<div class="callout">
<h3>Prediction</h3>
<p>Does the model generalise to new data?</p>
</div>
<div class="callout">
<h3>Explanation</h3>
<p>What causes what?</p>
</div>
</div>

This course is about prediction. Causal inference is a different course.

---

## Where classical ML still wins

![Classical ML vs deep learning quadrant](../figures/ml-quadrant.svg)

---

## By December, you can

- **Frame** a real problem as a predictive modelling task
- **Build** a reproducible pipeline you can defend choice by choice
- **Compare** model families and explain the tradeoffs honestly

---

## How the course works

---

## The 16-lecture arc

![Semester timeline with four phases](../figures/semester-timeline.svg)

---

## How the lectures work

- **Lectures** = ideas and intuition
- **The project** = where the doing happens
- No live coding — code on slides, you run it yourself

---

## Tools you will use

**Core**   `Python 3.11+` `NumPy` `pandas` `matplotlib` `scikit-learn`

**Editor**   `Jupyter` `VS Code` — your choice

**Environment**   `pip + venv` `conda`

</div>

Boosting libraries (XGBoost, LightGBM) join later in the semester.

---

Everything lives on the **course site**.

tem0052.vegardlarsen.com

---

## Communication and getting help

- **Shared channel** — questions of general interest
- **Email** — personal matters
- **Office hours** — TBA

---

## How you're graded

---

## How you are graded

<div class="cols">
<div class="stat stat--blue">
<span class="stat-value">30%</span>
<span class="stat-label">Semester project</span>
</div>
<div class="stat">
<span class="stat-value">70%</span>
<span class="stat-label">Oral exam</span>
</div>
</div>

<p class="muted">You must pass <strong>both</strong>. Full detail: assessment guidelines on the course site.</p>

---

## What the project is graded on

- **Deliverable:** PDF report + reproducible code
- Quality of analysis, clarity of reasoning, honesty about limitations
- **Quality beats length** — no page count to hit

---

## What the oral exam tests

<div class="cols">
<div class="callout">
<h3>Part A — your project</h3>
<p>Explain every choice, method, and number.</p>
</div>
<div class="callout">
<h3>Part B — the syllabus</h3>
<p>Conceptual questions across the course.</p>
</div>
</div>

**Closed book.** No notes, no laptop, no AI.

---

## Grades, anchored

<div class="cols-3">
<div class="stat stat--green">
<span class="stat-value">A</span>
<span class="stat-label">Ambitious, rigorous,<br>well defended</span>
</div>
<div class="stat stat--yellow">
<span class="stat-value">C</span>
<span class="stat-label">Standard methods,<br>correctly applied</span>
</div>
<div class="stat stat--red">
<span class="stat-value">F</span>
<span class="stat-label">Cannot defend<br>the work</span>
</div>
</div>

Full A–F descriptors in the backup slides and the assessment guidelines.

---

## The semester project

---

## The project, in one slide

- **Individual.** Real dataset. Real predictive problem. **End-to-end.**
- Due **20 November 2026 at 12:00** — PDF report + code
- Runs alongside the course; you apply each method as it is taught

---

## What counts as a valid project

- A **target** you are predicting — regression or classification
- Models trained and **evaluated on held-out data**
- A **reproducible pipeline** from raw data to predictions
- Dashboards or clustering-only studies

</div>

---

## Methods in scope

**Linear models**   `OLS` `ridge` `lasso`

**Classification**   `logistic` `k-NN` `naive Bayes`

**Trees & ensembles**   `decision trees` `random forests`

**Boosting**   `XGBoost` `LightGBM` `CatBoost`

**Kernel methods**   `SVM`

**Unsupervised, as support**   `PCA` `k-means`

</div>

Going beyond the syllabus within this family is encouraged.

---

## Out of scope — no deep learning

- No multi-layer neural networks, CNNs, RNNs, Transformers
- No fine-tuning large pretrained models
- No LLM-as-predictor

Borderline case? **Ask first.**

</div>

---

## Example project shapes

- **Churn prediction** — telecom customers
- **Sales forecasting** — retail, per store
- **News classification** — by topic
- **Oslo rental prices** — scraped listings
- **Dropout warning** — students at risk

Shapes, not topics to copy. Pick something you would want to defend.

---

## Project milestones

![Project milestones from L1 to the December oral](../figures/project-milestones.svg)

No formal scope check — want feedback? Come find me.

---

## Lecture 8 — optional project showcase

<p class="kicker">Volunteer · 5-minute slots</p>

- Present your project-in-progress to the class
- **Feedback from me** + **ideas from what others are trying**
- Not presenting? **Come anyway** — the room is half the value

**Signup deadline: TBA.** Slots fill in order.

</div>

---

## AI tools & integrity

---

**No restrictions** on what you use AI for.

<p class="muted">One requirement: you can <strong>defend every line</strong> at the oral.</p>

---

## What good AI use looks like

- Use AI to **learn**, **debug**, **draft**, **explain papers**
- Never paste output you can't explain
- Include an **AI usage statement** in your report

---

## Collaboration and integrity

- **Discussion encouraged** — talk, compare, help each other debug
- **Final submission is your own** — no shared code, text, or reports
- BI's academic misconduct rules apply

---

## Wrap-up

---

## Before Lecture 2

- **Install** Python and a working virtual environment
- **Skim** the project and assessment guidelines on the course site
- **Start thinking** about a dataset you'd like to work on

</div>

---

## Dataset starter pointers

**General**   `UCI ML Repository` `Kaggle` `OpenML` `Hugging Face`

**Open government data**   `SSB` `Eurostat` `Norges Bank`

</div>

Aim for hundreds to hundreds-of-thousands of rows. Check the licence, cite the source.

---

## Common pitfalls (preview)

- **Data leakage** — the future sneaking into training
- **No target** — a purely descriptive problem
- **Too small or too messy** — not enough signal
- **Needs deep learning** — images, raw text, sequences

---

## Questions

Backup slides below — press down ↓

## Predictive vs explanatory (expanded)

Same data, two goals:

- **Predictive:** rent = f(size, location, floor) — does it predict new listings?
- **Explanatory:** does an extra floor raise rent, all else equal? Causal.

Prediction tolerates correlated features. Explanation must isolate effects.

## Reproducibility

- **Pinned dependencies** — `requirements.txt` or `environment.yml`
- **Random seeds** — set them, document them
- **A README** — how to run the pipeline from scratch
- **A pipeline** — not a notebook of scattered cells

</div>

## Grade descriptors in full

| Grade | What it looks like |
|-------|--------------------|
| **A** | Ambitious, rigorous, well defended. Independent thinking. |
| **B** | Solid work with at least one aspect beyond the basics. |
| **C** | Standard methods, correctly applied. Competent but not deep. |
| **D** | Works end-to-end but noticeable gaps in reasoning or validation. |
| **E** | Minimal viable submission. Significant methodology flaws. |
| **F** | Cannot defend the work or missing major components. |

## Office hours and feedback

- Day, time, and location: TBA
- Drop in — no appointment needed
- Email for things that are personal or can't wait

## Mini scikit-learn example

A five-line pipeline as a teaser:

```python
from sklearn.pipeline import make_pipeline
from sklearn.preprocessing import StandardScaler
from sklearn.linear_model import Ridge

pipe = make_pipeline(StandardScaler(), Ridge())
pipe.fit(X_train, y_train)
```

We'll unpack every piece of this in Lecture 3.

---

## What's next

**Lecture 2:** AI tools in a machine-learning workflow

- Where AI helps — and where it fails silently
- Prompts that work vs prompts that don't
- Documenting AI use in your project