Machine Learning — You Already Know More Than You Think 🤖

Scene: week one of your first analyst job. The chief economist stops at your desk: "Unemployment just jumped this quarter. Give me a GDP growth forecast by lunch." You panic-Google "machine learning", and drown in neural networks and Greek letters. Breathe. The tool you need is one you already met in econometrics class — you just know it under a different name.

The secret nobody tells economics students: the workhorse of machine learning, linear regression, is literally OLS — the same estimator from your econometrics course, same normal equations, same β̂. What changes is not the tool. It is the question you ask of it.

Econometrics vs machine learning — same engine, different exam

An econometrician fits on 60 quarters and writes a paper about β̂ and its p-value. An ML practitioner fits on 50 quarters and is graded only on how well the model forecasts the 10 quarters it never saw. Same regression — different scoreboard.

🧠 The mental model: study the past, sit the exam

• ›fit() = studying. The model reads the historical data and distils it into what it has learned — for linear regression, just a slope and an intercept. Econometricians call this estimation.
• ›predict() = exam day. The model faces a case it has never seen and must answer using only what it learned. Econometricians call this forecasting.

A student who memorises past papers word-for-word aces the mock exam but flunks the real one. In ML that failure is called overfitting — public enemy #1, and it gets its own lesson soon. For now, remember: the grade that counts is the exam on unseen questions.

The Rosetta Stone: ML vocabulary for economists

Learn this mapping once, and every ML tutorial suddenly reads like econometrics:

Meet scikit-learn — three lines, every model

scikit-learn is Python's standard ML library, and its genius is a uniform ritual that never changes, whatever the model:

 Python
model = LinearRegression()   # 1. choose the recipe (nothing learned yet!)
model.fit(X, y)              # 2. study the past   (econometrics: "estimate")
model.predict(new_X)         # 3. sit the exam     (econometrics: "forecast")

One quirk to know: scikit-learn wants X as a 2D structure — rows are samples (quarters), columns are features — even when you have just one feature. That is why the code below builds X as a list of one-element lists.

Your dataset: Okun's law

In 1962, economist Arthur Okun documented one of macro's most reliable regularities: when unemployment rises, GDP growth falls below trend. His original estimate was about 3 points of output per point of unemployment; modern re-estimates (e.g. Ball, Leigh & Loungani) put the coefficient nearer 2 — the value your data follows. Your training set is 12 quarters of stylised data with exactly this pattern, and you will make the machine rediscover Okun's law from the data — which is precisely what "learning" means here.

The learned parameters then have a clean econometric reading:

• ›Intercept ≈ trend growth when unemployment is flat (about 3%)
• ›Slope ≈ the Okun coefficient (about −2)

Your Task

The starter code loads the 12 quarters and creates a LinearRegression model — but never trains it. Run it as-is: it politely refuses to forecast. Your job is the single most important line in machine learning: the fit call.

Predict before you run: with an intercept near 3 and a slope near −2, what growth should the model forecast for a new quarter where unemployment rises by 0.6pp? Do the mental arithmetic (3 − 2 × 0.6), then run and see how close you were — congratulations, you just sanity-checked an ML model like a professional.