Economics graduate to data scientist: econometrics that transfers

Run this first

A hypothetical store runs promotions that cut the price and also bring extra customers through advertising. Regress units sold on price alone and you get the wrong price effect, because price is correlated with the promotion. This is omitted-variable bias, and you studied it in your econometrics course.

import random
import pandas as pd

random.seed(7)

# Synthetic weekly store data (hypothetical): promotions cut the price AND add footfall
rows = []
for week in range(200):
    promo = 1 if random.random() < 0.4 else 0
    price = 100 - 12 * promo + random.gauss(0, 4)
    sales = 900 - 5 * price + 80 * promo + random.gauss(0, 15)
    rows.append({"week": week, "price": price, "promo": promo, "units": sales})
df = pd.DataFrame(rows)

# Model 1: units on price only
b_naive = df["units"].cov(df["price"]) / df["price"].var()

# Model 2: units on price and promo, via the two-regressor OLS formula
c = df[["units", "price", "promo"]].cov()
det = c.loc["price", "price"] * c.loc["promo", "promo"] - c.loc["price", "promo"] ** 2
b_price = (c.loc["units", "price"] * c.loc["promo", "promo"] - c.loc["units", "promo"] * c.loc["price", "promo"]) / det
b_promo = (c.loc["units", "promo"] * c.loc["price", "price"] - c.loc["units", "price"] * c.loc["price", "promo"]) / det

print(f"price effect, price only:        {b_naive:.2f} units per rupee")
print(f"price effect, controlling promo: {b_price:.2f} units per rupee")
print(f"promo effect, controlling price: {b_promo:.1f} units per week")

mean_price = df["price"].mean()
for label, b in [("naive", b_naive), ("controlled", b_price)]:
    elasticity = b * mean_price / df["units"].mean()
    print(f"{label} price elasticity at the means: {elasticity:.2f}")

assert b_naive < b_price < 0
assert abs(b_price - (-5)) < 1.5

On Python 3.12 the output is:

price effect, price only:        -9.84 units per rupee
price effect, controlling promo: -5.29 units per rupee
promo effect, controlling price: 79.8 units per week
naive price elasticity at the means: -2.07
controlled price elasticity at the means: -1.11

The true price effect built into the synthetic data is -5 units per rupee. The price-only model nearly doubles it, because it credits the price cut with customers who actually came for the promotion. A pricing team that trusted the naive elasticity of about -2 would conclude demand is highly price-sensitive and cut prices further, when most of the lift came from the promotion.

Why this matters in a data science job

Many business questions a data scientist is asked are really causal questions dressed as prediction: "what happens to sales if we raise price", "did the new onboarding flow reduce churn", "which channel should get more budget". A candidate who fits a gradient-boosted model and reads feature importance as causal effect makes precisely the mistake above. You were trained to ask what else moved at the same time. Say so in interviews, with this example.

The artefacts from your degree that carry over:

• Regression output tables you interpreted for coursework or a dissertation: coefficients, standard errors, R-squared, and what heteroskedasticity does to them.
• Stata, EViews or R do-files that cleaned survey data such as NSS or PLFS extracts, or panel data from a term paper.
• Your dissertation itself, especially the section defending your identification strategy.
• Index number and elasticity work from microeconomics and applied courses.

What economics training usually does not cover

Be precise about the gap, because it is different from the gap a computer science graduate has.

The mindset shift is from "is this coefficient significant?" to "does this model make better decisions on data it has not seen?". Both matter; data science roles usually weigh the second more. The Data Science programme at /courses/ds covers Python, machine learning, validation and deployment over six months, and suits someone who already has the statistics foundation.

A first portfolio project that combines both halves

Extend the synthetic example into a real portfolio piece:

1. 1Add a third confounder, such as a festival week, and show how the price estimate moves as you add or omit controls.
2. 2Build a prediction model for weekly units with a proper time-based train/test split, and compare its error with a seasonal naive baseline.
3. 3Write a one-page note separating the two uses: the prediction model forecasts demand for stock planning; only the controlled regression, with its stated assumptions, informs pricing.

That note is the part most portfolios lack. It shows you know which tool answers which question.

Pricing and experimentation teams as a first home

With strong Python, a data scientist role on a pricing, marketing analytics, experimentation or economics-heavy team (fintech, e-commerce, policy research) uses your training most directly. If your coding is still mostly Stata, a data analyst or decision scientist role is a sensible bridge; see data analytics vs data science for how those jobs differ.

Explaining your background

"I studied economics, and my dissertation estimated how state electricity tariffs affected small-firm output, so most of my training was about separating cause from correlation. What I lacked was prediction and engineering, so I moved my work from Stata to Python, learned scikit-learn and cross-validation, and built a project that uses the same data twice: once to forecast demand, once to estimate price response, with a note on why those are different questions. I think that combination is useful on a team where people ask 'what will happen' and 'what should we change' in the same meeting."

To check which of these skills your resume currently shows, use the free skill check.