This is the table of contents for Week 8. The interactives below are self-paced — you drag sliders, toggle predictors, and read live output right in your browser. Walk the five core trainers in order; they build up gently from correlation. The Excel-OLS path is an optional companion. Then the deck, the Lab, and the Homework pull it together.
A model is just prediction = f(inputs) — you fit it, then you judge it. This week you meet two
of them: OLS for a continuous answer (how many days a customer takes to pay) and logistic
regression for a yes/no answer (does the invoice go bad). The Python that fits them is a few lines an AI
will happily write for you. That is not where the points are.
The graded skill is specification — picking the right target and predictors, refusing a feature that leaks the future, spotting when two predictors are really the same thing — plus validation and interpreting the output in accounting terms. The trainers exist to build that judgment, on the Summit Gear customer data you will use all week.
Start at the top and work down. The sequence builds from the root idea — correlation — up to the multicollinearity centerpiece. Most trainers are anchored to the same Summit Gear customers table you fit models on this week.
r is — one number from -1 to +1 for how tightly two things move
together — then read r off five real ski and Summit Gear pairs. The Excel tie-in is
=CORREL, with a practice workbook to download.
Why first: correlation → regression → logistic. Regression turns
r into a predictive line; get r right and the next steps are extensions of it.
days_to_pay.
Builds on correlation: in a one-predictor fit the slope is r rescaled
and R-squared equals r squared — the trendline from Week 5, now numeric and multi-variable.
visits: as a single day_number (one meaningless slope, a straight line
that misses the real shape) versus one-hot dummies with a base case (a coefficient per day, predictions
that land). Then trigger the dummy-variable trap on purpose and watch the solve fail.
Why it matters: most accounting predictors are categories — region, account type,
weekday — and every one needs a base case.
p/(1-p) —
then the sigmoid that squashes a linear score into a probability. Read a coefficient as direction of
effect (and exp(coef) as an odds multiplier) on defaulted, then move a cutoff
slider and read one confusion matrix. Evaluation proper is Week 9.
Why it matters: some accounting questions are yes/no, so you need a model
that outputs a probability, not a dollar amount.
days_to_pay and watch the
fitted coefficients, R-squared, adjusted R-squared, and a correlation heatmap move. Add a
collinear predictor and see a coefficient blow up while R-squared sits still; a hot off-diagonal cell in
the heatmap is the warning sign.
Why last: it pulls everything together — correlation, fitting, and reading the
output — into the single most important idea of the week. Slow down here.
=LINEST(...) and the Data Analysis
ToolPak regression on the same Summit Gear data, reading the coefficient and R-squared output, and doing
feature selection plus a multicollinearity check with a CORREL matrix.
Excel bridge: Python is the spine, but a quick regression in a spreadsheet is
a real, fast secondary path — not an afterthought. Pairs with trainer 2 (OLS).
Start at correlation and build up. Feature selection, the last of the five, is the one to slow down on — it pulls fitting, encoding, and reading the output together. The Excel path above is optional.
The trainers are practice. These three are the week itself — the lecture to watch, the Lab to build, and the Homework to submit.
days_to_pay, run feature selection and drop a collinear
predictor, fit logistic for defaulted, and write the one-paragraph spec. Opens as a Canvas
New Quiz with a notebook deliverable.Week 8 of ACCTG 5150, Accounting Analytics. The trainers compute everything in your browser (nothing leaves your machine).