Segment customers, optimize reach, run A/B tests, and read what moves the numbers.
Business analytics turns customer and campaign data into decisions: who to target, which product mix to carry, whether a change actually helped. Most of these questions are classical statistics wearing a business hat — an A/B test is a two-proportion test, key-driver analysis is regression, a perceptual map is correspondence analysis.
A couple of practices keep the conclusions trustworthy. Size an A/B test up front and resist “peeking” — repeatedly checking significance as data streams in inflates the false-positive rate, so use a fixed sample size or a sequential-testing correction. And when you slice a result across many segments or variants, correct for multiple comparisons before crowning a winner. Analytics projects are often organized around the CRISP-DM cycle (business understanding → data → modeling → evaluation → deployment).
Segmenting customers (Wedel & Kamakura)
Group customers by behavior so you can treat them differently. RFM scores each on how recently, how often, and how much they buy; clustering finds natural segments from any set of features.
▸ — Model which of several segments or brands a customer picks.
▸ — Recency / Frequency / Monetary segmentation.
▸ — Behavioral clustering into k segments.
▸ — Dendrogram of segments — choose the cut.
▸ — Auto-selects the number of segments (BIC).
Product mix & reach (TURF)
TURF (Total Unduplicated Reach and Frequency) finds the smallest set of products, flavors, or messages that reaches the most distinct people — the classic assortment / line-up optimization question.
▸ — Maximize unduplicated reach from a candidate set.
A/B tests & conversion (Kohavi et al. 2020)
Did version B convert better than A? That is a two-proportion test. For richer breakdowns use a contingency table; the 2×2 calculator adds the lift, relative risk, and risk difference behind the headline.
▸ — Compare two conversion rates (A/B test).
▸ — Lift / relative risk / risk difference from a 2×2.
▸ — Association across several segments or variants.
▸ — Declare two variants equivalent, rather than merely "not significant".
▸ — Benford's law screen for fabricated numbers in submitted data.
What drives the outcome? (key-driver regression)
Key-driver analysis is regression: relate satisfaction, spend, or churn to its predictors and rank their influence. Logistic regression handles yes/no outcomes like churn; regularization tames many correlated survey items.
▸ — Many correlated drivers — lasso/ridge to stabilize.
Perceptual maps & item structure (Greenacre)
Correspondence analysis turns a brand × attribute table into a 2D perceptual map; factor analysis and PCA collapse many survey items into a few underlying dimensions you can name and track.
▸ — Brand–attribute perceptual map (biplot).
▸ — Reduce survey items to latent dimensions.
▸ — Index construction / dimension reduction.
▸ — Show what built the change between two totals, step by step.
▸ — Are two correlations significantly different?
Forecasting demand (Box–Jenkins · Hyndman)
Project sales, traffic, or demand forward from history, capturing trend and seasonality.
▸ — Separate the trend from the seasonal swing before you forecast.