[03] / 06

Highlighted projects

From the commercial question to the decision it changed.

[01] The promotions were paying for themselves and nothing more A coupon programme that looked like a success — and was rescheduling sales, not creating them. 0incremental return
[02] Knowing when a bar will reorder Daily demand across 20,000+ outlets, most of them buying rarely and irregularly. ~80%no-purchase calls correct
[03] Why insurance customers were unhappy Everyone blamed thin coverage. The data pointed at the process instead. 2 in 5unhappy with process, not cover

Project [01] / 03

The promotions were paying for themselves and nothing more

Spain’s leading brewer — one of the country’s top FMCG companies — ran discount coupons on its ordering platform for bars and restaurants. Sales and discounts moved together almost perfectly, so the programme looked like a success.

88%

correlation between sales and discounts

incremental return

The trap

Correlation isn’t causation.

Customers who would have ordered anyway also use the coupon — and a discount can simply pull forward an order that was coming next week.

What I did

Bayesian Ridge over hundreds of thousands of orders.

Separating the first coupon order from the ones that followed, with different methods for percentage and fixed-amount coupons.

What came out

Net return: indistinguishable from zero.

The first coupon order brought extra volume; the following orders gave it all back through cannibalisation. The programme was rescheduling sales, not creating them.

Where it landed

An alternative plan, taken to the steering committee.

Five lifecycle segments with data-derived thresholds, a look-alike model to find offline customers worth acquiring, and an activation plan with its own success measure.

Key insight

Correlation was 88%. Incremental return was zero.

Project [02] / 03

Knowing when a bar will reorder

The same platform needed to tell each outlet what to order and when. Standard forecasting struggles here: a small bar might order one product every three weeks, with long gaps that are not zeroes so much as silence.

20k+

outlets forecast daily

~80%

no-purchase calls correct

~1%

false negatives

The problem

Irregular buyers across 20,000+ outlets.

Daily demand for every distributor, outlet and product line — most of them buying irregularly.

What I did

Each outlet’s rhythm, plus the world outside.

Recency, moving averages and the most comparable week a year earlier, combined with weather, public holidays and football fixtures weighted by how big the match was.

The hard part

Knowing when a gap really means no purchase.

Interpolation and a statistical rule, tuned against the error that costs money: telling an outlet it doesn’t need stock when it does.

What came out

4 in 5 no-purchase calls correct.

False negatives near one percent, validated by replaying four months of real order history.

Key insight

The interesting variable turned out to be the football calendar.

Project [03] / 03

Why insurance customers were unhappy

An insurer wanted to reduce detractors. The assumption inside the business was that people were dissatisfied because the medical coverage was too thin.

2 in 5

detractors unhappy with the process, not the cover

What I did

Two models and a friction score.

A predictive detractor model on authorisation data and a satisfaction model beside it, scoring friction by province and specialty across satisfaction, complaints, usage and operations.

What came out

It wasn’t the coverage.

Among detractors who barely used their policy, roughly two in five were unhappy with the process — delays and authorisations — not with what was covered.

Where it landed

From widening coverage to fixing operations.

Prioritised by the friction score, and delivered as committee summaries, dashboards and plain-language explanations inside the file front-line agents actually use.