DataQuant Consulting

AI pricing, churn prediction & demand forecasting | Retail, FMCG & Logistics | Results in weeks | Deep dives →

[DataQuant Consulting @DataQuantAI]

What to watch: how the save was made.

36% of merchants saved with a fee discount left within the next 12 months, against 15% of those saved by fixing the problem.

How do you decide who gets the call?

[DataQuant Consulting @DataQuantAI]

A merchant acquirer processes card payments for 42,000 small businesses and keeps about €0.18 of every €100.

It loses 15% of them a year: about €2.1M of net revenue.

Three account managers, about 1,800 calls a quarter. Who should they call? A deep dive, in seven charts.

#Analytics #FinancialServices #Payments #CustomerRetention #ChurnModelling

[DataQuant Consulting @DataQuantAI]

The business case.
Calling every flagged merchant keeps €279K of revenue for €182K of manager time: €97K net.

Calling only where a call helps, 71% of the calls, keeps 96% of the revenue: €140K net.

[DataQuant Consulting @DataQuantAI]

The result, against a random holdout:
called merchants kept 70% of their volume after 12 months;
those held back kept 61%.
E-commerce barely moves (47% vs 44%): they leave for platform payments, not service.

[DataQuant Consulting @DataQuantAI]

The method: take the season out before you score.

A raw month-on-month model flags hospitality in October, when most are having a quiet month.

Adjusted, false alarms fall from 2,570 to 1,210, at the same overall hit rate.

[DataQuant Consulting @DataQuantAI]

Which signals come first? Of merchants whose volume fell 20%+ (seasonally adjusted), 75% left within a year, against 15% overall.

A terminal idle 14+ days: 56%.
Repeated support tickets matter less than assumed: 36%.

[DataQuant Consulting @DataQuantAI]

Smaller merchants leave.

E-commerce loses 21% of its merchants a year; hospitality 17%.

But leavers are 21% smaller than stayers, so counting merchants overstates the loss. In volume it is 12% a year.

[DataQuant Consulting @DataQuantAI]

Read the burn-down, cut by this week’s elasticity, stage the depth, and never price below the floor without a human.

The models decide depth and timing; the agent re-plans weekly; the merchandiser owns the exceptions.

[DataQuant Consulting @DataQuantAI]

‘Validated models decide.’ Here’s what that means for clearance pricing.

The Markdown Agent, torn down: every signal it reads, every model it calls, one cut walked end to end, and the 40% markdown the guardrail stopped.

Teardown ↓

#Retail #Markdown #RevenueManagement #Pricing #AgenticAI

[DataQuant Consulting @DataQuantAI]

The one it caught: a 40% cut in week 3, below landed cost with eleven weeks to run.

The guardrail held it at the 20% cap and escalated. The merchandiser found the line was mis-ranged in two stores — and moved the stock.

[DataQuant Consulting @DataQuantAI]

One cut: 2,400 shirts left in week 6,
900 still on hand at season end at full price.

A 15% cut now does twice the work of one in week 12, so the plan is 15% now and 25% in week 10 — 60 units left, not a 50% slash.

[DataQuant Consulting @DataQuantAI]

The models: an elasticity model that changes through the season, a burn-down forecast of what is left at the end, a schedule optimizer for depth and timing, and a margin guardrail setting the deepest cut allowed.

[DataQuant Consulting @DataQuantAI]

What it reads: sell-through, stock on hand, price history and the season calendar.

The decision is never ‘how much discount’ in isolation — it is how much stock, how many weeks, and how fast it is actually moving.

[DataQuant Consulting @DataQuantAI]

Model the drivers, borrow a shape for new items, make the levels add up, and keep one-offs out of the baseline.

The models decide the number; the agent re-forecasts as events land; the planner owns the exceptions.

[DataQuant Consulting @DataQuantAI]

‘Validated models decide the number.’ Here’s what that means for a demand forecast.

The Forecasting Agent, torn down: every input, every model, one promotion re-forecast end to end, and the bulk order it refused to learn from.

Teardown ↓

#DemandForecasting #DemandPlanning #SupplyChain #AgenticAI #B2B

[DataQuant Consulting @DataQuantAI]

The one it caught: a one-off order of 3,000 units — over seven weeks’ demand in a day.

The model wanted to raise the baseline. The guardrail kept it out; learn from it and the Planning Agent orders against a ghost.

[DataQuant Consulting @DataQuantAI]

One re-forecast: a 400-a-week item goes on promotion; the driver model lifts it 35% to 540, and finds 30% of the lift is switching from sister products.

Over the three weeks the category rises 294, not 420. A planner approves.

[DataQuant Consulting @DataQuantAI]

The models: a driver model for what price, promotions and events do;
an analog model for items with no history;
a reconciler so SKUs add up to the category plan;
and a guardrail on what gets into the baseline.

[DataQuant Consulting @DataQuantAI]

What it reads: sales, the price and promotion calendar, events, and the product hierarchy.

A forecast that only sees its own history misses every spike someone in the business planned on purpose.

[DataQuant Consulting @DataQuantAI]

Before reading any result, write down the smallest effect that would change the decision, then ask whether the interval rules it out.

And if a segment ‘works’ after the fact, it is a hypothesis for the next test, not a finding.

[DataQuant Consulting @DataQuantAI]

‘No significant difference’ is the most misread sentence in experimentation.

These three tests all earned it. One rules out anything worth having. One can’t tell. One is probably a win.

Reading a null, taken apart ↓

#Experimentation #ABTesting #Statistics #DecisionScience #CommercialExcellence