Voucher Analyzer

A voucher analyzer for a small Southeast Asian online seller, the kind that runs monthly store vouchers on top of the 11.11 and 12.12 mega-campaigns. It answers the question raw campaign sales cannot: of everything that sold during a promotion, how much would have sold anyway?

Every campaign spikes sales, so every voucher looks like a winner. But part of each discount lands on buyers who were going to purchase regardless. That part is subsidy, not lift, and a sales report cannot see it. Separating the two is the entire job.

The engine estimates incrementality blind to the answer. It learns each category's ordinary order rate from non-campaign days, then treats the excess during campaigns as voucher-driven, a baseline-counterfactual in the difference-in-differences spirit. On a simulated book of 2,678 redemptions across 7 voucher types, the estimate recovers a 45% incremental share against a true 46%. The naive read, counting every redemption as new, would have claimed 100%.

From there each voucher is costed out and ranked:

• F1, a small RM5 fixed voucher, adds the most real margin.
• P3, the 20% voucher, is the only one that loses money: the discount runs deeper than the margin it returns.
• Grocery and Toys lose on every voucher, while the fatter categories carry the whole program.

The recovery check is the part that matters. An estimate of something unobservable is only worth trusting if it can be shown to land on a known answer, so the true was_incremental flag is withheld from everything the tool can query and used only to grade the estimate. The gap between the 46% truth and the 100% naive read is exactly why one flat voucher rule quietly leaks margin.

The output is a per-category plan: which mechanic, how deep, and what to stop. The questions run as fixed, read-only SQL in the browser, so the demo is free to host and has no key to leak.