Find the 1% of customers who will buy a new product

For every new product we rank all 1 million customers. The model reads each customer's receipts from before launch day, so what they buy afterwards is only used to check.

The model only sees the past, and the window moves with each launch

Every customer is described by the 8 weeks of receipts before that product's launch day; what they buy afterwards is only used to check.

Tested in 3 categories

Test launches were drawn at random within each category, using only rules known before the outcome: first time on sale in Thailand, 100+ stores within 2 weeks, one product per brand, class and week. Learn, tune and test launches never overlap in time. Universe: 1,000,000 loyalty card customers.

Real example. Khao Shong Instant Coffee Espresso 45 g · coffee · launched 24 Aug 2026 in 223 stores · 3 of the 1,000,000 customers we ranked.

A ranks high because coffee is a habit (5 trips, the last one yesterday) and they already buy Khao Shong. C bought no coffee in 8 weeks.

Shoppers A–C are real loyalty card customers; values rounded, no IDs. Rank = position among the 1M customers sorted by our forecast. Tuned ML = gradient boosting on the same inputs. Random = 1% of 258.

Buyers caught in the top 10,000 · share of all real buyers · axis 0 – 25%

Ranking score (AUC) · 0.5 = coin flip, 1.0 = perfect

• Vs tuned ML: ahead in all 3 categories, but within the noise (ranking +0.007).
• Vs category rule: +0.08 ranking, 142 more buyers.
• No training: ours reads 16,384 past-launch rows as examples; tuned ML was fitted and tuned on 457k rows.

3 of the 4 tested categories. DG Non-Food is excluded; there, tuned ML caught more buyers (352 vs 226), and across all 4 categories the two models tie (ranking 0.808 vs 0.808; 1,297 vs 1,408 buyers caught). Every model is scored on the same customers.

On buyers caught, the category rule wins the snack and candy launches (Geemong, Line Friends).

All 10 launches: caught ours 273 · tuned ML 262 · rule 274 · ranking ours .823 · tuned ML .815 · rule .734

Real buyers = customers among the 1M who bought that exact product within 28 days of launch. Highlighted = best in the row. Rule = rank by past spend in the product’s category. Ranking score: 0.5 = coin flip, 1.0 = perfect.

Tuned ML wins the drink and ice-cream launches (Proten, Magnum); one launch had no buyers.

All 10 launches: caught ours 745 · tuned ML 744 · rule 619 · ranking ours .854 · tuned ML .850 · rule .789

Real buyers = customers among the 1M who bought that exact product within 28 days of launch. Highlighted = best in the row. Rule = rank by past spend in the product’s category. Ranking score: 0.5 = coin flip, 1.0 = perfect.

Most Hardline launches had under 20 buyers in 1M, too few to separate the methods.

All 10 launches: caught ours 53 · tuned ML 50 · rule 36 · ranking ours .765 · tuned ML .756 · rule .694

Real buyers = customers among the 1M who bought that exact product within 28 days of launch. Highlighted = best in the row. Rule = rank by past spend in the product’s category. Ranking score: 0.5 = coin flip, 1.0 = perfect.

The model ranks all 1,000,000 customers for each new product. You choose how far down the ranking to go. A short list is the most precise; a longer list reaches more of the buyers at a lower hit rate. Pick the size that fits the sampling, SMS or coupon budget.

Suggested use: a sampling or SMS budget of 1,000 per launch finds buyers at 45× random; 10,000 catches 1 in 5 buyers; 100,000 reaches over half of them at 5.5×.

Buyers caught = customers in the list who bought that exact product within 28 days of launch, summed over 30 August 2026 launches. × random = hit rate vs picking the same number of customers at random. 3 of 4 categories (DG Non-Food excluded).

Two things put a customer on the list: they already buy the category (4.9 vs 0.7 trips in it) and they keep exploring. That makes the list a good starting audience for the next launch too.

Picks = the top 10,000 per launch across the 30 launches, weighted to the 1M universe; average/all customers = the whole 1M. Another August launch = any of the other 39 tested August launches, bought within 28 days.

Not just busier or bigger spenders · at the same visits or the same spend, our picks still try more products

Different products bought in the 8 weeks before launch.

Under 1k THB left out: almost none of our picks spend that little.