Adding distinct customer counts changed product pair forecast loss from 0.276025 to 0.276024 at one UK retailer in 2011

We tested whether distinct past customer counts add information to forecasts of later product pair purchases. We used positive UK invoices from UCI Online Retail II. A pair is two products in one invoice. We fitted logistic models on January to June 2010 counts and July to November 2010 outcomes. We tested the same months in 2011. The main outcome required a later customer who did not buy that pair in the first six months. For 42,972 test pairs, adding distinct customer counts changed mean log loss from 0.276025 to 0.276024. The change was -0.00000144. A descriptive 95% range from product resampling was -0.00000304 to +0.00000019, with fixed fits. Other product set sizes and penalties gave small changes in both directions. The main counts had a correlation of 0.982 after log transformation. Later purchases occurred for 91.9% of tested pairs. This test did not establish a stable gain or equal performance. It applies to selected popular products at one retailer with many wholesale buyers. The package has invoice baskets, source hashes, code, all forecasts, model checks, figures, and repeat instructions.

retail · product pairs · customer counts · link prediction · temporal test · null result · open data