Applied Predictive Technologies’ software helps businesses run internal experiments on offline data like in-store purchases. In this case competition, a client was interested in determining the effect of several varieties of poster campaign on profits. We used hypothesis testing on several months of sales data from a test and control group to determine whether various iterations of a sales campaign had a positive effect on profits. Our presentation won third place.
The full slides are available here.
Identify possible confounding variables available in data set
Verify that treatment group and control group have similar distributions of possible confounding variables
Plot and compare profits by product across treatment and control groups
Choose best statistic for hypothesis test: percent mean difference for each store over weeks before and weeks after campaign.
Run t-tests on each campaign
Check correlations between test statistic and possible confounding variables
Analyze relationship between sales of different products to determine overall effect on products
Provide recommendation for continuing poster campaign
For modeling and plotting, we used R and Python, including Pandas for some data manipulation and standardization.