The RAMSTAT analytics system team developed an architecture to calculate marketplace revenue. Initially, the project relied on daily snapshots of product cards, including prices, ratings, and review counts, as historical data was unavailable. To assess market competition, the team utilized the Herfindahl-Hirschman Index and cumulative market share metrics.
The integration of historical data from Wildberries, specifically daily stock snapshots, allowed for revenue estimation based on inventory changes. However, this method produced inaccurate results due to technical data gaps and the 'Dachnik Syndrome,' where sellers manually adjust stock levels to avoid logistics penalties, creating phantom sales spikes.
To address these anomalies, the team moved away from standard statistical filters, which failed due to non-normal data distribution. They adopted a three-level verification algorithm using logarithmic scale metrics and quantile ranges. This approach identifies artifacts and zeroes out unreliable revenue data, which previously accounted for up to 95.2% of calculated figures in some samples.
The team implemented Bayesian smoothing with pseudo-observations to mitigate the impact of rare keywords on performance rankings. By assigning a confidence weight to revenue based on review counts, they prevented unvalidated items from topping search results. This method ensures that rankings are driven by statistically significant data rather than outliers.
Further investigation revealed that the 'phantom sales' were caused by Wildberries' interface capping stock visibility in the shopping cart. The platform displays an artificially limited number of items regardless of actual warehouse stock, which fluctuates based on geolocation, user sessions, and warehouse types, leading to false delta calculations.
