Optimizing Advertising Revenue Forecasting for a Global E-Commerce Leader

The company runs a large online auction platform and shopping website that connects buyers and sellers in over 190 countries. Their website has millions of sellers with 1.9 billion global listings and over 132 million active buyers.

THE CHALLENGE

The company earns revenue from fees on paid sales, as well as from promoted listings, which is a suite of first-party advertising offerings. The different types of promoted listings include a cost per acquisition product, a cost per acquisition product for auction listings, a cost-per-click product and an off-platform advertising product. One of the key performance metrics is the Gross Merchandise Volume (GMV) which is defined as the total value of all paid transactions between users on the platform inclusive of shipping fees and taxes. Business users rely on historical GMV data to derive forecasts of advertising revenue for the next year. They use a complex Microsoft Excel based solution to forecast quarterly advertising revenue on an ongoing basis. However, while comparing the actual revenue figures with the predicted revenue values, the company found that there were large variations, especially during the holiday season and other months with peak demand. As a result, the existing model did not allow the advertising team to accurately determine the expected revenue flow for several months of the year. The company needed an analytics partner to develop a new advertising revenue forecasting solution that would effectively factor in the expected increases during certain months of the year while predicting quarterly, weekly and daily revenues.

THE SOLUTION

The Prescience team explored different machine learning (ML) models to identify which of them would be a best fit for the historical GMV data. Among the ML models that were evaluated were the Seasonal Autoregressive Integrated Moving Average (SARIMA) model, which is designed to manage seasonal data patterns. However, given the nature of their advertising business, none of these shortlisted ML models were suitable for revenue forecasting. Next, the team adopted an Indexing based approach with revenue as the key metric for forecasting, instead of GMV. The solution created a weighted average of historical revenue, based on which the corresponding index was calculated for that time period. The weighted average of the preceding years was determined with a higher weight assigned for the most recent year. To achieve accurate forecasting at a daily level, the solution factored in higher sales for Thursdays and Fridays, with lower sales for weekends. It also had built in multipliers for annual seasonality and specific holidays which fell on the same day and same date every year. With this new Indexed solution, business users have access to forecasted advertising revenue numbers at a quarterly, monthly and daily level for the next 12 months. The technology used for this engagement was,

1. Microsoft SQL
2. Microsoft Excel

THE IMPACT

Business users from the advertising division are now able to analyse the forecasted advertising revenue figures at different granularity levels. While the old solution had significant variances between forecasted and actual revenue figures in several months, the overall accuracy of this new solution is between 85% – 90% across all scenarios on a daily basis.