Forecasting Method Selection for Digital Marketing Budget Pacing: A Seasonality-Aware Comparison of MAPE and RMSE Trade-offs on Public Retail and Sponsored-Search Benchmarks

Abstract

Digital marketing budget pacing is an operationally sensitive task for U.S. e-commerce teams, in which misallocated spend can cause both over-delivery waste and under-delivery revenue loss. Short-horizon forecasts of demand and conversion activity are a key input to this task, yet practitioners face an often-overlooked tension between error metrics that can disagree on which forecasting method should be preferred. This paper reports a seasonality-aware empirical comparison of forecasting approaches on three public datasets that serve as proxies for U.S. retail demand, European retail demand, and sponsored-search conversion activity, benchmarking classical statistical models (SARIMA/SARIMAX and exponential smoothing state-space models), the Prophet additive framework, gradient boosting machines (XGBoost, LightGBM), and neural architectures (DeepAR, Temporal Fusion Transformer). A rolling-origin protocol is adopted with dual-metric reporting in Mean Absolute Percentage Error (MAPE) and Root Mean Squared Error (RMSE). The results show systematic disagreement between MAPE and RMSE rankings on series with strong promotional spikes: gradient boosting methods lead under RMSE on the retail-style series, while on the intermittent sponsored-search series the Temporal Fusion Transformer and DeepAR produce the lowest MAPE values. The analysis yields a metric-aware selection heuristic that links series-level characteristics to method--metric pairings, informing forecasting method choice for short-horizon demand and conversion tasks relevant to budget pacing rather than prescribing a direct pacing controller.