Rakuten Advertising currently builds a separate revenue forecasting model per client with open-ended feature adjustment, producing inaccurate forecasts. This was addressed by benchmarking five time series foundation models, both zero-shot and under seven parameter-efficient fine-tuning methods, against statistical and machine-learning baselines on NielsenIQ receipt-level e-commerce data across three verticals. Covariate-aware, zero-shot Chronos was strongest in every vertical, cutting Weighted Absolute Percentage Error by roughly 45% against the best baseline. Adaptation helped in only 17% of runs and degraded the strongest backbones. The project delivers a standardized modeling and evaluation pipeline and a client-facing dashboard improving forecast accuracy, consistency, and interpretability.
Watch the team present this project at 00:59:18 in the session recording here.
Keywords: time series forecasting, foundation models, parameter-efficient fine-tuning, LoRA, affiliate marketing, consumer behavior data, revenue forecasting, interpretability
Faculty Advisor
Anil Chaturvedi has over 35 years of professional experience at companies such as AT&T Bell Labs, Kraft Foods, Capital One, and Accenture. He has provided consulting services to Bank of America, Fannie Mae, Johnson & Johnson, and Procter & Gamble. His general research interests include enhancing business value using data science. He has patented and published advanced algorithms for predictive modeling, market segmentation, new product development, product positioning, customer loyalty, consumer promotion mix optimization, and brand strategy. He earned his PhD from Rutgers University and an MBA from IIM Ahmedabad, India.