LLM-Augmented Machine Learning for Cross-Channel Media Allocation in E-Commerce Marketing

This paper develops and evaluates an LLM-augmented machine-learning workflow for cross-channel media allocation in e-commerce marketing. The empirical analysis uses the Multi-Region Marketing Mix Modelling dataset released on Figshare, which contains anonymized marketing and purchase observations for e-commerce brands across regions and digital channels. Daily observations were aggregated to a weekly panel of 19,082 brand-region time-series records. Media spend was represented at the Google, Meta, and TikTok platform level, with adstock and saturation transformations, non-paid demand controls, seasonality, panel fixed effects, and a time-based train-validation-test design. Four models were compared: Ridge MMM, Elastic Net MMM, Random Forest, and XGBoost. Random Forest achieved the best out-of-sample sMAPE of 0.322 and log-scale R2 of 0.908 on the held-out test period, outperforming regularized linear MMM baselines while preserving an interpretable allocation layer. Counterfactual fixed-budget simulations shifted average allocation from 69.3% Google, 30.5% Meta, and 0.3% TikTok to 87.1% Google, 11.5% Meta, and 1.4% TikTok. The median predicted lift was 5.17%, and the 5%-95% trimmed mean lift was 24.56%. A constrained language-model memo layer converted the empirical outputs into channel-specific recommendations and guardrails. The study shows that MMM-style feature engineering, supervised learning, privacy-aware incrementality analysis, and LLM-assisted explanation can be combined into a reproducible decision-support pipeline for e-commerce media planning.