Abstract
Digital short-term-rental platforms combine structured property information, historical engagement, platform-management information, and host-written text, yet it is unclear whether multimodal gains persist in later periods. This study develops a privacy-aware longitudinal forecasting framework using five nationwide New Zealand Inside Airbnb snapshots (October 2025 to August 2026). Listings are matched across time, and cumulative review growth is standardised to a 90-day horizon as a review-based proxy for platform activity. Two windows (W1, W2) are used for development and a later window (W3; N=46,471) is the held-out temporal test. Five models, from a recent-activity persistence baseline to deep multimodal fusion, are compared. The persistence baseline achieved the lowest W3 MAE (1.974). Machine-learning fusion was the strongest learned model under the primary specification (MAE 2.476, R2=0.237, Spearman ρ=0.696), but only narrowly ahead of structured XGBoost, and removing text improved W3 error, so the gain cannot be attributed to text. Historical engagement dominated SHAP importance and ablation, deep fusion did not improve temporal generalisation, and eight of 33 regions had negative R2. The contribution is methodological: model-level gain, modality-specific value, and temporal robustness are evaluated separately. The target is not interpreted as bookings, occupancy, demand, or revenue.