Conformal Tabular Forecasts of African Protected-Area Visitation

Abstract

Protected areas depend on stable visitor flows for funding and local livelihoods. Public logs in the new protected area dataset cover pre-pandemic visitation to hundreds of African protected areas. We study next-year forecasting with a strong tabular neural baseline, the FT-Transformer, and quantify uncertainty with split conformal prediction. Using a strict temporal split (train to 2015, validation 2016–2017, calibration 2018–2019, test 2020–2023), the FT-Transformer improves accuracy over a multi-layer perceptron by a wide margin. On the held-out test years we obtain root mean squared error near 2.95×105 visitors and median prediction-interval width around 6.2×104 for 90% target coverage. Year-wise coverage stays near target for 2020–2022 and softens in 2023. Simple ablations show that adding only area-identity and basic geography plus three lags of visitors captures most of the signal, while a plain MLP is less reliable. Our contributions are a reproducible forecasting and uncertainty pipeline for protected area visitation, calibrated prediction intervals for decision support, and ablations that clarify which tabular features matter.