Forecasting Nigerian Foreign Direct Investment One Quarter Ahead: A Leakage-Free Comparison of Six Regression Models Ofierohor Ufuoma Earnest

Abstract: This study compares six supervised regression models for forecasting Nigerian foreign direct investment one quarter ahead, using 67 quarterly observations from 2008Q2 to 2024Q4. Two feature sets are tested: a compact core set of six contemporaneous predictors, and an enhanced set adding lags, four-quarter rolling means and percentage changes, all shifted so that no future information enters. Both sets are scored on an identical evaluation window, which matters because an earlier version of this work compared them on different windows and reached the opposite conclusion. Models are tuned with four-fold expanding-window cross-validation and evaluated by expanding-window walk-forward prediction. Gradient boosting on the enhanced feature set performs best, with MAE 385.59, RMSE 523.37 and out-of- sample R-squared of 0.0257. It is the only one of twelve model-specification pairs to beat a holdout-mean benchmark. Naive persistence returns RMSE 691.51 under both feature sets, confirming that the two are being scored on the same observations. Enhanced features improve on core features in three of five model classes. The main finding is negative and, we think, useful: at this sample size and horizon, quarterly Nigerian FDI is close to unforecastable, and eleven of twelve approaches do worse than predicting the mean.

Keywords: Foreign Direct Investment; Time-Series Forecasting; Feature Engineering; Walk-Forward Validation; Nigeria.