Explanation versus Prediction in Macroeconomic Modelling: Reconciling Econometric and Machine-Learning Evidence on Nigerian FDI Ofierohor Ufuoma Earnest

Abstract: A variable can be statistically significant in a structural model and carry no weight in a predictive one, or the reverse. This paper works through that problem using two companion analyses of the same Nigerian quarterly dataset, 2008Q2 to 2024Q4: an ARDL bounds-testing model of FDI and a leakage-free machine-learning forecasting pipeline. The interest rate makes the point cleanly. In the ARDL model it is significant and negative under both classical and HAC standard errors (coefficient -57.24, HAC p = 0.005). In the forecasting model its permutation importance has a bootstrap 95% confidence interval of [-29.34, 42.94], which spans zero, so its predictive contribution is not established at all. Exchange-rate volatility fails on both axes: insignificant in the ARDL model (p = 0.70) and with a bootstrap interval of [-12.35, 44.06]. Neither pattern is a contradiction. The two frameworks answer different questions, one about a stable structural association over the full sample and one about whether a variable improves forecasts on held-out data, and neither licenses a causal claim. We set out a three-way distinction between association, predictive usefulness and causation, and draw practical guidance for reading the two kinds of output side by side.

Keywords: Explanation versus Prediction; ARDL; Machine Learning; Causality; Nigeria.