Car Sales in a Fully Import-Dependent, High-Ownership Country, New Zealand: Evaluating Forecasting with Deep Learning and Econometric Methods

This study examined the dynamics of private car sales in New Zealand, a fully import-dependent and tariff-free economy with one of the highest car ownership rates globally. Using monthly data on small vehicle sales (cars, sport utility vehicles, and passenger vans) from 1998 to 2024 (324 observations) obtained from the Motor Industry Association, the study investigated market behavior and external shocks through both a deep learning approach, Long Short-Term Memory (LSTM), and an econometric framework, Exponential Generalized Autoregressive Conditional Heteroskedasticity (EGARCH). The study further examined methodological challenges associated with validating the LSTM model, with particular emphasis on the application of time-series k-fold cross-validation (TSCV) to data characterized by a long-term upward trend. Specifically, it assessed potential distributional inconsistencies between the train and test datasets of individual folds and showed reliability issues in k-fold cross-validation outcomes. Moreover, by construction, the TSCV framework systematically excludes important data segments from model estimation. This exclusion limits the ability of the validation procedure to provide a fully robust assessment of a model ultimately trained on the complete dataset. This study has also explained an underlying cause of negative out-of-sample R-square in k-fold cross-validation tests using a methodological framework and empirical evidence. Forecast performance comparisons between the EGARCH and LSTM models revealed statistically significant differences in their predictions. The study also identified statistically significant differences between models trained on split datasets and those trained on the full dataset. Results suggested that models trained on the full dataset provide more robust predictions and perform comparably to EGARCH, particularly at the beginning of the forecast horizon. The LSTM model effectively captured recent upward drifts and nonlinear patterns, generating point forecasts of expected sales and reflecting long-term trends. In contrast, the EGARCH model captured volatility persistence and mean reversion, offering forecasts that quantify uncertainty and risk. An integrated LSTM-EGARCH framework successfully captures both nonlinear and stochastic dynamics, producing forecasts that are more informative for policy formulation and managerial decision-making. From a policy perspective, the findings indicated that car sales in New Zealand have a long-term upward trend, potentially supported in part by tariff relaxation. However, sales display pronounced volatility clustering, strong mean reversion, and a high likelihood of extreme fluctuations in response to external shocks. Full import dependency is argued to be one contributing factor to this volatility. Sales declined sharply during the COVID-19 pandemic but rebounded rapidly, possibly reflecting strong institutional foundations. Forecasts for 2025-2027 suggest that sales growth is likely to continue, albeit at a marginal pace, approaching a potential peak level. In terms of overall lessons for other tariff-free, import-dependent economies, tariff relaxation, probably supported by favorable measures such as roads and effective institutional backups, can sustain car sales growth and support recovery, even amid market volatility and external shocks.