Support Vector Regression for Forecasting Ghana’s Usd to Ghs Exchange Rate_A Comparative Study With Econometric and ML/QML Baselines.
No Thumbnail Available
Date
2025-11
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
UENR
Abstract
Ghana continues to struggle with accurate exchange-rate forecasting because to structural breaks,
nonlinear macroeconomic dynamics, and the shortcomings of conventional econometric models
like ARIMA, ARIMAX, and VAR. These Bank of Ghana institutional workhorses are predicated
on linearity, stationarity, and steady parameters—conditions that are increasingly broken in the
fluctuating USD/GHS market. Studies that have already been conducted in Ghana frequently
benchmark forecasting models inconsistently, lack unified assessment pipelines, and fail to
evaluate operational usability or computational efficiency. The question of whether contemporary
machine-learning approaches can produce much better and more trustworthy forecasts for
decision-making is thus left open from a methodological and policy standpoint.
This study investigates whether a kernel-based Support Vector Regression (SVR) model provides
better operational robustness and prediction accuracy than the baseline models used by the Bank
of Ghana. We employ a two-phase strategy using a monthly macroeconomic panel spanning 2014–
2023: (i) an ex-post backtest (train 2014–2022; test 2023) and (ii) an ex-ante operational forecast
for January–December 2024 following model refitting on data through December 2023. Using
accuracy metrics (MAE, RMSE, MSE, MAPE, R²), computational cost indicators (training time,
inference time, peak memory), and uncertainty quantification (native intervals for econometric
models, bootstrap intervals for ML/QML), all models, econometric, classical ML, and exploratory
quantum ML, are assessed within a single, leakage-safe pipeline.
According to the 2023 backtest results, SVR significantly outperforms ARIMA and VAR,
although ARIMAX is still the most effective conventional comparator. While quantum variations
show promise but inconsistent accuracy, classical machine learning methods (LSTM, XGBoost)
offer moderate increases. A policy-ready Streamlit prototype that operationalizes model selection,
scenario analysis, uncertainty reporting, and exportable outputs for institutional usage is shown in
the study's conclusion.
Description
Keywords
Quantum Machine Learning, Classical Algorithms, Micro-Economic Forecasting