Foreign Exchange Cash Demand Forecasting: A Machine Learning-Driven Multi-Currency Intelligent Scheduling System at Bank Branch Level

This project built a multi-currency foreign exchange cash demand forecasting system for Nigerian banks, using machine learning technology to predict the daily net cash flow of USD, GBP, and EUR at the branch level. It aims to address the limitations of traditional manual forecasting and optimize cash distribution and capital allocation efficiency. The project covers the complete process from data construction, feature engineering, model selection to business application, providing a reference framework for financial operation optimization.

Foreign exchange cash management is a complex challenge for banks in emerging markets like Nigeria. Traditional manual forecasting relies on intuition and spreadsheet averages, which cannot capture the interaction of complex factors. The core problem is to accurately predict the next day's foreign currency cash delivery/withdrawal volume for each branch, which directly affects operational efficiency, customer satisfaction, and capital utilization. Banks face a double dilemma: insufficient cash leads to customer churn, while excess cash ties up non-interest-bearing capital. The unique characteristics of the Nigerian market include: multi-currency parallelism (different demand patterns for USD/GBP/EUR), significant branch differences, seasonal fluctuations (pay weeks/end of month/holidays), exchange rate impacts, and limitations of manual forecasting.

The project uses a synthetic dataset to simulate the statistical characteristics of foreign exchange transactions of Nigerian banks, covering working day data from January 2022 to March 2026 (excluding weekends/holidays). There are 253 branches for USD, 153 each for GBP and EUR, totaling 598,130 records. Transaction amounts are standardized by denomination (multiples of USD 100, multiples of GBP/EUR 50). Core features include: lag and rolling features (capturing short-term momentum, weekly/monthly cycles), calendar features (day of the week, end of month/quarter, pay week, exchange rate returns), holiday features (markers and distances before and after Nigerian holidays). At the same time, attention is paid to data leakage prevention, using only the previous day's data, pre-known calendar/holiday features, and the previous day's exchange rate returns.

Various models were evaluated (leakage-safe random forest, magnitude-weighted RF, Huber gradient boosting, etc.), and Huber GBR (robust to peak days) was finally selected. Branch-level strategy: branches with sufficient data use hybrid models or magnitude-weighted RF, while branches with sparse data fall back to the global model. Performance: USD uses robust Huber GBR (WAPE 0.3238, MAE 901.51, R² 0.8363); EUR/GBP both use Huber GBR (EUR WAPE 0.4083, MAE 550.08, R² 0.7729; GBP WAPE 0.4625, MAE 545.94, R² 0.7415). The main metric is WAPE, and the USD model performs best due to rich data.

The dates with the largest prediction errors are concentrated at the end of the month, end of the quarter, and before holidays (e.g., December 31 and March 31 at the end of the year). On these dates, customer behavior is driven by external factors (salaries, regulations, travel), making historical averages difficult to predict. The pay week feature (10 days before the end of the month) is a strong predictor. Standard random forests perform poorly on extreme demand days because they optimize average error rather than peak day accuracy; magnitude weighting and Huber loss solve this problem.

The model generates 7-day forward forecasts (April 7-13, 2026), outputting the daily predicted net cash flow for each branch to drive scheduling. Operational optimizations include: optimizing delivery routes to reduce logistics costs, reducing capital occupation, improving service levels on high-demand days, and providing decision support for management.

Technical highlights: multi-currency reuse architecture (USD pipeline reused for GBP/EUR), synthetic data generation (methodological value when no real data is available). Conclusion: The project provides a complete technical solution, demonstrating the value of ML in financial operation optimization. Its data leakage prevention, business cycle mining, and pursuit of model robustness reflect engineering rigor and can serve as a reference for other financial institutions.