Banks and other organizations in the financial sector, particularly payment service providers (PSPs), require accurate forecasts of transaction counts to optimize procurement planning for point-of-sale (POS) devices under budget constraints. To address this need, this study proposes a two-level hierarchical forecasting framework for POS transaction counts to support tactical decision-making. At the first level, a time-weighted Long Short-Term Memory (LSTM) model is developed to forecast monthly transaction counts over a one-year horizon. A set of candidate exogenous variables is initially identified and subsequently refined using Goal Programming Best-Worst Method (GP-BWM) to improve model performance. At the second level, a Random Forest (RF) model is employed to estimate the relative weights of daily transactions within each month based on intrinsic daily characteristics. A forward feature selection approach is further applied to identify the most influential daily features affecting transaction distribution. The proposed framework is validated through a real-world case study of a PSP in Iran. The results indicate that the time-weighted LSTM achieves high predictive accuracy with R² values of 0.90 and 0.89 on validation and test datasets for monthly forecasting, respectively. Additionally, the daily feature selection process demonstrates strong explanatory power with an R² of 0.9625. These findings confirm the effectiveness and practical applicability of the proposed framework in supporting data-driven planning.