Operational data were obtained from UNP Mart’s internal records, including: (1) Master data (products, suppliers, categories); (2) Historical transaction data (goods incoming and outgoing). Time-series sales data were used to train the ARIMA forecasting model. These data represent chronological demand observations required for time-series modeling .

This phase focused on defining the structural and behavioral architecture of the proposed inventory system using Unified Modeling Language (UML). The modeling process was intended to ensure clarity in system logic, user interaction, and data structure prior to implementation. The diagrams developed include a system flowchart, use case diagram, activity diagram, and class diagram. Each diagram represents a different level of abstraction in the system design, ranging from procedural workflows to object-oriented structures.

2.3.1 System Flowchart

The system flowchart was developed to represent the overall logical sequence of operations within the inventory application. The flow begins with user authentication, where registered users log in based on their assigned roles. After successful login, users can access master data management features, record inbound and outbound transactions, and initiate forecasting and inventory optimization processes. The forecasting module uses the ARIMA model to process historical sales data and generate predicted demand. The predicted demand is then forwarded to the EOQ module to compute optimal order quantity, Reorder Point (ROP), and Safety Stock values. The final stage involves generating reports and visualizing inventory status.

2.3.2 Use Case Diagram

The Use Case diagram defines the interaction between system actors and core functionalities. Three primary actors were identified: Admin, Staff, and Manager. The Admin has full access to manage master data, configure system parameters, and supervise forecasting and EOQ execution. The Staff is responsible for recording inbound and outbound goods transactions and initiating calculations as necessary. The Manager primarily accesses reporting features to review forecasting results, inventory levels, and optimization outputs.

2.3.3 Activity Diagram

The Activity diagram was created to visualize the system’s operational workflow. It describes the step-by-step activities performed by users, from login verification to transaction recording, and extends to forecasting and EOQ computation procedures. Decision nodes represent validation checks, such as verifying transaction completeness before updating stock levels or confirming that forecasting is ready before running an ARIMA model.

2.3.4 Class Diagram

The Class diagram models the system’s object-oriented structure by defining core entities, attributes, methods, and relationships. Key classes include User, Product, Supplier, Category, Transaction, Forecast, and InventoryCalculation. Relationships among these classes reflect database associations such as one-to-many and relational dependencies. For instance, the Transaction class is linked to the Product class, while forecasting outputs are associated with historical transaction records.

The Autoregressive Integrated Moving Average (ARIMA) model was employed for demand forecasting due to its robustness in modeling univariate time-series data with autocorrelation structures .

The modeling process consisted of:

These metrics are standard measures for evaluating forecasting performance in inventory management contexts. The final forecast output provided the Annual Demand (D) parameter for EOQ calculation.

The Economic Order Quantity (EOQ) model was applied to determine the optimal order quantity minimizing total inventory cost . The EOQ formula (1) is:

In the EOQ formulation, D represents the annual demand obtained from the ARIMA forecasting output, S denotes the ordering cost per order, and H refers to the holding cost per unit per year. The Reorder Point (ROP) was calculated as ROP = dL + SS, where d is the average daily demand, L is the lead time. To account for demand uncertainty during the replenishment period, Safety Stock (SS) was determined using a probabilistic demand-variability approach (2):

Where Z denotes the service level factor corresponding to the desired probability of avoiding stockouts, σd​ represents the standard deviation of daily demand, and Ladjusts demand variability over the lead-time horizon, assuming independent daily demand.

The system’s integration mechanism follows a sequential analytical flow: historical transaction data are processed through the ARIMA forecasting engine to generate predicted demand (D), which then serves as the primary input for the EOQ calculation to determine the optimal order quantity and corresponding ROP values. While the EOQ model assumes deterministic cost parameters and relatively stable ordering conditions , incorporating ARIMA forecasting enhances the system’s responsiveness to demand variability and improves the robustness of inventory decision-making under dynamic conditions .

Safety Stock (SS) was determined using a demand-variability approach based on the standard deviation of demand during lead time, formulated as:

The application was implemented using a Model-View-Controller (MVC) architecture with CodeIgniter 3 (PHP) for the backend logic and MySQL for the relational database. Bootstrap was utilized to ensure a responsive interface design. A Python-based forecasting engine performed ARIMA modeling and communicated with the web application via an API.

We validated the system using multiple complementary approaches. We conducted Black-Box Testing to verify all features functionally without examining the source code, ensuring compliance with predefined specifications. We also conducted User Acceptance Testing (UAT), allowing end users to assess system usability and the appropriateness of functional operations. To evaluate forecasting accuracy, we compared ARIMA performance metrics, including Mean Absolute Percentage Error (MAPE) and Root Mean Square Error (RMSE), with actual demand values. We assessed system effectiveness by examining the functional pass rate from Black Box Testing to confirm that all features operated according to specifications, the ARIMA model’s forecast accuracy to ensure predicted demand closely matched historical patterns, and the logical consistency of EOQ-derived Reorder Point (ROP) and Safety Stock values to verify that optimization outputs were mathematically sound and operationally feasible. Together, these validation steps demonstrate that we developed a system that is technically robust, analytically reliable, and practically applicable for data-driven procurement decision-making.

The system integrates a Python-based forecasting engine that uses the ARIMA model to predict time-series demand. The computational process begins by extracting historical sales data from the “Goods Out” transaction table, which serves as the primary time-series input. The ARIMA engine processes these data to estimate future demand levels and produces the projected Annual Demand parameter (D). The predicted demand is subsequently transmitted to the EOQ module, where it is combined with ordering cost (S) and holding cost (H) parameters stored in the master data to compute optimal inventory decisions.

To evaluate forecasting performance, model accuracy was assessed using the Mean Absolute Percentage Error (MAPE), the Root Mean Square Error (RMSE), and the Mean Absolute Error (MAE). The evaluation results are presented in Table 1.

Based on AIC and BIC comparison, the ARIMA(1,1,1) model was selected as the optimal specification for forecasting demand. The MAPE value below 10% indicates that the ARIMA model provides high forecasting accuracy and is suitable for operational inventory planning. The relatively low MAE and RMSE values further confirm that prediction deviations remain within an acceptable operational tolerance range.

Based on this integrated analytical process, the system generates three principal optimization outputs: the Optimal Order Quantity (Q*), representing the most economical purchase quantity per replenishment cycle; the Reorder Point (ROP), calculated as the product of average daily demand and lead time plus Safety Stock; and the Safety Stock (SS), which functions as a protective buffer against uncertainty during replenishment periods. Safety Stock was computed using the demand standard deviation during the lead time to accommodate demand variability and reduce stockout risk.

To evaluate the financial implications of the optimization approach, a comparative cost simulation was conducted between the existing ordering policy and the EOQ-based policy. The comparison was simulated over a one-year operational period using projected annual demand derived from the ARIMA forecast. The results are summarized in Table 2.

The implementation of the EOQ-based policy resulted in a total inventory cost reduction of approximately 22.73%, indicating improved cost efficiency through optimized order sizing and inventory balancing. All calculated values are presented on a dedicated EOQ Calculation page, where users can review detailed results and export them as PDFs for reporting and managerial evaluation.

We conducted functional validation using Black-Box Testing to verify that the system complied with all predefined requirements. Our evaluation emphasized input–process–output verification across the system’s critical modules, including user authentication, CRUD operations for goods, transaction processing, ARIMA-based forecasting, and EOQ-based inventory optimization. Each module was tested against realistic operational scenarios to ensure that the system could handle typical and exceptional use cases effectively. Table 3 summarizes the tested features, corresponding scenarios, expected results, and validation outcomes.

During testing, we simulated various interactions, such as adding, editing, and deleting product data; recording incoming and outgoing stock transactions; executing EOQ calculations to generate optimal order quantities, Reorder Points (ROP), and Safety Stock (SS); and running ARIMA-based demand forecasting. We also verified that the system correctly handled invalid inputs, such as incorrect login credentials or transaction anomalies, and produced appropriate error messages. All modules successfully met their expected outcomes, indicating that the system processed data accurately, maintained data integrity, and performed computations reliably.

We observed that all tested features passed the validation criteria, confirming that the system functions as intended and satisfies both operational and analytical requirements. Beyond functional correctness, we verified that users can access detailed test results and export them as PDFs for reporting and managerial evaluation, supporting transparency and traceability of system operations. These results demonstrate that we successfully implemented a robust, reliable, and user-oriented inventory management system that supports data-driven decision-making, enhances operational efficiency, and provides actionable insights for inventory control in MSME contexts.

Integrating ARIMA forecasting with the EOQ model establishes a structured link between predictive analytics and prescriptive optimization. Classical EOQ theory assumes deterministic demand and constant cost parameters [12], [13]. In practice, however, demand uncertainty undermines the validity of static inputs. Embedding forecast-derived annual demand into EOQ makes a traditionally fixed parameter dynamically estimated, improving replenishment decision accuracy.

This integration reflects a broader methodological shift in operations management, where predictive models inform optimization engines to enhance decision robustness. Historical data inform demand forecasts, forecasts inform replenishment parameters, and optimization outputs guide procurement actions. Such a closed-loop architecture aligns with contemporary research advocating data-driven inventory control frameworks that combine forecasting accuracy with cost-minimization logic [10], [17].

The incorporation of probabilistic Safety Stock enhances inventory resilience by explicitly accounting for demand variability during the replenishment period. By scaling the standard deviation of daily demand over lead time and adjusting with the service level factor Z, the system translates stochastic demand patterns into quantifiable inventory buffers. This approach aligns with established principles in stochastic inventory control, where safety stock functions as a risk-mitigation mechanism that balances the likelihood of stockouts against holding-cost exposure [11], [30].

Crucially, the buffer is calibrated to predefined service level objectives rather than relying on arbitrary heuristics or fixed percentages. This calibration ensures that the probability of stockouts remains within managerial tolerance limits, providing a defensible and analytically grounded method for mitigating risk. In contexts where demand variability constitutes the principal source of uncertainty, probabilistic adjustments offer superior control compared with ad hoc or reactive buffer policies. As a result, inventory management shifts from corrective responses to proactive, quantitatively justified risk management, enhancing both operational reliability and cost efficiency [5], [6], [17].

The incorporation of probabilistic Safety Stock further strengthens inventory resilience. By scaling the standard deviation of daily demand over the lead time and adjusting it with the service level factor Z, the system explicitly models demand variability over the replenishment horizon. This approach is consistent with stochastic inventory control theory, in which safety stock serves as a risk-mitigation mechanism that balances the probability of stockouts and holding-cost exposure.

Importantly, the buffering mechanism is calibrated to service level objectives rather than arbitrary percentage adjustments. This ensures that stockout risk remains within managerial tolerance thresholds. In environments where demand variability is the primary source of uncertainty, such probabilistic adjustments are more defensible than heuristic buffer additions. Consequently, inventory control shifts from reactive correction toward quantitatively grounded risk management.

The one-year cost simulation revealed a 22.73% reduction in total inventory costs under the EOQ-based policy compared with the existing ordering approach. This improvement suggests that prior procurement practices relied on heuristic rules, fixed intervals, or experience-based judgment rather than total-cost minimization principles [9], [10].

The observed cost reduction arises not solely from order quantity optimization but also from the interaction among accurate demand estimation, probabilistic buffering, and optimized replenishment intervals. This confirms that forecasting and optimization function as interdependent components of an integrated system, producing measurable financial benefits and improving transparency in decision-making [11], [12], [17].

From an implementation standpoint, the system demonstrates that advanced analytical integration can be operationalized using standard web technologies. The Model-View-Controller (MVC) architecture in CodeIgniter 3 separates business logic and presentation layers, facilitating maintainability and scalability [7]. Integrating a Python-based forecasting engine with a PHP-based web framework confirms that traditional web stacks can effectively accommodate modern data science tools [9], [23].

This hybrid architecture enhances accessibility and enables role-based access to real-time inventory information across devices. Such accessibility is particularly relevant for MSMEs transitioning from manual or spreadsheet-based systems toward digital platforms [1], [2], [3].

Despite positive results, several limitations warrant acknowledgment. The EOQ model assumes constant ordering and holding costs, which may not reflect real-world fluctuations such as quantity discounts or supply chain disruptions [12], [30]. ARIMA captures linear temporal dependencies but does not explicitly incorporate exogenous factors such as promotions or macroeconomic shocks [24], [26]. In demand environments characterized by seasonality or structural breaks, extended models such as SARIMA or hybrid machine-learning approaches may yield superior forecasting performance [28].

The cost simulation was conducted over a one-year operational horizon. Longer-term validation is necessary to assess robustness under variable economic conditions. Future research may explore multi-period validation, multi-warehouse coordination, comparative analysis with advanced forecasting techniques, and integration with emerging AI-driven predictive models [14], [19], [20], [26].