StockSense: Solving Inventory Management Challenges for SMEs with Machine Learning

StockSense is an open-source project developed by IssaMuiz, released on GitHub (link: https://github.com/IssaMuiz/StockSense, release date: May 23, 2026). It is a lightweight inventory forecasting tool based on Random Forest Regression that helps SMEs analyze historical sales data, predict product demand, generate replenishment recommendations, reduce stockouts and overstock issues, and promote data-driven decision-making. The tool uses Streamlit to build its interface, lowering the barrier to use for non-technical users.

SMEs face two major challenges in inventory management: excessive inventory ties up cash flow and storage space, while insufficient inventory leads to stockouts and missed sales opportunities. Traditional experience-based management struggles to handle demand fluctuations, and StockSense is an open-source ML tool created to address this pain point.

The design goal of StockSense is to enable SMEs to easily implement data-driven inventory decisions. Its core features include:

1. CSV Data Upload: Supports users to upload their own sales data;
2. Sales Insights and EDA: Automatically analyzes data features and trends;
3. Demand Forecasting: Predicts future sales volume for each product;
4. Replenishment Recommendation Engine: Provides specific replenishment quantities based on forecasts;
5. Lightweight Interface: Built with Streamlit, no complex deployment required.

Data Processing

A synthetic dataset of 901 rows and 9 columns (simulating supermarket inventory sales from January to June 2024) was used. Preprocessing steps: remove irrelevant columns, convert dates to datetime format, handle duplicates/empty values, sort by date and product ID, check for negative values, add lag features. Finally, 895 rows and 7 columns of data were obtained, divided into training set (626 rows), validation set (134 rows), and test set (135 rows).

Model Selection

RandomForestRegressor was adopted for the following reasons: handles non-linear relationships, is insensitive to feature scaling, has strong interpretability, and is stable on small datasets. Feature engineering: remove product names (retain ID encoding), experimentally select features (e.g., day-of-week feature improves performance).

Performance Evaluation

Metrics after GridSearchCV tuning:

Metric	Score
MAE	2.66
MSE	11.47
R²	-0.006
Although the R² is close to zero, the MVP prototype on synthetic data has an error of about 3 units, which has practical value.

Experimental Insights

• The 7-day rolling average feature increased MAE from 2.81 to 2.98 (performance degradation);
• The day-of-week feature reduced MAE to 2.75 (performance improvement), indicating that feature engineering requires experimental validation.

StockSense is suitable for:

1. Small retailers: No professional data team, need ready-to-use tools;
2. Startup e-commerce businesses: Simple operations, regular sales data;
3. Inventory optimization beginners: Understand the application of ML in inventory management. Its value lies in solving practical problems with simple technology and providing a lightweight solution for SMEs with limited budgets.

Future improvement directions:

1. Introduce more lag features and rolling statistics;
2. Add seasonal features (months, holidays);
3. Integrate external signals (promotions, weather);
4. Train the model using larger-scale real datasets. The aim is to enhance the model's ability to adapt to dynamic changes in real business environments.