Machine Learning-Based Retail Sales Forecasting System: From Data Integration to Random Forest Practice

This article introduces an open-source end-to-end retail sales forecasting project. It integrates multi-source data to build a random forest regression model for weekly sales prediction, reveals the dominant role of store size and seasonal factors in sales performance, helps enterprises optimize inventory allocation, reduce unsold inventory costs, and arrange staffing reasonably, with direct commercial value.

The retail environment is highly dynamic and significantly influenced by localized external factors (such as holidays and inflation). The goal of this project is to build an end-to-end regression machine learning framework to accurately predict weekly sales of different stores and departments. Through high-precision demand forecasting, operators can optimize inventory allocation, reduce unsold inventory costs, and arrange staffing reasonably during holidays, which has direct commercial value for chain retail enterprises.

The project uses over 400,000 historical records covering data from 45 stores. The raw data is scattered across three tables: store information table (area, type), feature data table (temperature, oil price, CPI, unemployment rate, promotion status), and sales data table (weekly sales). Preprocessing includes: extracting year/month/week information from dates to capture micro-seasonality; filling missing promotion values with 0 and economic indicators with median values; merging the three tables into a unified data frame based on store and date indices.

Exploratory Data Analysis: Visualize variable relationships through KDE histograms (right-skewed sales distribution), box plots (performance differences between store types), scatter plots (relationship between sales and macro indicators), heatmaps (feature correlations), etc. Statistical Validation: Use ANOVA F-test to verify mean differences between store types, two-sample T-test to prove higher sales in holiday weeks, and Pearson correlation coefficient to confirm the linear relationship between store area and sales. Model Evolution: Gradually compare algorithm performance from linear regression (baseline) → decision tree → random forest (n_estimators=50, max_depth=10).

Core Findings: 1. Store size (area) and micro-seasonality dominate sales, with extremely low correlation to macroeconomic indicators; 2. Departments 92/95/38 are the core of revenue; 3. The random forest model performs best and can effectively handle non-linear holiday peaks. Engineering Implementation: Provide deployment-ready Jupyter Notebook, build.py parsing environment, serialized model files, and a full-stack interactive dashboard using FastAPI+Vite+React (supports dark theme and glassmorphism UI).

Technical Insights: When building prediction models, priority should be given to basic business features (such as store area, seasonality) rather than over-pursuing complex macro indicators. Industry Applications: This solution provides a directly implementable reference architecture for retail enterprises, with clear paths in all links from data cleaning, feature engineering to model deployment, helping enterprises improve operational efficiency.