Retail businesses face a constant dilemma: how much inventory to stock. The challenges are significant:

• Overstocking: Ties up capital in unsold inventory, increases storage costs, and leads to markdowns that erode profit margins
• Understocking: Causes lost sales, damages customer satisfaction, and gives competitors an opportunity to capture market share
• Seasonal Variability: Sales patterns fluctuate dramatically by season, making it hard to predict future demand
• Regional Differences: Different US markets have vastly different buying patterns and preferences

Traditional forecasting methods (like "last year plus 10%") fail to capture complex seasonal patterns, trend changes, and market dynamics. Manual forecasting in Excel is time-consuming, inconsistent, and doesn't provide confidence intervals to quantify uncertainty.

I built a comprehensive time-series forecasting solution that combines statistical rigor with practical business application:

• Exploratory Data Analysis: Analyzed historical sales data to identify trends, seasonality, and outliers using Pandas and visualization libraries
• Feature Engineering: Created time-based features (month, quarter, holiday indicators) and calculated rolling averages to capture momentum
• Prophet Modeling: Implemented Facebook Prophet for automatic detection of yearly seasonality, weekly patterns, and holiday effects
• ARIMA Modeling: Built traditional ARIMA models for comparison and to capture autocorrelation patterns in the time series
• Model Evaluation: Compared models using RMSE, MAE, and MAPE on holdout test data to select the best performer
• Power BI Integration: Exported forecasts with confidence intervals to Power BI for stakeholder dashboards

The solution provides 12-month forward forecasts with upper and lower bounds, allowing planners to balance risk tolerance with inventory decisions.

Key Insights Delivered:

• Q4 Sales Spike: Model identified 38% increase in Q4 sales, enabling proactive inventory buildup for holiday season
• Regional Patterns: West Coast shows 23% higher summer sales compared to other regions, informing regional allocation
• Product Mix Shifts: Footwear sales growing 15% YoY while apparel flat, suggesting portfolio rebalancing
• Confidence Intervals: Upper/lower bounds help planners make risk-adjusted decisions (order conservatively vs aggressively)
• Early Warning System: When actual sales deviate from forecast by 15%+, alerts trigger investigation into market changes

The forecasting model is now used monthly by supply chain planners to make purchasing decisions 6-12 months in advance. The confidence intervals allow them to order base quantities at the lower bound, with option to order more if demand signals strengthen.

Modeling Workflow: Data ingestion → EDA & visualization → Feature engineering → Train/test split → Prophet & ARIMA training → Model evaluation → Hyperparameter tuning → Forecast generation → Power BI export

Technical Highlights:

• Prophet Configuration: Tuned yearly seasonality strength, added custom holiday calendar (Black Friday, Christmas), and adjusted changepoint sensitivity
• ARIMA Selection: Used auto ARIMA to find optimal (p,d,q) parameters through AIC minimization
• Cross-Validation: Implemented time-series cross-validation with expanding window to avoid data leakage
• Ensemble Approach: Combined Prophet and ARIMA forecasts weighted by historical performance for robust predictions
• Automated Reporting: Script exports forecast CSV that Power BI imports automatically, updating dashboards monthly

All code is well-documented in Jupyter notebooks with markdown explanations of methodology, model selection rationale, and interpretation of results. The GitHub repository includes sample data and step-by-step instructions for reproduction.