Back to Projects
Forecasting / Retail Analytics

Demand Forecasting with ARIMA and Prophet

A comparative forecasting system that applies multiple time-series methods to retail sales data, achieving a 31% reduction in mean absolute error and providing clear demand projections for inventory planning.

Type
Time Series Analysis
Domain
Retail / Demand Planning
Best MAPE
22.79%
Status
Completed
TIME SERIES FORECASTING
+15%

The Challenge

Retailers face volatile demand fluctuations driven by seasonality, promotions, external events, and shifting consumer behaviour. Baseline forecasting methods often fail to capture these temporal patterns, leading to either overstocking (tying up capital) or understocking (losing sales).

Accurate demand forecasting directly impacts profitability. Even small improvements in forecast accuracy translate into significant savings in inventory costs and reduced lost revenue from stockouts.

Approach

01
Time Series Decomposition
Decomposed the historical sales data into trend, seasonal, and residual components to understand the underlying patterns driving demand fluctuations.
02
Feature Engineering
Created temporal features including lag variables, rolling statistics, holiday indicators, and seasonal encodings to enrich the forecasting models.
03
Model Development
Built and compared ARIMA/SARIMA models for statistical rigour and Prophet for handling multiple seasonality patterns and holiday effects.
04
Evaluation and Selection
Evaluated models using RMSE, MAE, and MAPE against baseline methods, selecting the optimal approach for each product category.

Results

31%
Reduction in mean absolute error over baseline forecasting methods
22.79%
MAPE achieved - the average forecast deviation from actual demand
Multi-Model
Best model selected per product category for optimal accuracy

The best-performing model achieved a mean absolute percentage error (MAPE) of 22.79% and a 31% reduction in MAE over the baseline, giving demand planners materially tighter projections for inventory and staffing decisions.

The Prophet-based approach performed best on products with strong seasonal patterns and holiday effects, while ARIMA excelled on more stationary series. The hybrid approach of selecting the best model per category delivered the overall improvement, confirming that no single forecasting method dominates across all demand profiles.

For retail operations, this level of forecast improvement translates directly into reduced overstocking costs and fewer lost sales from stockouts - the two largest controllable drags on inventory-dependent profitability.

Technology Stack

Python Statsmodels Prophet ARIMA SARIMA Pandas Matplotlib
Interested in this work or something similar?