This project applies classical time-series forecasting methods (ETS, ARIMA, and TSLM) to Australian wine sales data to compare model performance across wine varietals. By evaluating forecast accuracy using multiple metrics, the analysis demonstrates how demand patterns differ by product type and why model selection should be varietal-specific. An interactive Shiny application enables users to explore trends, seasonality, and forecasts in real time. The live application is available here: Australian Wine Sales Forecasting App

Motivation & Objectives

ARIMA

ETS

TSLM

December 2025

This project analyzes monthly Australian wine sales across multiple wine varietals with the goal of understanding underlying time-series patterns and identifying the most accurate forecasting approach for each varietal. Using classical forecasting methods, the project seeks to answer the following questions:

• How do trend and seasonality differ across wine varietals?
• Which forecasting models (ETS, ARIMA, or TSLM) perform best for each varietal?
• How does model performance change based on forecast horizon and evaluation metric?
• How can interactive tools improve interpretability and model comparison for non-technical users?

Accurate demand forecasting is critical for inventory planning, production scheduling, and distribution strategy in consumer goods markets. By comparing models across varietals, this project demonstrates how forecast performance is context-dependent, and why a single modeling approach may not be optimal across all product categories.

Data & Exploratory Analysis

The dataset consists of monthly Australian wine sales by varietal, measured in thousands of liters. Initial exploratory analysis revealed clear seasonal patterns, long-term trends, and differences in volatility across varietals.

Wines sales by varietal (1980-1995)

To better understand these dynamics, STL decomposition was applied to individual time series, separating each into trend, seasonal, and remainder components. This decomposition highlighted how some varietals exhibit strong, stable seasonality, while others show weaker or evolving seasonal effects.

STL decomposition of Sweet Wine varietal

Modeling Approach

Three classes of forecasting models were implemented and compared:

• ETS (Exponential Smoothing) models to capture trend and seasonality through state-space formulations
• ARIMA models to account for autoregressive and moving-average structure
• TSLM (Time Series Linear Models) incorporating trend and seasonal terms

Models were trained on historical data with user-defined training windows and evaluated across multiple forecast horizons. Forecast accuracy was assessed using RMSE, MAE, and MAPE, allowing performance comparisons under different business-relevant criteria.

Rather than selecting a single best model globally, performance was evaluated by varietal, reinforcing that different products require different forecasting strategies.

12-month Sweet Wine sales forecast overlaid over actual sales (1994-1995), using an ARIMA model. Includes 95% confidence intervals.