Retail Forecasting Models: Beyond Algorithms (Part 1)

In today's data-driven retail environment, effective sales forecasting has become a critical capability for maintaining a competitive advantage. While algorithms like ARIMA, Prophet, and machine learning techniques get much attention, the actual forecasting models - the conceptual frameworks that structure how we approach prediction - are equally important.

This blog explores the various forecasting models retailers use to predict demand, examining their applications, strengths, and limitations.

1. Time Series Models

Description: The most straightforward approach to retail forecasting, time series models analyze historical sales data to identify patterns and project them into the future.

How it works: These models treat historical sales as a sequential dataset, identifying trends, seasonality, and cyclical patterns. They rely primarily on the target variable's own past values.

Best for: Stable products with consistent demand patterns and sufficient historical data.

Retail Examples:

• Costco: With relatively stable demand for staple products and bulk grocery items
• Dollar General: For forecasting consistent demand of essential household items
• Chewy.com: For recurring pet supply purchases with predictable consumption patterns

Limitations:

• Assumes future patterns will resemble past patterns
• May struggle with new products or volatile markets
• Often misses external influences unless explicitly incorporated

2. Causal/Regression Models

Description: These models incorporate cause-and-effect relationships between sales and various influencing factors.

How it works: Regression models establish mathematical relationships between sales (dependent variable) and influencing factors (independent variables) such as pricing, promotions, economic indicators, and competitor actions.

Best for: Understanding and quantifying specific drivers of demand, pricing optimization, and promotional planning.

Retail Examples:

• Best Buy: For understanding how product price changes and how competitors' promotions affect electronics sales
• Walmart: For measuring impacts of economic indicators on consumer spending patterns
• Wayfair: For quantifying how shipping costs and delivery times influence furniture purchases

Limitations:

• Requires identifying and measuring all relevant variables
• Can be complex to maintain as relationships change
• Needs continuous recalibration

3. Hierarchical Forecasting Models

Description: These models create forecasts at multiple levels of aggregation (company, region, store, department, category, SKU) and reconcile them for consistency.

How it works: Forecasts are generated at different levels of the product and location hierarchy, then reconciled using top-down, bottom-up, or middle-out approaches.

Best for: Complex retail operations with thousands of SKUs across multiple locations.

Retail Examples:

• Target: For coordinating forecasts across departments, categories, and individual stores
• Kroger: For managing grocery forecasts from company level down to individual store departments
• Amazon: For managing inventory across vast product categories and distribution centers

Limitations:

• Reconciliation can introduce distortions
• Computational intensity increases with hierarchy complexity
• Requires a clear hierarchical structure

As we've seen, forecasting models like time series, causal regression, and hierarchical forecasting play a crucial role in predicting retail demand based on historical patterns and structured frameworks. However, not all products have stable sales trends, and many external factors can heavily influence consumer behavior.

In the next part of this discussion, we'll explore more advanced and specialized AI forecasting models designed to address new product launches, promotional events, external influences, and AI-driven predictions.