The COVID-19 pandemic fundamentally shifted consumer habits, as confirmed by an October 8, 2020, UNCTAD study involving 3,700 respondents across nine countries. The survey highlighted a surge in online shopping frequency, creating a disconnect between historical forecasting models and actual customer behavior. Relying solely on historical data risks skewing your perspective, as old patterns no longer reflect current market dynamics.
Forecasting Challenges and Data Quality Control
The core issue today isn't so much algorithm accuracy as it is data quality. For instance, when a retail chain closes due to a lockdown, zero sales in the system are often misinterpreted as a lack of demand, when in reality it is a «supply deficit.» A situation where a product is physically unavailable in stock is known as a stockout. It is critical to distinguish between actual sales and unmet demand; otherwise, the model will train on false signals. Use channel segmentation—modeling demand separately for brick-and-mortar locations and online channels, as customer behavior in these segments currently varies significantly.
Methodologies for Operating Under Uncertainty
To verify model reliability, implement rolling backtests. This technique involves testing forecasts across successive historical intervals (e.g., forecasting March based on February, then April based on March) without peeking at «future» data. This allows you to determine if your model is «drifting.» Concept drift occurs when the logic behind purchases changes (e.g., customers stop buying office attire due to the shift to remote work), while distribution shift occurs when the structure of sales changes (e.g., a sharp rise in the share of online channels).
Consider a hypothetical example: an appliance retailer. Before the pandemic, the model predicted high demand for small appliances in November based on 2019 data. However, in 2020, the pandemic shifted demand toward home-office equipment, and physical stores were closed in certain regions. If you don't introduce new dimensions to the model, such as «store operating status» or «delivery availability,» the forecast will be inaccurate. It is essential to implement scenario-based forecasting: a base case, an optimistic case, and an adverse case (accounting for potential supply chain constraints).
Coordination and Infrastructure
Regular model retraining must be supported by Data Governance checks. With the support of companies specializing in infrastructure and integration—such as Data Management IG, IQusion IT, Softline IT, or SL Global Service—enterprises can establish a robust architecture for data collection and cleansing. Remember that any automation is useless without an understanding of the business context. Develop procedures where sales department experts review model results before they serve as a basis for procurement or budgeting. This helps avoid automated errors caused by brief, anomalous spikes or dips in sales.
Author