Recent advancements in time series forecasting have led to the development of foundation models, but they frequently overlook domain-specific features that are crucial for accuracy, particularly in volatile markets such as African retail. Despite the African retail sector's rapid growth, there is a lack of benchmarks and models tailored to its unique conditions. We present Lag-Llama Retail, an adaptation of Lag-Llama, a state-of-the-art foundation model, capable of effectively modelling covariates like promotions and pricing. We pretrain this model on a large-scale, private dataset comprising sales data from four African retailers over two years. Our results demonstrate significant improvements in forecasting accuracy and bias, especially in capturing sales spikes caused by promotions, compared to fine-tuned Lag-Llama, DeepAR and Temporal Fusion Transformer (TFT). This work positions Lag-Llama Retail as a new baseline for time series forecasting in the African retail sector, highlighting the potential of the approach in high-volatility settings and the limitations of foundation models lacking domain-specific covariates.