<p>Accurate demand forecasting is essential in the retail sector to ensure customer satisfaction, reduce costs, and promote sustainability. Many predictive variables, such as weather and calendar features, have been incorporated into forecasting models. However, demand is also influenced by market states, such as shifts in consumer preferences or economic changes, which the existing literature has not rigorously considered alongside predictive variables. Furthermore, little is known about the effect of market states on short-term relationships between predictive variables and sales. This study addresses this gap by developing a novel forecasting model that incorporates predictive variables and market states, which are typically long-lasting and hidden. The proposed approach combines a hidden Markov model with a deep neural network, named the Hidden Markov Modulated Deep Neural Network (HMMDNN), to predict sales. We apply the model to grocery sales data from a major Canadian retailer, achieving significant forecast error reductions of 26.4% for daily data and 17.5% for weekly data compared to the best benchmark models. We identify two market states: a <i>high</i> market, characterized by elevated sales and variations, and a <i>low</i> market, characterized by reduced sales and smaller variations, relative to trend and seasonality. We find that market states play a more significant role than predictive variables in sales dynamics. Additionally, state-specific interpretations reveal that in the high market, weather has a more significant impact, whereas in the low market, sales follow calendric recurring patterns and longer-term trends. These findings provide retailers with valuable insights into evolving market states alongside short-term effects.</p>