The Goal
Food prices change over time, making future costs difficult to anticipate. We explored whether historical supermarket data could forecast average fruit, vegetable, and meat prices in Italy over six months. We treated the forecasts as a potential aid for planning, not as a tool already used in practice.
What I Built
Working as a two-person team, we aggregated product prices into daily averages and developed forecasts using a simple baseline, statistical methods, and machine learning models. For the machine learning models, we created features from previous prices and calendar dates. We evaluated predictions against a 180-day test period. The models selected in our notebook achieved MAPE below our 5% target across all three categories.
Process and Reflection
This was our first time working with time series data, so we first had to learn how to explore trends and seasonal patterns while preserving the order of observations. The results taught us to look beyond a single accuracy score: despite its low error, the vegetable model’s longer-range forecast became nearly flat. If we continued the project, we would test other time intervals, evaluate across multiple periods, and consider external factors that affect prices.