An Attention-Based Encoder-Decoder for Week-Ahead Sales Forecasts Applied to a Norwegian Zoo

Abstract

Accurate sales forecasts are vital for companies to plan efficiently and make strategic economic decisions. However, predicting future sales is a complex problem due to the high variance and strong seasonality of sales data. Dyreparken is a zoo and water park in Norway that operates over 30 stores and restaurants throughout the two parks. Accurate automated sales forecasts can enable Dyreparken to more efficiently manage staffing by providing a more reliable decision basis. Such forecasting is challenging among others because there are significant differences between departments with varying activity levels, opening hours, and the uncertainty created by the weather. This thesis investigates how we can apply machine learning-based sales forecasting to amusement parks and zoo’s to forecast hourly turnover for the coming week. By structuring the problem as a multi-output prediction problem, we propose an attention-based encoder-decoder model with gated recurrent units. The proposed model is compared toa range of baseline models. Experimental results show that the suggested model outperforms models such as XGBoost, vanilla artificial neural networks, and neural networks with long short-term memory. Adding weather-related features to the model enables it to predict large deviations in sales caused by poor weather, which can help Dyreparken prevent large deficits in sales generated by unexpectedly low visitor numbers due to rain. In our best-case scenario, our proposed model can predict future turnover with a mean absolute error of 0.135, or 0.452 weighted MAPE.