Predicting Demand in Food Delivery: A Hybrid Model for Promotional and Post-Promotional Periods

Abstract

Accurate forecasting of daily order volumes is critical for businesses facing rapid fluctuations in demand, particularly during promotional events such as the Subsidy War. This paper proposes a hybrid forecasting model combining log-normal distribution, ARIMA, and
LSTM to predict order volumes during and after the promotional period for a milk tea shop. For the promotional period, we model the demand using a log-normal distribution to capture the skewed and heavy-tailed nature of the order volume. Post-promotion, we use an ARIMA
model to account for trend in the order data, augmented with external temperature data to handle residual errors via an LSTM neural network.
A weighted sum of the ARIMA forecast and a decaying trend function is used to generate the final predictions, ensuring stationarity and
reducing overfitting. The proposed hybrid approach provides a robust solution for managing demand fluctuations, offering valuable insights
into both promotional and non-promotional periods.