Demand Analysis of Daily Necessities Based on Machine Learning Technology

In response to the small sample size and irregular time series data of household goods orders, this study proposes a data repair and multi-model collaborative prediction framework based on the BGCP algorithm, focusing on addressing the limitations of traditional statistical methods in terms of data missingness, feature correlation, and adaptability to multiple time granularities. Innovative proposals include: (1) employing the BGCP algorithm to construct a Bayesian tensor decomposition model with shared regional parameters, achieving anomaly correction and missing value imputation (regional average RMSE=0.482); (2) designing the MMIFS multivariate mutual information feature selction mechanism, effectively capturing nonlinear features such as holidays and promotional activities (lag feature mutual information reaches 0.95); (3) establishing a multi-model prediction system with time granularity adaptability, where the CRU model enhances daily granularity prediction accuracy through a continuous time increment module (MSE=0.643×10^7), the DeepAR model optimizes weekly granularity prediction using an autoregressive network (MSE=0.475×10^7), and the Prophet model performs best at monthly granularity prediction due to its trend decomposition capability (MSE=0.310×10^7). Experiments show that compared to traditional methods, the proposed framework reduces prediction errors at daily/weekly/monthly time granularities by 18.3%/23.6%/34.7%, respectively, verifying the adaptability of the multi-model collaborative mechanism to complex time series features. The research results provide a fine-grained demand forecasting method for supply chain management, particularly suitable for short-life-cycle products with promotional fluctuations and seasonal abrupt changes.