Scholarlink

The study examines challenges in forecasting time series with sudden peaks and large fluctuations, common in logistics and supply data of the cosmetics industry. Such irregularities violate key assumptions of classical models, including stationarity and predictable seasonality. Daily delivery data from selected cosmetics suppliers were analyzed using three models: ARIMA, ETS, and TBATS, applied to both raw and preprocessed datasets. Preprocessing included outlier adjustment via decomposition-based cleaning and winsorization. Results showed strong weekly seasonality across all suppliers and, in some cases, additional monthly effects. Data cleaning improved forecast accuracy, reducing the mean absolute error (MAE) by over 50%. ARIMA performed best on cleaned data, while TBATS handled multiple seasonalities most effectively. The findings highlight the importance of preprocessing for reliable forecasting in volatile environments, though excessive smoothing can obscure meaningful irregular patterns.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Time Series Forecasting Under Irregular Seasonality and Anomalies

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Patrycja Guzanek,
  • Mykola Karpenko,
  • Anna Borucka

摘要

The study examines challenges in forecasting time series with sudden peaks and large fluctuations, common in logistics and supply data of the cosmetics industry. Such irregularities violate key assumptions of classical models, including stationarity and predictable seasonality. Daily delivery data from selected cosmetics suppliers were analyzed using three models: ARIMA, ETS, and TBATS, applied to both raw and preprocessed datasets. Preprocessing included outlier adjustment via decomposition-based cleaning and winsorization. Results showed strong weekly seasonality across all suppliers and, in some cases, additional monthly effects. Data cleaning improved forecast accuracy, reducing the mean absolute error (MAE) by over 50%. ARIMA performed best on cleaned data, while TBATS handled multiple seasonalities most effectively. The findings highlight the importance of preprocessing for reliable forecasting in volatile environments, though excessive smoothing can obscure meaningful irregular patterns.