Evaluating LSTM Performance on Multivariate Time Series with One-Class SVM Outlier Detection

Weekly sales forecasting plays a crucial role in retail business planning and inventory management.This study evaluates the prediction performance of a Long Short-Term Memory (LSTM) model for weekly sales forecasting after data preprocessing using standardization and outlier detection with One-Class Support Vector Machine (OCSVM) method. The independent variables used include temperature, fuel price, holidays, Consumer Price Index (CPI), and unemployment rate, with weekly sales as the target variable. The dataset is preprocessed using StandardScaler and OCSVM to detect and remove outliers before model training. The evaluation shows that the LSTM model on the clean data achieves an MSE of 0.03, an RMSE of 0.18, and an MAE of 0.11. The LSTM model demonstrates good forecasting performance when trained on cleaned data without outliers. This study provides practical insights into applying data preprocessing with OCSVM to improve the consistency of prediction models in retail time series analysis.