<![CDATA[This research aims to develop and optimize a daily sales prediction model based on time series data using the Light Gradient Boosting Machine (LightGBM) algorithm on online retail data from the Olist marketplace. The research process began with merging and aggregating e-commerce transaction data into a daily format, followed by outlier handling using the Interquartile Range (IQR) capping method, and feature engineering to add temporal and historical information such as prev_day_sales and day_of_week. The dataset was then split into training and testing sets using a time-based split approach. A baseline model was trained with default parameters and subsequently optimized through hyperparameter tuning using GridSearchCV with TimeSeriesSplit cross-validation. Evaluation was conducted using MAE, RMSE, and R² metrics. The results show that the tuned model improved prediction accuracy, with MAE reduced by 5.34%, RMSE decreased by 8.34%, and R² increased by 0.76%. The one-day-ahead daily sales prediction reached R$ 1,676.86 and closely followed the actual sales pattern. This study demonstrates that a systematic approach involving data preprocessing, feature engineering, and parameter tuning can produce a more accurate, stable, and practical sales prediction model to support decision-making in the e-commerce sector. Theoretically, this research contributes to strengthening the understanding of the effectiveness of the LightGBM algorithm in daily time series modeling, particularly through the integration of temporal feature engineering and systematic parameter tuning strategies. These findings underscore the importance of a comprehensive approach in building accurate sales prediction models.]]>
