Sales Forecasting for Provider Cards Using Seq2Seq LSTM Architecture

Abstract

This study proposes a sales forecasting model for telecommunication card products using a Sequence-to-Sequence Long Short-Term Memory (Seq2Seq LSTM) architecture. The model aims to address the high volatility and nonlinear patterns in internet service provider sales data by learning long-term dependencies and focusing on relevant historical information. The dataset consists of monthly sales data from 2013 to 2024. Data preprocessing includes date formatting, normalization, and dataset splitting into training, validation, and testing groups. The model is trained using the AdamW optimizer, and performance is evaluated using RMSE, MAE, and MAPE metrics. Results indicate that the proposed model achieves RMSE of 47.65, MAE of 37.72, and MAPE of 8.79%, demonstrating strong predictive capability with high accuracy and stability. The forecasting results suggest increasing sales trends for the next three months, showing the model’s effectiveness in capturing realistic market patterns.