Feature Engineering for Virtual Currency Price Prediction in Online Games Using a Multiple Linear Regression Approach

Abstract

Changes in the price of Growtopia's virtual currency can influence players' decisions when buying, selling, or storing items, while price information circulating in the community is often inconsistent. This condition confirms the need for transparent, data-driven predictions. This study aims to predict the prices of World Lock (WL), Diamond Lock (DL), and Blue Gem Lock (BGL) using multiple linear regression. The novelty of the study lies in the application of three separate models with a combination of lag features and cyclic time features, accompanied by chronological data sharing to avoid information leakage. The initial dataset contained 925 daily observations from 2024–2026, while the modeling was limited to 194 data points from 2026 that had a consistent scale. After the formation of Lag_1, Lag_2, Lag_7, Day_Sin, and Day_Cos, 187 complete lines are available, divided into 150 training data and 37 test data. The models were evaluated using MSE, RMSE, MAPE, and R². All three models yield an R² of 0.9318, a MAPE of 2.3191%, as well as a MAPE-based accuracy of 97.6809%. The RMSE is 0.1400 for WL, 14.0017 for DL, and 1,400.1705 for BGL, respectively. As of July 13, 2026, the actual and predicted differences are 0.1743 WL, 17.4286 DL, and 1,742.8648 BGL, respectively. The 123-day recursive prediction yields a final value of 2.8357 WL, 283.5670 DL, and 28,356.6979 BGL. The results show the model follows a general pattern, but its response weakens when price changes occur suddenly. The findings can be used as companion information to transactions, rather than as price certainty or investment recommendations.