A study on the prediction of user participation and game outcome of Wordle guessing game based on Arima time series and BERT neural network model

Abstract

Wordle is a popular daily puzzle game in which players try to guess a five-letter word six times or less. This report develops predictive models for the distribution of reported outcomes and the number of reported outcomes reported. The data were first preprocessed, and then an ARIMA(1,1) prediction model was developed by ADF testing and plotting ACF and PACF plots, and calculated that there would be 7,260 reported outcomes on March 1, 2023. In addition, two attributes were extracted, namely the number of repeated letters and the information entropy. Through Pearson correlation analysis, the report investigates the relationship between the attributes and the variance of the scores. It can be seen that entropy is significantly negatively correlated with variance and the number of repeated letters is significantly positively correlated. The report uses the One-hot method to encode words, applies the BERT neural network architecture, and uses the SGD algorithm for optimization to predict the distribution of the reported results. For the word EERIE, the calculated percentages were 1.16%, 5.75%, 20.68%, 32.60%, 24.72%, 12.39%, and 2.70%. The calculated MSE errors indicate the high confidence level of the model. The study provides insight into user behavior and improves game design.