Predicting Block Delays in Bitcoin Blockchain via Bayesian and Non-Bayesian Logistic Regression

Blockchain has found many uses beyond cryptocurrency trading. However, this technology faces the challenge of low transactions per second (TPS), which hinders it from competing at scale with other established digital payment technologies. This low TPS problem is worsened by occasional block generation delays, which usually hike transaction fees as transactors must compete using transaction fees to have their transactions added to a block in a blockchain. These block delays can also be indicative of other underlying issues, such as dishonest mining. This study focuses on the prediction of such delays in a Bitcoin blockchain. Such predictions can help transactors strategically place transactions at moments when delays are unlikely to occur, and the predictions can also help avoid resulting higher transaction fees and potential dishonest miner behavior. The problem is modeled as a prediction of a binary outcome, that is, block delay or block-on-time. Three approaches for training the logistic regression model were compared: frequentist maximum likelihood estimation (MLE), Bayesian Hamiltonian Monte Carlo (HMC), and Bayesian Gibbs sampler. The precision-recall (PR) area under the curve (AUC) was chosen as the main comparison metric because the dataset exhibited a class imbalance problem. The PR AUC values showed that the fitted models performed better than the random classifier, with the Gibbs sampler model performing the best among the three. The PR AUC values also showed that there was a big room for improvement on all the fitted models.