Predicting Transaction Latency with Deep Learning in Proof-of-Work Blockchains

Abstract

Proof-of-work based cryptocurrencies, like Bitcoin, have a fee market where transactions are included in the blockchain according to a first-price auction for block space. Many attempts have been made to adjust and predict the fee volatility, but even well-formed transactions sometimes experience delays and evictions unless an enormous fee is paid. In this paper, we present a novel machine-learning model, solving a binary classification problem, that can predict transaction fee volatility in the Bitcoin network so that users can optimize their fees expenses and the approval time for their transactions. The model’s output will give a confidence score whether a new incoming transaction will be included in the next mined block. The model is trained on data from a longitudinal study of the Bitcoin blockchain, containing more than 10 million transactions. New features that we generate include information on how many bytes were already occupied by other transactions in the mempool, assuming they are ordered by fee density in each mining pool. The collected dataset allows to generate a model for transaction inclusion pattern prediction in the Bitcoin network, hence telling whether a transaction is well formed or not, according to the previous transactions analyzed. With this, we obtain a prediction score for up to 86%.