Master's thesis presentation -Martijn Bredewold

Title: Transformer-Based Next-Event-Time Prediction for Banking Transactions

Abstract: 

Retail banking transactions form irregular event sequences in which the timing of past transactions may provide information about future financial events. 

This thesis investigates whether Transformer-based sequence models can predict the number of days until the next incoming transaction within constructed relationship-level transaction sequences. 

Using 134.52 million transactions from 1.57 million customers, three sequence constructions are evaluated, namely customer–counterparty, customer–counterparty–amount, and customer–counterparty–amount–text.
A causal Transformer inspired by SASRec is adapted to continuous next-event-time prediction by modelling the next inter-transaction interval with a Log-Normal distribution. 

The baseline Transformer is compared with Time2Vec and TiSASRec-style temporal extensions, as well as with last-value, window-mean, and window-median recurrence baselines.