Reconstruction of payment obligation trajectories from event logs of a digital service platform

Abstract

This paper addresses the scientific and applied problem of reconstructing payment-obligation trajectories on digital service platforms from operational logs and partially aggregated records. Raw platform data do not normally form ready-to-use state sequences: information about obligation creation, payment requests, repeated contacts, due dates, payments, and observation termination may be stored in different fields, have different temporal granularity, and may not match the business meaning suggested by technical attribute names. A reconstruction procedure is proposed that includes identifying records belonging to the same payment obligation, semantic validation of time fields, conversion of source records into events, event ordering, reconstruction of states and admissible transitions, handling of aggregated repeated contacts, separation of confirmed terminal outcomes from censoring, and quality control of the resulting trajectory. A key principle is that timestamps are not synthetically generated for events whose occurrence times were not directly observed. The procedure was evaluated on a dataset of 15,566 payment obligations from a digital service platform. For 5,580 records, the number and temporal boundaries of payment requests were available. The sent_count field represented 7,320 requests in total. Exact timestamps could be reconstructed for 6,390 requests (87.3%), whereas 930 intermediate repeated requests remained temporally unresolved. An exact payment time was available for 1,979 obligations. The analysis also showed that created_at cannot be automatically interpreted as the trajectory start: in 5,564 of 5,580 cases with a known first request, created_at was not earlier than min_sent_at. Two cases contained payment timestamps earlier than the last recorded request and were marked as conflicting rather than automatically corrected. The scientific contribution is the formal reconstruction of payment-obligation trajectories from partially aggregated event data while explicitly preserving the degree of temporal certainty and separating censored observations from confirmed completion without payment. The resulting trajectories can serve as a consistent data basis for multi-state modelling, time-to-event analysis, and subsequent dynamic prediction.