Feedzai drops causal discovery's regular-lag assumption

Temporal causal discovery tools assume a regular clock, but real event data rarely obliges, per a new paper from Feedzai's Penim, Pereira and Bono.

That's a problem for root-cause analysis on logs, telemetry, or user events. Production systems don't emit data on a fixed cadence. Standard causal discovery methods quietly assume otherwise.

In arXiv 2607.18226, the team's method operates directly on irregularly sampled event series, skipping the step of forcing data onto a regular time grid. They point to Feedzai's own transaction data as the kind of stream that breaks the old assumption: bursty, uneven, never on a clean interval.

The paper doesn't say how the method compares to lag-based ones on evenly spaced data. So it's not clear whether this replaces the standard toolkit or just handles the messy cases it can't.

If your incident tooling assumes evenly spaced events, its root-cause rankings are already wrong the moment traffic gets bursty.