The Watch Bell Doesn't Always Ring on Time

Opening Scene

The ship’s watch bell rings every half hour, without fail, marking regular intervals a navigator can plan calculations around. But a lookout’s shout of “sail on the horizon” doesn’t wait for the bell — it happens the instant it happens, at whatever irregular moment the sighting occurs. Both belong in the logbook. Both are genuinely time-series data. But a navigator who tries to analyze the lookout’s sightings as if they arrived on the bell’s steady half-hour schedule will draw conclusions the data never actually supports.

Time-series modelling has to know, deliberately, which kind of interval it’s dealing with.

In Plain English

Regular-interval time-series data arrives at fixed, predictable spacing — every minute, every hour, every day — like a sensor polled on a schedule. Irregular-interval time-series data arrives whenever the underlying event happens — a user click, a trade execution, an alert firing — with no fixed spacing at all. The two require genuinely different modelling choices: regular data supports straightforward windowing and gap detection, while irregular data needs event-based thinking and can’t simply assume the next record is “on schedule.”

The Old Way

A great deal of time-series tooling was built with an unspoken assumption baked into it, and that assumption has caused real friction wherever the data didn’t cooperate:

Many time-series tools and models assumed regular intervals by default, because regularly-sampled sensor and metrics data was the dominant early use case, making functions like fixed-window aggregation trivial to reason about.
Applying regular-interval assumptions to genuinely irregular, event-driven data produced quietly wrong results — a missing bell ring meant a genuine gap needing attention, but a missing lookout shout just meant nothing happened, and conflating the two led to false alarms or, worse, real gaps going unnoticed.
Distinguishing “no event occurred” from “an event occurred but wasn’t recorded” has always required irregular-interval data to be modeled with real event semantics, rather than being forced into a regularly-spaced grid it doesn’t actually belong on.

Getting this distinction right early has always saved a lot of downstream confusion about what a missing data point actually means.

What’s Changing (and Why AI Is the Reason)

AI-generated event streams — an agent’s decision points, a model’s inference triggers — are overwhelmingly irregular by nature, arriving exactly when something happens rather than on a fixed schedule, pushing more of the time-series world toward irregular-interval modelling than the historically sensor-dominated field was built for. Tooling and habits built around regular polling increasingly need to accommodate this shift.
AI-assisted anomaly detection, covered in Article 14, needs to know explicitly whether it’s evaluating regular or irregular data, because the definition of “a gap” and “an outlier” differs fundamentally between the two, and treating irregular data as if it were regularly sampled produces a flood of false positives. Correctly labeling interval regularity has become a prerequisite input to reliable automated monitoring.
AI-assisted resampling and interval-classification tools can now automatically detect whether an incoming stream is regular, quasi-regular, or genuinely irregular, and recommend the appropriate modelling approach, reducing what used to be a manual judgment call made once at design time into something continuously verified as real-world data streams evolve.

The Metaphor, Fully Extended

Ship’s Chronometer & Logbook Element	Time-Series Modelling Concept
The watch bell, ringing on a fixed half-hour schedule	Regular-interval time-series data, sampled at fixed spacing
The lookout’s shout, occurring whenever a sighting happens	Irregular-interval, event-driven time-series data
A missing bell ring signaling something is genuinely wrong	A gap in regular data, a real anomaly worth investigating
A quiet hour with no sightings simply meaning nothing happened	The absence of an irregular event, not itself a gap or anomaly
The navigator choosing different calculations for bell-marked versus shouted entries	Modelling regular and irregular time-series with genuinely different techniques

For Beginners: What to Actually Do

Before modelling any time-series source, determine explicitly whether it’s regular or irregular — don’t assume based on how it’s stored.
Practice distinguishing “a missing expected reading” from “an event that simply didn’t occur,” since conflating the two produces wrong conclusions in both directions.
Get comfortable with the idea that irregular data needs event-based tools, not fixed-window ones borrowed from regular-interval workflows.
When in doubt about a source’s regularity, plot the actual time gaps between consecutive records before choosing a modelling approach.

For Practitioners and Leaders: The Deeper Layer

Require interval regularity to be documented explicitly for every time-series data source in your systems, rather than left as an implicit assumption.
Use AI-assisted interval classification to continuously verify that a source you modeled as regular hasn’t quietly drifted into irregularity, or vice versa.
Tune anomaly detection and gap-monitoring systems specifically to each source’s actual regularity, since a one-size-fits-all approach reliably produces either false alarms or missed real gaps.
Recognize that AI-generated event streams are shifting more of your time-series estate toward irregular intervals, and plan tooling investment accordingly.

Quick Recap

Regular-interval time-series data arrives on a fixed schedule; irregular-interval data arrives whenever the underlying event happens.
Applying regular-interval assumptions to irregular data confuses “nothing happened” with “something went wrong,” producing false alarms or missed gaps.
AI-generated event streams are overwhelmingly irregular, pushing more time-series work toward event-based modelling.
AI-assisted interval classification can now continuously verify a source’s actual regularity rather than relying on a one-time manual judgment.

Where This Fits in the Series

This article extends Article 1 and 2’s foundation by addressing how timestamps space themselves out. Article 4 turns to a related foundational problem: timezones and clock drift.