Does intrinsic time advance every second?

No. It advances according to the defined events.

Does a local maximum trigger an immediate downward event?

No. The downward reversal must reach the chosen threshold before confirmation.

Is one threshold correct for all datasets?

No. The threshold defines the scale of the representation and requires justification for the task.

AI Glossary / Data science and analytics / Advanced

Directional-change intrinsic time

Definition

Directional-change intrinsic time is an event-based way to represent a time series in which the clock advances when an observed value reverses from a running extreme by a specified threshold. It has been studied particularly for financial price series. Instead of recording one observation every fixed number of seconds, the representation focuses on alternating upward and downward changes of the chosen size.

What it means

A wall clock moves at a constant rate regardless of activity. An intrinsic event clock can record many events during a rapidly changing period and few during a quiet period. The threshold defines what counts as a meaningful reversal for the representation. Smaller movements still occur in the underlying data, but they do not necessarily trigger a new directional-change event. This is a way to organize observations, not a claim that the threshold identifies an economically important trend in every setting.

How it works

The procedure tracks an established direction and the relevant running maximum or minimum. During an upward movement, it updates the maximum and waits for a sufficient decline from that maximum to confirm a downward directional change. During a downward movement, it updates the minimum and waits for a sufficient rise to confirm an upward change. Movement continuing beyond a confirmed change toward a new extreme is commonly discussed as an overshoot. Specify initialization, threshold units, equality rules, observation frequency, and how gaps are handled before implementing the event clock.

Three practical examples

Market-data research: an analyst compares activity using directional events alongside ordinary minute-by-minute summaries.
Simulation research: an agent-based model uses threshold-defined changes as part of an event-based timing framework.
Monitoring prototype: a team experiments with an analogous reversal-based summary of a changing numerical signal, validating whether it preserves information relevant to the monitoring purpose.

Think of it this way

Think of a walking diary that records when you reverse direction after moving far enough, rather than making an entry every minute. A busy walk can create many entries quickly, while standing still creates none. The analogy captures event timing. Real series require exact numerical rules, and a confirmed reversal does not reveal in advance where the previous turning point will occur.

Worked example

Use a fictional positive-valued series and a relative reversal threshold of 10%. Assume the procedure is already in an upward state with a running maximum of 100. Values of 106 and 110 update that maximum to 110. A later value of 104 is only about 5.45% below 110, so it does not confirm a downward event. A value of 99 is exactly 10% below 110 and confirms one under an inclusive threshold rule.
Now track the downward state. If the value falls to 95, that becomes the running minimum. A subsequent value of 104 is about 9.47% above 95 and does not confirm an upward event. A value of 104.5 is 10% above 95 and does. The two confirmation events can be separated by very different amounts of clock time. Also, a 10% fall followed by a 10% rise does not return to the same value, because the percentage bases differ. These invented observations explain the rule and are not a trading signal.

Illustrative case study

In a fictional research team, analysts compare periods that look similarly volatile on a daily chart but have different intraday reversal patterns. They create directional-change summaries at several documented thresholds and keep the original timestamps so event counts can be related to elapsed time. They inspect missing observations and abrupt gaps rather than treating every detected event as equivalent evidence. Before using the representation in a predictive experiment, they ensure that turning-point labels do not introduce future information into features. The study illustrates a research workflow; it makes no claim about profitable decisions or superior forecasting.

Common misconceptions and mistakes

Intrinsic time does not replace physical timestamps for every purpose. A reversal is confirmed only after the threshold is reached, so the earlier local extreme is recognized retrospectively. Ignoring that delay can create look-ahead bias. A percentage threshold and an absolute-value threshold produce different event definitions. Event counts depend on the threshold and the observed path, including data resolution. A pattern reported in one market dataset is not a universal law, and filtering small fluctuations can remove information as well as reduce detail.

Business use

The practical value is in asking whether an event-based representation describes activity more usefully for a specific analytical task. Preserve physical timestamps, threshold definitions, and the underlying data version. Compare results across reasonable thresholds and check whether apparent findings persist outside the development sample. An organization evaluating such a method should distinguish a descriptive activity measure from a prediction model or operational decision rule. Any downstream use needs its own validation, costs, constraints, and clear account of what is known at each time.

Check your understanding

A series reaches 110 and then falls to 104 under a 10% relative reversal threshold. Is a downward event confirmed?

Answer: No. The decline is about 5.45% from the running maximum. Under the stated inclusive rule, 99 would reach the 10% threshold.

Related terms

Time series
Threshold (for decision trees)
Event time
directional change
running maximum
running minimum
overshoot
threshold
look-ahead bias
volatility

Learn next

Practice tracing running maxima and minima through a short series, keeping both event times and clock times. Then compare absolute and relative thresholds and study the distinction between an extreme's timestamp and its confirmation time. Read the original research before adopting method-specific volatility or scaling claims.

Sources

Petrov, Golub, and Olsen, agent-based model abstract: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3240456 . Their intrinsic-time volatility paper is available in the publisher collection: https://mdpi-res.com/bookfiles/book/2890/Computational_Finance.pdf?v=1754442211 . Primary abstract and publisher excerpts were consulted. Direct article access returned a rate limit and the repository copy presented an access challenge; full article review is pending. Examples are original and non-predictive.

Previous: Directional arc consistency Next: Directional component analysis