A price bar looks like a settled fact: open, high, low and close. While its interval is running, three of those values can change. The boundary between observation and confirmation matters more than the appearance of the chart.
A result is useful only if the information behind it could actually have been known at the time.
A five-minute example
Imagine a bar covering 10:00 to 10:05. At 10:03 its high may have crossed a reference level. Its eventual close is unavailable. A rule requiring a close beyond the level cannot honestly be confirmed at 10:03 using the value eventually recorded at 10:05.
This does not mean every strategy must wait for a close. An intrabar rule needs intrabar evidence and an explicit specification. It cannot be tested using only the final OHLC bar and an assumed path through it.
The hidden future in a swing label
Some swing definitions require bars on both sides of a turning point. That creates two times: the price extreme and the time the definition becomes confirmable.
If a study places a label at the extreme and behaves as though it was available immediately, it can leak future information. A careful implementation records both timestamps and exposes the label only after confirmation. This is an engineering example, not a claim about a particular product.
Revisions need a history
Data can arrive late or be corrected. Quietly replacing a bar makes a historical explanation difficult to reproduce: today's chart may differ from the data behind the original decision.
A useful record keeps the original value, revision time and affected observation's identity. Re-running a study against corrected data should be distinguishable from replaying the original record.
Three different tests
Ask whether costs are represented, whether intervals are missing and whether the same period was repeatedly used to tune parameters. Ask what determines order timing and price. Touching a level does not establish a fill at that level.
- Rule correctness: Did the code apply the stated conditions?
- Historical evaluation: How did the rules behave under the study's assumptions?
- Operational readiness: Can those conditions be maintained with current data and constraints?
Success in one category does not establish success in the other two.
Keep an uncertainty state
“Waiting for close,” “stale data” and “insufficient history” are useful outputs. Hiding them behind a score simplifies the interface at the expense of clarity.
NIST's AI Risk Management Framework identifies validity and reliability as characteristics of trustworthy systems. Applied to a market study, that means testing when an output should be withheld as carefully as when it should appear.
Further reading
The candle examples are original engineering illustrations. They do not establish investment performance.