A report covers the quarter ending March 31 but appears publicly in May. A backtest improves after its revenue is entered under the March date. Did the researcher understand history better, or let a March investor read a May report?
Data can describe an earlier period while becoming known later. Retaining a number and one date can turn “occurred then” into “usable then.” Before finding better parameters, give five records the clocks needed to distinguish those claims.
What you will be able to check
- Distinguish the observation period, first publication, personal receipt, and completed processing.
- Locate the earliest permitted decision under given rules, separating usable information from executable trades.
- Handle revisions, delays, and missing timestamps in a versioned evidence table.
The essential skill is respecting historical information boundaries. Tools can manage storage, but the evidence supporting when a version was usable must remain inspectable.
One datum needs several clocks
The reporting period describes what information covers. Quarterly revenue concerns specified business activity; publication records when the information became public. Daily sales do not mean the public received the final quarterly total each day.
Your own process may receive it later. A company releases at 8:00, a service transmits at 8:05, your program receives at 8:07, and checking units and calculating finishes at 8:10. A reconstruction of your executable process cannot use 8:00 as though every later step were already complete.
Processing completion here means the checks, transformations, and calculations required by the rule have finished. It is not a time invented after observing the outcome. Public availability can be studied separately; a claim about your executable workflow must include receipt and processing too.
CFA Institute's backtesting overview discusses approximating actual investing and avoiding look-ahead bias. [1] The violation is using information unavailable to the specified decision, even if the gap is only a few minutes. A date that looks close enough is not a timing policy.
Define the decision clock first
This independent C4 timing exercise does not change the earlier short-account ledger or C2's month-end trend specification. Announcements, data, delays, and execution conditions are synthetic, not actual security events or historical performance.
Assume May 6, 7, 8, and 9, 2024, are applicable normal trading days with no halts or outages. Unless explicitly labeled UTC, all times are local U.S. Eastern Time, or ET. For these dates the exercise supplies the conversion ET equals UTC minus four hours.
The rule checks only once daily, at 08:45. It uses data received and fully processed strictly before that time, produces the day's decision, and plans attempted execution at 09:30. Data completed exactly at 08:45 or later wait until the next applicable trading day. This explicit boundary avoids guessing the sequence of events with indistinguishable timestamps. No midday or closing check is added.
NYSE distinguishes core and extended sessions, with the core session ordinarily beginning at 09:30. Actual markets, instruments, and dates require their own checks. [2] The supplied 09:30 tests ordering here; it does not guarantee an opening fill at a particular price.
Which records arrive in time?
Completion in the table already includes required processing. Unknown means evidence is missing, rather than permission to select a convenient time. First determine eligibility, without evaluating returns.
| Record | Observation described | First publication | Receipt | Processing complete |
|---|---|---|---|---|
| A: original quarter | Revenue through March 31 | May 6, 08:00 | May 6, 08:07 | May 6, 08:10 |
| B: initial macro value | An April economic observation | May 6, 12:05 | May 6, 12:12 | May 6, 12:15 |
| C: company announcement | A May 6 operating event | May 6, 16:05 | May 6, 16:07 | May 6, 16:12 |
| D: correction to A | Revised quarterly revenue | May 7, 08:30 | May 7, 08:37 | May 7, 08:40 |
| E: daily close field | Labeled only May 6 | Not saved | Not saved | Not saved |
A qualifies for May 6's 08:45 check and that day's 09:30 attempt. Its March reporting period does not move eligibility back to March. Publication at 08:00 does not establish that the program had completed its work then.
B finishes at midday and C after the close, both missing May 6's check. Their earliest decision is May 7 at 08:45, followed by the 09:30 attempt. Public release during the day cannot add a midday trade to a rule that checks only once in the morning.
D finishes at 08:40 on May 7, before that day's check. It is a newly usable version from then, not a correction to what May 6's process knew. Keep both A and D to distinguish the original judgment from its later revision.
E is especially vulnerable to an automatic assumption. Its date labels an observation day, without documenting when the final field arrived. A close must at least form before its final value is available, but missing receipt and processing evidence prevents locating a confirmed checkpoint. “Surely available tomorrow” is still an assumption, not a recorded fact.
A revision cannot supply yesterday's answer
Give A a current-quarter revenue value of 100, with the prior year's comparable quarter also 100, in millions of dollars. Its reported growth is 0%. D revises current revenue to 115 while the prior value remains 100, producing 15% growth. These values and version dates are exercise inputs.
If the signal condition is known year-over-year growth strictly above 10%, A does not trigger it on May 6, while D does on May 7. Backfilling 115 under March 31 or the original May 6 timestamp creates a signal absent from the then-available version.
Revisions are not forbidden. Today's analysis can use the current version; past-decision reconstruction needs the version available then. ALFRED preserves historical vintages so users can retrieve what a source reported on past dates. [3] Those are different questions, with different appropriate inputs.
A vintage label does not fully reconstruct intraday availability. ALFRED describes tracking releases and revisions into its database, with source release information distinct from database availability. Minute-level tasks need minute-level evidence; a daily archive cannot substitute for your reception log.
A time-zone error can shift the decision day
Preserve the original time with its zone, then convert to one comparison clock. Eastern, UTC, and Beijing 08:30 refer to different instants. Eastern daylight saving also prevents using one fixed offset all year.
Here, UTC 12:35 becomes ET 08:35 under the supplied four-hour conversion. It does not become 12:35 ET. Then check receipt and processing too. Being publicly released before 08:45 does not establish readiness for that decision.
A trading calendar is more than Monday through Friday. Holidays, early closes, halts, and product-specific windows can postpone an attempt. Establish when the information was usable, then whether the rule permits the execution window. Even both checks together do not guarantee quantity or price.
Can every missing timestamp be delayed one day?
A one-day lag can be a disclosed conservative research convention without being historical fact. An actual three-day delay would still make that convention premature. Different sources' delays also cannot be established through one average.
Verified delay bounds can support explicit scenarios. Without historical versions or receipt evidence, narrow the claim to a mechanism exercise with a fixed-delay assumption, rather than a complete reconstruction of a historically executable strategy. The convention should appear beside the result it limits.
Missing, backfilled, and genuine zero values differ too. Undisclosed revenue cannot become zero and manufacture a collapse. A disconnected program missing a quote does not establish an unchanged market price. Skipping a decision or retaining a position may be appropriate only under a policy specified in advance, with a recorded reason.
For conflicting sources, compare entity, period, unit, version, and publication scope. A news flash may contain an initial value while a later database holds a revision. Greater numerical precision does not determine which belongs in a historical decision. Retain unexplained conflicts instead of generating a plausible release time with AI.
Connect the clock to the next order
C2-A's teaching model forms signals after month-end closing data and executes at the next normal trading day's open. This exercise instead checks at 08:45 daily. Both need required information and calculation to precede the permitted execution window. Neither can backdate an ordinary trade to an already completed close.
No real backtest runs here, and the five records receive no invented returns. We locate data within decisions. Historical universes, corporate actions, expenses, and execution conditions still need checks before evaluating a strategy.
Reorder the records yourself
Explain why A's March 31 period-end is not its decision time. Diagnose a program that reads B at 13:00 on May 6 and rewrites that day's 08:45 decision. Finally, explain why E's May 6 label does not establish timely receipt. Write before checking the answer.
The period describes coverage, not availability. The program violates its fixed morning checkpoint and changes a past decision. A midday strategy could be studied separately if defined in advance, not inserted after seeing a favorable outcome. E lacks release, receipt, and processing evidence; the date label cannot perform those fields' jobs.
Change a delay by a few minutes
A new independent message is released on May 8, 2024, at UTC 12:35. You receive it at ET 08:40 and finish processing at 08:50. Keep the other rules and calendar. Can it enter that day's 08:45 decision? What is the earliest decision and attempt? Can D's revised 115 replace A's 100 in May 6's record?
Release converts to ET 08:35. Although release and receipt precede the check, completion is five minutes late. The earliest decision is May 9 at 08:45, followed by the 09:30 attempt. Do not round processing backward because the delay seems small.
Nor can 115 enter May 6: its correction becomes usable May 7. If the original version is unavailable, acknowledge the inability to reproduce the actual historical input. A newer value cannot masquerade as the missing original.
Save a versioned timing-evidence table
For all five records, enter object identity, metric, observation period, value unit, original source, publication, receipt, processing completion, unified time zone, version and revision relationships, decision window, and earliest permitted execution attempt.
Give each time its evidence: original announcement, reception log, or explicit teaching assumption. Do not copy a file date into all three clocks. Add missing-data and conflict policies, preserving earlier versions and updates. A changed delay should update the table before deciding whether research must be rerun.
Cover the values and inspect only the time chain. Does evidence show this version was usable before the specified decision? Where a field is a guess, narrow the conclusion instead of running the model first and finding a favorable timing story later.
Dates must tell you when it was known
Separate observation, publication, receipt, and processing, then apply your decision frequency and trading calendar. Public information need not have arrived in your process; received information need not be immediately executable under your rule.
A revision can improve current understanding without rewriting past inputs. Mark assumptions separately from historical evidence when times or versions are missing. An incomplete but honest record prevents a tidy historical table from silently acquiring future information.
References
- CFA Institute: Backtesting & Simulation, 2026 public overview. Realistic process and look-ahead bias; no paid course reviewed.
- NYSE: Holidays & Trading Hours. Sessions, holidays, and early closes. The exercise supplies its 2024 trading conditions; today's calendar does not establish them historically.
- Federal Reserve Bank of St. Louis: ALFRED Help. Vintages, release information, and database collection, without minute-level availability claims.
Sources checked October 30, 2026. AI-assisted draft; independent AI full-text and bilingual review is complete. Human financial review and target-reader testing have not been conducted and are not scheduled in this phase. The records and time calculations are synthetic, not actual announcements or strategy performance.