How should a backtest handle 53-week fiscal years? | Quantery Blog
A backtest should treat a 53-week fiscal year as a longer measurement period. Keep the filed number, flag the unusual period length, compare it with care, and rerun the test without affected observations. Don't let an extra week masquerade as revenue growth or cash-flow improvement.
This matters because plenty of companies close their books on a weekday instead of a fixed calendar date. Apple's fiscal year, for example, ends on the last Saturday of September and may contain either 52 or 53 weeks, according to its fiscal 2023 Form 10-K. The filing is correct. A naive comparison can still be wrong.
Why does an extra week distort a stock screen?
Revenue, expenses, operating cash flow, and capital expenditures are flow values. They accumulate over a period. Give a business another week to sell products and collect cash, and the annual total gets another week of activity. Balance-sheet values such as cash and debt are point-in-time values. They describe the closing date, so the longer reporting year doesn't add to them by itself.
That distinction creates a narrow but real comparability problem. Apple's fiscal 2023 contained 53 weeks while the two preceding fiscal years contained 52. The company reported total net sales of $383.285 billion for fiscal 2023, down 3% from $394.328 billion in fiscal 2022, and said the decline was partially offset by the additional week of sales in fiscal 2023 in its annual filing. The reported decline is real. It just combines business performance with a calendar benefit.
The comparison reverses the next year. Apple's fiscal 2024 Form 10-K says fiscal 2024 spanned 52 weeks, fiscal 2023 spanned 53, and the extra week in fiscal 2023 sat in the first quarter, which ran for 14 weeks. The same filing reports that total net sales rose 2%, or $7.8 billion, in fiscal 2024. That growth cleared the burden of comparing a shorter year with a longer one.
A screen that reads only the annual growth rate can't separate those effects. It may reward the extra week when it enters the comparison, then penalize its absence a year later. No accounting error occurred. The research rule misunderstood the denominator of time.
Why doesn't TTM automatically fix the problem?
Trailing twelve months, or TTM, usually means adding the latest four reported quarters. It is a strong default because the window contains each seasonal quarter once. Our guide to trailing versus forward numbers explains why this beats annualizing the latest quarter.
But “four reported quarters” and “a constant number of weeks” can diverge. If one quarter contains an extra week, a four-quarter sum that includes it covers more trading activity than the next four-quarter sum after it drops out. The TTM series can show a step down even when the business's weekly pace is unchanged.
The same issue reaches ratios in different ways:
- Margins usually hold up better. Revenue and operating income cover the same period, so dividing one by the other cancels much of the duration effect. Mix a flow numerator with a point-in-time denominator, though, and the extra activity remains.
- Turnover ratios need matching periods. Annual revenue divided by ending inventory will be a little more generous in a longer year even if inventory discipline hasn't changed.
- Cash conversion can move for timing reasons. An extra payroll, supplier-payment, or collection cycle may land inside the period. Dividing annual operating cash flow by annual net income doesn't guarantee that those timing effects cancel.
- Growth rates take the direct hit. A longer current period against a shorter prior period flatters growth. A shorter current period against a longer prior period suppresses it.
Weekly normalization looks like an easy repair: divide each annual flow by the weeks reported, then compare weekly rates. It works as a diagnostic. The filing remains the record. Retail weeks differ in economic importance, product launches bunch activity, and expense recognition doesn't arrive in neat weekly slices. The extra week may sit in the holiday quarter, as Apple's did, so treating every week as interchangeable can create a tidier number without creating a better one.
What should the backtest do with fiscal calendars?
Start with the period dates attached to each filing. The backtest should know how much time each flow covers before it computes growth. Fiscal labels alone aren't enough. “Fiscal 2023” describes the company's reporting year; it doesn't promise the same duration as fiscal 2022 or a peer's fiscal 2023.
Then use a layered policy.
First, preserve the filed values. The raw record should remain exactly what the company reported. Calendar adjustment belongs in a derived feature, where it can be inspected and changed.
Second, flag unusual duration. A period-length flag lets the screen distinguish ordinary year-over-year growth from a comparison involving an extra week. The flag should travel with the result even if the company passes.
Third, compute an adjusted companion measure. Put reported growth beside a per-week growth diagnostic. If both point in the same direction, the calendar probably isn't deciding the signal. If they disagree, the observation needs a filing read before it earns a score.
Fourth, test exclusion. Rerun the backtest after removing growth observations whose current or comparison period has unusual duration. This is the cleanest answer to the research question. If the result disappears, the rule may have learned fiscal calendars.
Fifth, keep availability point in time. Period length is known from the filed document. A simulation may use it only after that filing became public. Inferring it later and patching old observations with knowledge unavailable at the time would reintroduce look-ahead. The article on reporting lag in backtests covers the filing-date discipline behind that rule.
Can you express the calendar check in a thesis?
Not entirely in the current Quantery thesis language. The DSL can read financial-statement fields with period functions such as latest, ttm, sum, cagr, and lag, but it doesn't expose a filing period's start and end dates to an expression. It therefore can't calculate period duration or identify a 53-week year on its own. The bundled templates and the Quantery DSL documentation remain the reference for supported fields and exact thesis structure.
Do the calendar classification outside the thesis expression by reading the filed period dates and the company's disclosure. Then compare separate research variants: the reported growth rule, the same rule with affected observations excluded, and a duration-adjusted calculation prepared outside the DSL. Keep the adjustment visible. Don't bury it inside a vendor growth field or present external preprocessing as if the thesis language performed it.
The paired growth measures still do more work than a single adjusted value. They show whether the conclusion depends on normalization. That makes the assumption testable, even though the current DSL can't perform the duration calculation itself. You can compare agreement, flag disagreement in your research notes, or require ordinary duration for a specific experiment.
Don't silently discard every longer period. That can select against retailers and other companies using week-based calendars, changing the tested universe for a reason unrelated to the thesis. Track how many observations leave each rebalance and whether exclusions cluster by industry. A cleaner comparison isn't useful if it changes what kind of company gets compared without telling you.
What would a convincing robustness test show?
Run the original rule on reported values. Run it again with the duration flag excluded. Then run the paired-measure version that requires reported and adjusted growth to agree. Backtests are hypothetical and exclude costs, so these runs are research variants. They aren't forecasts.
Look past the headline return. Check whether the same entry cohorts drive the result, whether the affected observations concentrate in one industry, and whether the rule changes character when those observations leave. A backtest result needs robustness checks, and fiscal-duration sensitivity is one more perturbation with a clear economic reason behind it.
The test can establish that your result doesn't rely on a recurring calendar quirk. It can't prove that per-week normalization captures the underlying economics, or that every filer describes unusual periods consistently. For any important disagreement, open the filing and read the company's explanation.
A 53-week year isn't bad data. It's the company's real reporting period. The mistake is pretending every annual column covers the same amount of time. Keep the filed number, expose the period length, and make the rule survive both versions. Then the growth signal belongs to the business, not the calendar.
Research tooling, not investment advice. Nothing here is a recommendation to buy, sell, or hold any security. Screens, scores, and backtests are informational only; backtested results are hypothetical, exclude costs such as commissions and slippage, and do not guarantee future results. Verify against primary filings and make your own decisions.
Want to try this on your own rules? Quantery is free for 14 days: the full app, no card required.
← All articles