← Quantery Blog

How to choose a backtest start date

October 6, 2026 · 8 min readbacktestingresearch-methoddata-coverage

Choose a backtest start date by finding the first period when the test has usable coverage, then asking whether that period gives the thesis a fair chance to operate. Declare the date before looking at returns. After the baseline run, move the start forward by one fixed interval while preserving the thesis, end date, benchmark, and every other setting. A conclusion that disappears after one defensible shift belongs to that slice of history.

The start date is the first simulated date on which the backtest can evaluate your rules. It controls more than the chart's length. It decides which companies can enter, which market episodes receive the most weight, and whether early results come from a broad universe or a thin corner of the database. The earliest date a form accepts isn't automatically the earliest useful date.

Data coverage sets the earliest defensible date

Begin with coverage before thinking about a favored market episode. A fundamentals thesis needs filing data available as of each simulated decision. A price rule needs enough prior prices to compute its lookback. An index thesis needs point-in-time membership, or a clearly labeled fallback. If those inputs are sparse at the front of the window, extending the chart backward adds dates without adding a comparable test.

Quantery's point-in-time SEC fundamentals support backtests from 2013. Price depth depends on the user's FMP plan, so don't turn that product boundary into a blanket claim about every installation. Open the coverage information for the symbols and datasets your thesis uses. The data-coverage preflight explains how to inspect missing filing periods, price gaps, and the effective universe before a run.

A feature can require more history than its raw field. A trailing value needs enough filed periods to fill its window. A year-over-year change needs the comparison period. A price average or drawdown needs prior observations before the first date it can be evaluated. Starting on the first stored row may therefore produce null features and no eligible companies.

The filing clock matters too. Form 10-Q covers interim financial statements after each of the first three fiscal quarters, according to the SEC's Form 10-Q instructions. A fiscal quarter can end before its report becomes public. Point-in-time evaluation waits for the filing; it doesn't backdate the numbers to the period end. Trailing and point-in-time numbers covers that distinction in detail.

Do one diagnostic run over the broad period you think you want, then read the coverage note before reading performance. Look for the first stretch where the effective universe is reasonably populated and stays that way. This isn't permission to hunt for the first profitable date. You're establishing when the intended test actually exists.

A fair window must contain the thesis's mechanism

Coverage gives you a floor. The business claim decides whether the window above that floor is informative.

Suppose the thesis selects companies with strong balance sheets after an industry shock. A window containing only the recovery tells you how the rule behaved during recoveries. It can't show how those companies entered the shock or whether the same gate failed in calmer conditions. Conversely, beginning at the peak of one crisis can let a single exceptional cohort dominate the whole record.

Write the mechanism in one sentence before choosing among valid dates:

Firms with resilient cash generation and manageable leverage tend to recover better after operating pressure becomes visible in filed results.

That sentence suggests the test should include ordinary periods as well as pressure and recovery. It doesn't tell you to start on the day of a famous selloff. Famous dates are memorable precisely because they're unusual.

Match the claim to the sample. If your valid history starts after the episode most relevant to the idea, narrow the claim or gather better data. Don't describe a recent expansion as evidence across a full cycle.

A longer window can still be a worse window. Early sparse years may contribute few positions while making the report look comprehensive. They can also change cross-sectional ranks because the peer set is smaller. If the thesis uses percentile ranks, the universe determines what each rank means. A company can receive a different relative score when half its intended peers are absent.

Declare the baseline before seeing its curve

Pick the baseline from two written facts: the coverage floor and the mechanism you want represented. Round the boundary to a natural research interval, such as a calendar year, unless the thesis itself has an event-based reason for another choice. Record why the date belongs.

Then freeze the rest of the test:

• Keep the active thesis version unchanged.
• Keep the end date and benchmark unchanged.
• Keep the rebalance cadence, reporting lag, and exit policy unchanged.
• Keep the universe definition unchanged.

Those controls make the start-date comparison legible. Changing the start and the exit rule together creates a new backtest, but it doesn't isolate either decision.

The shape below is illustrative, and the bundled templates are the reference for exact fields. The dates are proposed test parameters. They make no claim about an optimal historical window.

# Baseline research window: illustrative
backtest:
  from: 2016-01-01
  to: 2026-01-01
  rebalance: monthly
  reporting_lag_days: 1
  entry: next_close
  exit: { max_hold_months: 12, on_gate_fail: true }
  benchmarks: [SPY]

Save the baseline result before running variants. Your research note should include the coverage reason, the thesis mechanism, and the variants you intend to run. Once you've seen the curve, the temptation to call its best stretch the proper start date is hard to separate from genuine judgment.

Trying enough configurations makes an impressive backtest easier to find by chance. Bailey, Borwein, López de Prado, and Zhu describe how selecting from many tested variants can produce apparent performance that fails on new data in their paper on backtest overfitting. Start dates are parameters. Treating them as cosmetic doesn't exempt them from that search problem.

Shift the start date without moving anything else

A useful sensitivity test needs only a few windows declared in advance. Start with the baseline, then move the boundary forward by one fixed interval and again by the same interval if coverage permits. Preserve the common end date. The later runs contain less history by design, so don't rank them by total return.

Compare the parts that answer the same question:

• Did excess performance keep the same direction against the same benchmark?
• Did the median qualification event tell the same broad story as the mean?
• Did maximum drawdown change enough to alter the practical claim?
• Did the effective universe remain comparable across the shared years?
• Did one early cohort account for most of the baseline's advantage?

The last question needs care. Quantery's report provides an event study and a naive portfolio, but it doesn't provide full cohort attribution for every slice of the equity curve. Use fixed subperiod runs and inspect the go-event dates. Don't claim exact cohort contribution when the report doesn't calculate it.

A later start naturally removes early events. The sensitivity test asks whether the thesis still points the same way after that history is removed. The shorter run has fewer observations, so a noisier estimate is expected. Read direction and dependence before demanding identical magnitudes.

Keep the end date fixed for this pass. Moving both boundaries creates rolling windows, which answer whether the thesis behaved similarly across different periods. That's a valuable second test, covered in the guide to market-regime tests, but it isn't a start-date isolation test.

Don't optimize the window around what you know happened

There are three common ways to turn a date choice into result shopping.

First, start immediately after a large loss the thesis would have suffered. The explanation usually sounds plausible after the fact: the data were unusual, the rules weren't meant for that environment, the market structure changed. If that restriction wasn't in the original mechanism, it's an exclusion written by the curve.

Second, start just before the thesis's strongest cohort. That gives the favorable episode maximum influence and may leave little ordinary time before it. A calendar boundary chosen before the run is easier to defend than a date anchored to the first winning signal.

Third, extend backward until the result crosses a desired threshold. Weak early coverage can make this especially misleading because the added dates may contribute a small and unrepresentative set of companies. The chart gets longer while the evidence barely broadens.

If a structural break genuinely changes what the thesis means, state the break using information outside the backtest result and test the pre-break and post-break periods separately. A new accounting rule, a changed universe definition, or a missing historical field can justify a boundary. The resulting claim is conditional on that boundary. Say so.

Read instability as information

If nearby start dates preserve the broad conclusion, the baseline is less likely to owe everything to its opening stretch. It still needs other checks: threshold perturbations, alternate benchmarks, sector controls, and a holdout period. A first backtest is only a baseline, and every hypothetical result excludes trading costs.

If the conclusion flips, don't average the runs and call the problem solved. Find the dependence. It may be one market episode, one industry cohort, sparse early coverage, or a rule whose economics changed. Ask a narrower question next. Selecting another date to restore the curve teaches you nothing.

Quantery keeps completed runs tied to their thesis versions and settings, so leave the awkward windows in the record. Write what changed and what didn't. The rules are yours, and the window is one of them. Make the date earn its place before the return gets a vote.

Want to try this on your own rules? Quantery is free for 14 days: the full app, no card required.

← All articles