Start and End Date Sensitivity
How much a backtest's headline numbers change just from shifting the start or end date by a few months — a quick way to tell a robust strategy from one whose performance depends on a lucky window.
Every backtest has to start and end somewhere, and that choice is far more consequential than it looks. A strategy tested from January 2015 to December 2019 and a strategy tested from January 2016 to June 2020 are answering two different questions, even if the underlying logic is identical — one window includes a crash, the other doesn't; one starts mid-rally, the other starts at a trough. Start and end date sensitivity is the practice of checking how much a backtest's results move when those boundaries shift, because a strategy whose Sharpe ratio swings wildly with a two-month change in window is telling you something important about how much to trust it.
Why the boundaries matter so much
The first and last few months of a backtest window carry outsized influence on compounded return metrics, especially for a short backtest: a strategy that happens to end its test right after a big winning trade, or right before a losing one, can show a very different total return depending on where the window is cut. There's also a subtler problem — researchers, even unintentionally, sometimes settle on a window after looking at results, nudging the end date a month later because it made the Sharpe ratio tick up. That's a form of the same overfitting that plagues parameter tuning, except the parameter being tuned is the test's own boundaries.
Worked example: a strategy that isn't as good as its favorite window
A strategy backtested from March 2016 to March 2020 shows an 11% annualized return. Shifting the end date forward six months to include the COVID crash (September 2020) drops it to 4%. Shifting the start date back a year, to March 2015, drops it further to 2%, because that earlier year happened to be a losing stretch for the strategy's core signal. A researcher who only reports the original March 2016–March 2020 window, without disclosing how sensitive the result is to that specific choice, is presenting a best-case number as if it were representative. Running the same strategy across a grid of, say, twelve different start dates a month apart and twelve different end dates gives 144 Sharpe ratios instead of one — and the spread of those 144 numbers is far more informative than any single one of them.
What this means in practice
Reporting date-sensitivity alongside a backtest's headline numbers — a range or a distribution rather than one Sharpe ratio — is one of the cheapest robustness checks available, since it requires no new data or model, just re-running the existing backtest on shifted windows. Strategies whose performance is stable across nearby windows are more likely to reflect a genuine, persistent effect; strategies that only look good in one narrow window are prime candidates for having been (even unconsciously) fit to that window's specific history.
Rerunning a backtest across a range of nearby start and end dates turns one Sharpe ratio into a distribution, and a wide, unstable distribution is a warning sign that the reported result depends on a lucky choice of window rather than a robust effect.
Silently adjusting a backtest's start or end date after seeing preliminary results — even with a plausible-sounding reason like "avoiding the financial crisis" — is a subtle form of overfitting to history. Any deliberate window choice should be justified before looking at performance, not after.
Further reading
- López de Prado, Advances in Financial Machine Learning, ch. 11