Decision Time Versus Data Timestamp

A dataframe row has one time column, so it feels like there is one time. There are at least three, and they can be months apart:

• Event time — when the thing happened. The quarter ended on 31 March.
• Arrival time — when the information reached you. The earnings were released on 12 May at 16:05, and the vendor's row appeared in your database on 13 May at 02:00.
• Decision time — when you are allowed to act on it. Some time strictly after arrival, plus however long your pipeline takes.

The rule the whole page reduces to: a backtest must be indexed by arrival time, and may only trade at a decision time after it. Vendors overwhelmingly index by event time, because that is what an analyst wants to look at. If you join their file to a price series on its native date column, you have built a time machine and you will not notice, because nothing errors.

Worked example: a fundamental signal that was never tradeable

A cross-sectional value strategy on 1,500 US stocks. Rank on earnings yield, long the cheapest quintile, short the most expensive, rebalance quarterly. The fundamentals file has a period_end column, so the join uses it: on 1 April, use the quarter ending 31 March.

Backtest result: 19.4% a year, volatility 11%, Sharpe 1.76, positive in 17 of 20 years. It looks like a career.

Now re-run with the vendor's report_date field instead, and add the two-day settling lag a real pipeline needs. The median US filer reports about 42 days after quarter end; the slow tail runs past 70. So the 31 March number becomes usable around 14 May, not 1 April.

Same rule, same universe, same 20 years: 3.1% a year, volatility 11%, Sharpe 0.28. Net of 30 bps round-trip costs at 160% turnover it is roughly 2.6%, indistinguishable from cash.

Where did 16 points of annual return go? Nowhere — it was never there. Prices move on the earnings release. Knowing the number six weeks before the market did means you were front-running the announcement in every one of the 80 quarters in the sample. The "value edge" was the announcement drift, harvested with a time machine.

Worked example: one minute of the future, at high frequency

The same bug in miniature, where it is even easier to miss. You have 1-minute bars. Your feed labels each bar with the time it started: the bar tagged 09:31:00 covers 09:31:00 to 09:32:00. Your research code assumes the label is the end of the bar, which is the other common convention.

The strategy trades a short-horizon reversal, entering at what the code believes is the bar close and exiting one bar later. Backtest: average +3.9 bps per trade, roughly 26,000 trades a year, hit rate 58%.

Correct the convention — the label is the bar's open, so a decision made "at" 09:31:00 is really made 60 seconds before that data exists — and re-run: +0.6 bps per trade, hit rate 50.4%. The spread on these names is about 1.4 bps, so half a bp of gross edge is a loss before you have paid anyone.

The arithmetic is worth sitting with. A US trading year holds about 5.9 million seconds, so a stock with 25% annualised volatility moves roughly $0.25 \times \sqrt{60 / 5900000} \approx 0.0008$ — about 8 bps — in a typical minute. Perfectly knowing the sign of that move would be worth around 6 bps a trade. A "discovered" edge of 3.9 bps is therefore not a small effect that got lucky; it is most of one minute of foresight, which is exactly what the off-by-one label handed you. Whenever a one-bar shift swings the result by several multiples, the timing is the result.

In practice

Store both times and index on arrival time, keeping event time as an ordinary column for analysis. Build the decision timestamp explicitly — arrival plus feed latency plus compute time plus the wait for the next tradeable moment — and make the simulator refuse any order stamped earlier. An Event-Driven Backtest Architecture enforces this for free, which is most of the argument for using one. When an interviewer asks how you would validate a fundamentals-driven backtest, your first sentence should be about report dates, not Sharpe.