Backtest bias: validate a trading strategy without hindsight · 1 / 5
Find when a pivot signal actually became available
A marker drawn on a turning point does not prove you could have traded there. Reconstruct what the rule knew at each bar close before assigning an entry price.
Athenum7 minUpdated:
Pivot low rule: candidate time vs available-at time
For this original exercise, a strict pivot low must be lower than the lows of two preceding and two following completed five-minute bars. Equal lows do not qualify. Evaluate only at bar close. The two preceding lows are 103 and 102; the candidate bar makes a low of 100. At that point it is only a candidate: neither following bar is known.
A chart may place its label on the candidate after later bars confirm the condition. Store candidate time and available-at time separately. An indicator that updates during an unfinished candle is not automatically defective, but a replay must reproduce the information its actual decision rule used. Do not replace live knowledge with a final cleaned chart.
Confirmation vs execution: fill price and failed candidates
Even after confirmation, an order needs submission and a fill. Record those timestamps independently. A rule evaluated after the 10:15 close cannot assume an earlier intrabar low as its entry, and it cannot guarantee execution at the closing price either. Latency, spread and a gap can change the executable price.
Replay rejected candidates as well as confirmed ones. If the second following bar makes a new low below 100, the candidate fails. Keeping only the historical markers that survived makes it impossible to audit what a trader saw while the setup was forming.
Example: 0.81R after confirmation vs 16R in hindsight
The candidate occupies 10:00–10:05 UTC. The following bars have lows 101 and 102, so confirmation occurs at 10:15. Assume the next executable purchase is 104.20, a fixed stop is 99.50 and a later hypothetical exit is 108. One unit risks 4.70 and earns 3.80: 3.80 / 4.70 = 0.8085R before costs. Prices are USDT per underlying unit with a linear payoff. R measures the planned entry-to-stop distance before costs, not the maximum possible loss.
Assigning a fictitious entry at 100 instead gives risk 0.50 and profit 8, or 16R. Both divisions are mathematically correct; only the first entry respects this signal's availability. Neither calculation predicts that the exit at 108 will occur.
| Interval | Low | State at close |
|---|---|---|
| 10:00–10:05 | 100 | Candidate; two future bars unknown |
| 10:05–10:10 | 101 | One qualifying following bar |
| 10:10–10:15 | 102 | Confirmed at 10:15 |
| 10:15–10:20 | 103 | After initial confirmation |
Open full-size diagram- Candidate low
- Following bars close
- Signal confirmed
- Order submitted
- Fill recorded
Hypothetical five-minute replay, all times UTC. Keep the marker's bar, knowledge timestamp and execution record separate.
Dots show interval lows, not their exact occurrence times or a continuous price path. The arrow spans the two following bars, from 10:05 to 10:15. Formulas below compare assumed entries, not observed fills.
Low: 100 → 101 → 102 → 103. Confirmed at 10:15.
Look-ahead bias: entering at a backdated pivot marker
If the 10:10–10:15 bar instead falls to 99.80, the original candidate fails. A backtest cannot keep the earlier 100 entry while discarding the failed confirmation. Write tie rules and confirmation delays before seeing outcomes.
Before acting
- Freeze pivot and tie rules.
- Store candidate and available-at timestamps.
- Separate confirmation, submission and fill.
- Replay rejected candidates.
- Recalculate risk from the executable entry.
Check your understanding
Under the same completed-bar rule, require three following bars instead of two. The third low is 103. When is the original candidate confirmed?
Show the explained answer
At 10:20 UTC, when the third following bar closes. The low 103 satisfies the strict comparison with 100, but its completed value was not available at 10:15. The entry must follow the new confirmation time.
Sources and further reading
Continue with Athenum
Related lessons
- Open interest backtests: handle delayed data and revisions
Replay open interest with the values actually available at decision time. Track delayed reports and revisions without leaking future information into the test.
- Orderflow lead and lag: align event time with arrival time
A 100 ms lead in exchange time can become a 180 ms lag on arrival. Replay trade messages as received and keep event, publication and receipt times apart.