Bollinger band reversion shows no significant edge
Daniel Okafor, Strategy analyst (AI-assisted) · · updated
The performance of the Bollinger band reversion strategy applied to the AUDUSD currency pair has been tested, and the results indicate that this approach does not provide a significant trading edge. The out-of-sample Sharpe ratio was recorded at 0.45, with a p-value of 0.065 after applying the Benjamini-Hochberg correction [1]. This suggests that while the strategy shows some positive returns, its edge over the zero benchmark is not statistically significant after the correction [1].
The Testing Framework
To evaluate the effectiveness of the Bollinger band reversion strategy in out-of-sample conditions, our framework involved conducting 90 distinct tests across 15 different currency markets [1]. We focused on various settings for the Bollinger bands, ultimately determining that a 10-day period was the most effective based on in-sample analysis, yielding a Sharpe ratio of 0.43 [1]. In contrast, alternative configurations—20 days and 50 days—resulted in negative Sharpe ratios of -0.15 and -0.32, respectively [1].
The benchmark for our tests was set at zero, reflecting the AUDUSD market's capability for both long and short positions. This choice illustrates the challenge of achieving consistent performance; a rule that earns nothing. Our testing period spanned from January 1, 2016, to November 1, 2026, during which we recorded 292 position changes, each incurring transaction costs of 1.5 basis points [1]. The pass/fail rule was written on 2026-11-04, after the 2026-11-02 run had been read, so this is a labelled rule applied to a finished run, not a pre-registration [1].
This approach required careful consideration of the best-of-N correction, which accounts for the multiple tests conducted. By comparing the performance of our chosen rule against the best outcomes from all 90 tests, we can better understand the likelihood of achieving a result by chance alone [1].
Interpretation of Results
The out-of-sample Sharpe ratio for the Bollinger band reversion strategy stood at 0.45, which at first glance might suggest potential profitability [1]. However, a closer examination unveils considerable limitations. Out of 90 tests performed, 27 resulted in a negative out-of-sample Sharpe ratio, so their out-of-sample Sharpe ratio was not above 0 [1]. Among the remaining 63 tests that showed a positive Sharpe ratio, none maintained significance following the Benjamini-Hochberg correction, indicating that the strategies tested cannot be reliably deemed effective based on the available data [1].
A critical point of this analysis is the observed lack of statistical significance in the measured edge. The p-value of 0.065 falls short of the level the correction required: at rank 1 of 90 it needed to be at most 0.0011 [1]. To illustrate this, if you were to conduct 90 independent tests, you would expect about 4.5 to yield a p-value of 0.05 or lower purely by chance [1]. The strategy in question ranks first among all tests with the smallest p-value of 0.065, but even this does not meet the threshold for significance after correction [1].
Ultimately, these results suggest that while there may be instances of short-term profitability, the evidence does not support a robust or reliable application of the Bollinger band reversion strategy across different trading scenarios. The findings highlight the importance of rigorous statistical validation, as successful outcomes in such tests can often be misleading without thorough scrutiny.
Questions readers ask
What does an out-of-sample Sharpe ratio of 0.45 imply?
A Sharpe ratio of 0.45 means the rule earned 0.45 units of excess return per unit of risk over the 11.2 unseen years (2016-01-01 to 2026-11-01), against 0.00 for the zero benchmark [1]. Its t-statistic was 1.51 and its one-sided p-value 0.065, which does not pass the best-of-90 correction [1].
How does the Benjamini-Hochberg correction work?
The Benjamini-Hochberg correction used here controls the share of false discoveries at 10% across all 90 tests [1]. It ranks the p-values and compares each with a threshold that depends on its rank; for rank 1 of 90 that is 10% divided by 90, about 0.0011 [1]. The best p-value in the run was 0.065, so none of the 90 tests passed [1].
Can Bollinger bands work in other markets?
This test cannot say. It covers one market and 11.2 unseen years, which is one path of history; a rule that failed here can work in another regime [1]. The 15 markets and 6 rule families of the run are in the table above, and none survived the correction [1].