动量扫描反转延续策略 (Momentum Scan Reversal Continuation Strategy): 90 segments tested, 49 profitable, deepest drawdown 38.6%

Bets on momentum continuation: long on a moving-average golden cross and out on the death cross, with RSI crossing above price as confirmation; sideways markets churn out repeated losses.

This strategy is not a steady-output machine; it is a strategy that lives off a small number of segments. In 90 backtests, 49 segments were profitable, and the median segment return was only +1.1%; the deepest drawdown across the 90 segments was 38.6%. The number of profitable segments is close to half, which does not look bad at first; but once the return distribution is laid out, the conclusion becomes much more cautious. Let us put the key numbers on the table. Across the 90 backtests, median segment return was +1.1%, the highest was +101.2%, the lowest was -25.7%, the deepest drawdown was 38.6%, and 0 segments were blown up. The best segment was 2023-09-01 to 2024-03-05, with segment return +101.2% and its own drawdown 9.9%; the worst segment was 2024-03-05 to 2024-08-21, with segment return -25.7% and segment drawdown 30.3%. Average profit per trade was $3,137, and average loss per trade was $1,678. Over the same period, “buy and hold” with no trading at all had an average segment return of +29.3%. These numbers together show one thing: profits are not evenly distributed; they are pulled up by a very small number of strong trend segments, and the returns of most segments are far below doing nothing. What is it betting on? The core assumption is that once momentum forms, it will continue—the strategy scans for momentum signals and enters after a trend starts, hoping to use a few large market moves to cover many small losses. It relies on profit/loss ratio rather than win rate: average profit per trade of $3,137 is clearly higher than average loss of $1,678, so even if only 49/90 segments are profitable, as long as the large segments are large enough, the total account can still be positive. Conversely, once the large segments do not arrive, this logic loses its support. The weakness is most obvious in range-bound and low-volatility environments. When momentum signals repeat within a range, they get interrupted back and forth, the “continuation” assumption fails, and loss segments tend to cluster. The worst segment of -25.7% corresponds to a 30.3% drawdown within the segment, showing that it does not lose slowly but gives back quickly when the trend reverses. Parameter sensitivity is also a hidden risk: the original text does not give the values for interval, fast/slow moving-average periods, RSI period, and oversold threshold; these were filled in according to industry convention. With another set of values, the 49/90 ratio could very possibly change significantly. The two numbers that most let readers judge are these. First is the huge gap between the median segment return of +1.1% and the highest of +101.2%: using 90 ways to split the same dataset, the conclusion can jump from “barely makes money” to “a single segment with triple-digit percentage return”—a typical signal of overfitting; change the split and the conclusion reverses. Second is the deepest drawdown of 38.6%: to get that +101.2%, one must first be able to withstand a drawdown close to four-tenths. Compared again with “buy and hold” average segment return of +29.3%, the strategy median is only +1.1%; under most splits it did not outperform “hold and do nothing.” The limitations of the method need to be stated. The 90 backtests are 90 ways to split the same dataset (15 rounds × 6 segments, with window boundaries randomly jittered), not 90 independent samples; statistically they cannot be treated as 90 independent trials. The random seed has been stored in the database, and the splits are reproducible, but the source of the above 5 parameter values is not in the original text, so this part cannot be reproduced. Readers should verify the sample period and the unreproduced parameter sensitivity themselves. This content is a reproduction verification of historical data and does not constitute investment advice; this platform does not recommend, rate, or rank.