Alpha decay (strategy decay)

Why edges fade

• Crowding: a pattern that is profitable attracts capital, and the trades that exploit it move prices until the pattern pays less.
• Structural change: new order types, tick size changes, a change in trading hours or a different mix of participants can remove the mechanism an edge relied on.
• Regime change: a rule that needs mean reversion stops working when the market trends for months, and may come back later.
• Rising costs: an edge worth one tick per trade disappears if fees or slippage grow by one tick.
• Overfitting mistaken for decay: an edge that never existed out of sample also looks like it faded right after deployment.

Telling decay from a drawdown

Every profitable strategy has losing periods, and a drawdown alone is weak evidence of decay. The useful comparison is between the recent distribution of results and the distribution the strategy was expected to produce. A rolling statistic, such as a 20-session Sharpe or a 30-trade expectancy, shows the trend, but its own noise is large: over 20 sessions, a rolling Sharpe near zero is compatible with a strategy whose long-run value is clearly positive. Decay is suggested when the rolling measure stays below its historical range for longer than the losing streaks and underwater periods seen in the backtest, and when the average trade falls rather than only the win rate.

A decision rule written in advance, such as stopping after a stated number of trades below a threshold, prevents the verdict from being taken on the most recent bad week.

In Senzoukria

The Gauntlet charts under an automatic backtest include a Rolling Sharpe over 20 session days, computed with the Gauntlet's convention of daily results scaled by the square root of 252; the chart is described in the source as answering whether the strategy still works. The Performance panel adds a Rolling expectancy over 30 trades among its time breakdowns, as long as the trades of the run are still in memory. The walk-forward fold table shows the out-of-sample result of each slice separately, so a strategy whose later slices weaken is visible without averaging. The Performance panel's list of runs also makes it possible to compare the same strategy over successive periods.

Common mistakes

• Declaring decay after a drawdown shorter than the backtest's own worst one.
• Reoptimizing parameters after every losing month, which turns monitoring into curve fitting.
• Ignoring cost changes when comparing old and new periods.
• Reading a rolling statistic computed on a handful of trades as a trend.