Edge Decay: Detect When a Strategy Stops Working

Every edge dies eventually — markets adapt, the setup gets crowded, or the regime that fed it disappears. The hard part isn’t accepting that. It’s telling the difference, in real time, between a strategy that’s temporarily cold and one that’s genuinely broken.

Drawdown vs Decay: The Crucial Difference

A drawdown is a losing stretch produced by variance in a system whose edge is still intact. Decay is a losing stretch produced by the edge itself shrinking or vanishing. They feel identical from inside — red trades, shrinking account, mounting doubt — but they demand opposite responses.

Respond to a drawdown by staying the course: the math still favors you, and quitting locks in the loss right before the recovery.
Respond to decay by reducing or retiring: the math no longer favors you, and staying the course just funds the market’s new efficiency.

Get it backwards and you either abandon a good system at its worst moment or keep feeding a dead one. The whole discipline of decay detection is about not confusing the two — which means you cannot rely on how the equity feels. You need a measured signal.

Rolling-Window Expectancy and Win Rate

The core tool is a rolling window: instead of one lifetime expectancy number, compute it over the last N trades and slide the window forward with each new trade. Plot that rolling expectancy and rolling win rate as their own time series.

A healthy system’s rolling expectancy oscillates around a stable mean — up in good stretches, down in bad ones, but always reverting. Decay looks different: the rolling line steps down to a new, lower level and stays there. Not a dip, a shelf.

Two practical choices decide whether this works:

Window size — too short and every cold streak screams “decay”; too long and you notice the break months late. Match the window to your trade frequency: enough trades for signal, recent enough to matter.
Units — track expectancy in R so position-size changes don’t distort the picture. If you’re not normalizing yet, R-multiple explains why raw dollar P&L makes decay nearly impossible to see. Feed the window through an expectancy calculator and compare each window against your long-run baseline from the trading expectancy definition.

Confidence Bands as a Broken-Edge Alarm

A rolling number alone still fools you, because a small window is noisy — twenty trades can print a scary expectancy purely by chance. The fix is a confidence band around your baseline edge.

The idea: given your historical win rate and sample size, statistics define a range the observed rate should stay inside if the true edge is unchanged. A Wilson confidence interval is the right tool here — it behaves well at small samples and near extreme win rates, where the naive interval breaks down. When your recent window drifts outside the band your baseline predicts, that’s a statistically meaningful signal that something changed — not just noise.

This is precisely the alarm Shibiki builds into live edge health: each strategy carries an expectancy wrapped in a Wilson interval, and the band widens or tightens with the sample so a hot or cold streak doesn’t trip a false alarm until the evidence supports it. It turns “this feels off” into a threshold you set in advance and can actually act on — instead of relitigating the decision every red day.

Regime Tags: What Changed in the Market

A statistical break tells you that the edge changed; it doesn’t tell you why. That’s where regime tags come in. Annotate your trades with the market context they happened in:

Volatility regime — was realized volatility high or low?
Trend vs range — directional market or chop?
Session / catalyst — normal session, or dominated by a news release?

When decay appears, slice the rolling expectancy by tag. Often the edge didn’t die everywhere — it died in one regime. A breakout system might still print in high-volatility weeks and bleed in quiet ones. That’s not a dead strategy; it’s a filter waiting to be written. This is the difference between “retire it” and “only trade it when conditions X hold.” Auto-journaling every trade with its context is what makes this slice possible after the fact — reconstructing regime from memory is hopeless.

Setting a Rules-Based Retire/Rework Trigger

Decide the exit before you’re in the drawdown, when you’re calm. A workable trigger has three parts:

The signal — rolling expectancy stays below the lower confidence band for a defined number of trades (not one window — persistence).
The action — cut size by a set fraction first, fully retire only if it persists further. A tripwire, then a kill switch.
The floor — a hard risk limit that caps damage while you decide, so a decaying edge can’t blow the account during the diagnosis.

That last piece is where enforcement beats intention. On a funded account, Shibiki can push hard risk limits down to the broker-side EA, so a strategy that’s quietly decaying still can’t exceed the loss cap you set — the limit holds even if you’re not watching or your resolve wavers. A rule you can override in the heat of a drawdown isn’t a rule.

Re-Validating Before You Redeploy

If you rework a decayed strategy — new filter, new regime constraint — it is a new hypothesis, not the old one resurrected. Treat it that way:

Validate out-of-sample — test the fix on data it wasn’t built from, or forward-test it small.
Wait for the sample — a reworked edge needs enough trades to clear the confidence band on the upside before you trust it, exactly as a fresh strategy would.
Watch for curve-fitting — a filter that perfectly explains the past decay may just be a story fit to noise. If you can’t explain why it should work, be suspicious.

Detecting decay well is mostly about pre-committing to measured triggers instead of arguing with a losing streak in real time. Tools like Edgewonk popularized post-trade analytics for exactly this; the leap is moving from reviewing decay after the fact to being alarmed at the moment the band breaks — and having a hard limit hold the line while you respond.