Edge Degradation
Trading Intelligence
8 min read
edgeStabilityedgeDecayRate
Detect when a strategy is losing its edge through regime change, crowding, or structural market shifts before it becomes a losing system.
Loading
8 min read
Detect when a strategy is losing its edge through regime change, crowding, or structural market shifts before it becomes a losing system.
No strategy lasts forever. Here's how to spot when your edge is fading — before your account does.
Prereq: Biases in Backtesting — without it you'll mistake an overfit strategy for a decaying one. Next: Outliers and Their Impact on Metrics — a single tail trade can fake a regime shift in your stability score.
Edge degradation is a statistically detectable decline in a strategy's expected value, caused by crowding, regime change, structural market change, or leakage discovery. Distinguish it from normal variance using a CUSUM chart or rolling t-test on per-trade R. When detected, choose between retire, refit, or wait based on the cause — not on emotion.
Every system has a lifecycle.
Most traders:
But smart traders detect edge degradation statistically — and adapt before a full breakdown.
This post shows you how.
Edge degradation means your once-profitable strategy:
Not due to a few random losses — but due to a statistical shift in the system’s performance.
Markets are competitive systems. Any signal that produces excess return is a target. The decay process is information diffusion: one trader finds an inefficiency → a desk replicates it → an academic paper publishes the factor → a retail platform builds a screener for it → the original alpha is now priced in by 10,000 participants front-running each other into the trade. McLean & Pontiff (2016) measured this explicitly: published equity factor returns drop ~58% after publication. Your edge has the same fate unless you find one nobody else is hunting.
McLean & Pontiff (2016). The headline crowding-decay benchmark — once a factor is published, roughly three-fifths of its excess return vanishes as capital crowds in.
Four distinct decay mechanisms:
Great traders don't just build edge — they monitor it continuously.
Instead of only checking lifetime stats, track:
Plot these over time to see statistical trends.
"My average EV dropped from +0.6R to +0.2R over the last 100 trades" → Time to slow down and re-evaluate.
Graph your PnL or equity curve.
Look for:
These are often the early symptoms of a fading edge.
Segment your trades by:
You might find:
"This strategy worked well in high-volatility BTC days… but fails during chop."
Now you can filter or evolve it, not abandon it blindly.
Loss streak ≠ broken system. That's just normal distribution noise — see Outliers and Their Impact on Metrics for why a single tail trade can also masquerade as a regime change. And before trusting your in-sample numbers at all, audit your backtest for the failure modes covered in Biases in Backtesting.
A 55% win rate strategy will produce a 10-trade losing streak somewhere in 200 trades roughly 30% of the time. A 12-trade streak: ~18%. If you retire your system every time you see one, you will retire winning systems and keep replacing them with new untested ones — the trader's equivalent of shooting your own dog because it barked. Run the test before you reach for the trigger.
Long losing streaks are normal variance, not edge decay
At a 55% win rate, even a 12-trade losing streak shows up almost one time in five over a 200-trade window. Retiring on a streak retires winning systems.
Use a test, not a vibe:
Without a pre-committed test, you will either capitulate in week 3 of a normal drawdown or bag-hold a dead strategy for a year.
| Signal | Variance | Decay |
|---|---|---|
| Loss streak length | Within bootstrap 95% CI of historical streaks | Beyond 99% CI |
| Rolling 50-trade EV | Oscillates around in-sample mean | Trends down monotonically over 100+ trades |
| Drawdown | < 1.5x worst in-sample | > 2x worst in-sample |
| Regime explanation | Performance constant across vol regimes | Performance bifurcates cleanly by regime |
| Welch's t-test (last 50 vs prior 200) | p > 0.10 | p < 0.05 |
You're likely in edge degradation — and must adapt.
Cut size in proportion to your loss of confidence in EV. If your prior was +0.6R and rolling 50-trade EV is now +0.2R, you have ~1/3 the edge and Kelly says ~1/3 the size. Do this before a forced cut from drawdown — voluntary de-risking preserves both capital and the ability to think clearly.
Don’t scrap everything. Instead:
Try:
Log these as separate strategies until they’re proven
Think of your system as:
A living framework you periodically tune — not a sacred formula you never touch.
Rolling metrics and equity curve analysis tell you that something is changing, but the Edge Stability Score quantifies how consistently your edge has performed across the life of your trade log.
The formula normalizes the standard deviation of segment means against the overall mean:
Edge Stability Score = 1 − (sigma_segments / |mu_R|)
This is an ad-hoc inverse coefficient-of-variation — useful as a triage flag, not a hypothesis test. The 0.85 / 0.65 / 0.50 cutoffs are heuristic thresholds tuned for ~200-trade logs; they have no closed-form distributional meaning. For inference, pair this with a CUSUM or t-test on the same segments.
| Edge Stability Score | Interpretation |
|---|---|
| 0.85 - 1.0 | Excellent. Your edge is remarkably consistent across time periods. |
| 0.65 - 0.84 | Good. Normal variation exists but edge persists throughout. |
| 0.50 - 0.64 | Marginal. Performance is noticeably uneven. Investigate which segments underperform and why. |
| Below 0.50 | Concerning. Your "edge" may be concentrated in one or two favorable periods rather than being a true, durable advantage. |
A stability score below 0.50 typically means one of the following:
Edge Stability pairs powerfully with rolling EV and drawdown analysis. Two honest caveats: (1) with 200 trades, a Welch's t-test has low power to detect a 30% EV drop — you will miss real decay. (2) A high stability score on a curve-fit backtest means nothing; the optimizer was rewarded for producing exactly that. Stability scores are diagnostic on live trades, not on backtests.
| Diagnosis | Statistical signal | Action |
|---|---|---|
| Normal variance | t-test p > 0.10, drawdown < 1.5x in-sample max | Wait. Do nothing. |
| Regime-attributable decay | Performance bifurcates cleanly by a regime tag (vol, trend) | Refit: add the regime as a filter, retest out-of-sample. |
| Crowding / publication | Slow monotonic decline over 200+ trades, no regime explains it | Retire. Crowding rarely reverses. |
| Structural change | Sharp break tied to a known event (fee change, listing, microstructure update) | Retire or rebuild from scratch — the old assumptions are dead. |
| Leakage discovered | Look-ahead or survivorship bug found in code | Retire immediately. The edge was never real. |
Use the simulator to explore how different win rates and payoff ratios produce different equity shapes. An edge that is decaying will show the curve bending downward over time — compare high and low win rates to see this effect.
Edge degradation is a statistically detectable decline in a strategy's expected value. The strategy no longer has positive EV, underperforms current market conditions, or shows a measurable change in win rate, R:R, or risk profile — not from a few random losses, but from a real shift in the system's performance distribution.
A losing streak is variance — a 55% win rate strategy will produce a 10-trade losing streak somewhere in 200 trades roughly 30% of the time. Edge decay is a sustained statistical shift: rolling 50-trade EV trends down monotonically, drawdown exceeds 2x your in-sample max, and a Welch's t-test of recent vs prior trades returns p < 0.05.
Diagnose the cause first, then decide. Normal variance: wait. Regime-attributable decay: refit with a regime filter and retest out-of-sample. Crowding or structural change: retire. Always cut size voluntarily in proportion to your loss of EV confidence before drawdown forces you to.
Strategies don't die -- they evolve, or they erode.
Your edge is not guaranteed forever. But with disciplined tracking and objective review, you can:
Don’t just trade the market. Trade your data.