Quant
Strategy decay — knowing when to turn it off
You did everything right. The backtest was honest, the ladder was climbed, the limits are in place, and the thing has been quietly working for eight months. Then it stops. Now you face the hardest question in the whole business, and it's one no backtest can answer: is this a normal drawdown, or is the edge gone? Get it wrong in one direction and you kill a working strategy at the worst possible moment. Get it wrong in the other and you fund a corpse.
Edges are perishable
Start by accepting the premise. Every edge decays, for reasons that are structural rather than unlucky:
- Competition. If it works, others find it, and their trading is what erases it. This is the random walk's central mechanism — exploitable structure destroys itself precisely because it's exploitable.
- Regime change. A strategy that fed on a decade of falling rates or low volatility was implicitly a bet on that regime, whether or not you knew it.
- Capacity. Your own size becomes the counterparty. The edge is fine; you're just too big for it now.
- Structural change. A fee schedule, a tick size, an exchange rule, a new participant. The market you tested no longer exists.
So the question is never "will it die?" It's "will I notice?"
Drawdown and death look identical
Here's what makes this genuinely hard rather than merely unpleasant. A perfectly healthy strategy with a real edge produces losing streaks that look exactly like death — that's what variance is. Staring at the equity curve cannot separate them, because both are a line going down.
But you already built the tool that can. The Monte Carlo you ran to validate the backtest didn't just give you a pass/fail — it gave you the distribution of outcomes a working version of this strategy produces. Shuffle the trade order a few thousand times and you learn the range of drawdowns, the length of losing streaks, and the shape of equity curves that are entirely normal for your edge. That range is your reference for "behaving like itself."
Kill on behaviour, not on P&L
This is the distinction that makes the whole thing tractable. Losing money and being broken are different events. A strategy can lose for months while behaving exactly as designed — that's a drawdown, and killing it there is how you lock in the bad half of your own distribution. A strategy can also make money while thoroughly broken, which is far more frightening, because nothing prompts you to look.
So monitor behaviour, and let P&L be a consequence rather than a trigger. The things worth watching are the ones your backtest predicted:
- Drawdown against the Monte Carlo floor — inside the band, do nothing; through the 95th percentile, act.
- Win rate and payoff ratio drifting outside their simulated bands — the shape of the edge changing even when the total hasn't yet.
- Trade frequency — firing half as often as expected means the setup itself is disappearing from the market.
- Tracking error vs expectation — the ladder's question, still running in production. Live diverging from predicted is a signal regardless of direction.
- Realised costs — slippage creeping above your model quietly converts a positive edge into a negative one without changing a single signal.
Every edge eventually dies, and a healthy strategy's bad month looks just like a dead one. The simulation you ran to check the backtest tells you which losing streaks are normal — so use it as the tripwire, and turn it off when it stops behaving like itself, not merely when it loses.
Write the kill criteria before you deploy
The most important thing here isn't the metric. It's the timing of the decision. In the moment — down 15%, eight months of work on the line, convinced the mean reversion is coming — you will not be a rational judge of your own strategy. Nobody is. You'll find a reason to wait one more week, and then another.
So decide while you're calm and disinterested, before there's any money on it. Write down the specific numbers: drawdown beyond the simulated floor, N consecutive losing weeks, a metric outside its band for M days. Then treat the tripwire as binding. The rule you set in advance is smarter than the person you'll be when it fires — that's not a comment on you, it's the entire reason the rule exists.
Halt is not delete
One relief: the decision isn't permanent. Halting costs you the option value of a strategy that was about to recover — which is real but small. Not halting costs you the whole account if the edge is genuinely gone. Given that asymmetry, halt early and investigate honestly. Sometimes the answer is capacity, and it comes back at a third of the size. Sometimes it's a regime, and it waits. Sometimes it's dead, and you've learned something for the next one.
And when a strategy dies, resist the urge to immediately re-fit it on the recent data. That's not a repair — it's a fresh backtest on the smallest, noisiest sample you own, with a strong motive to find something. It starts back at the bottom of the ladder like everything else, and it needs the same evidence the original did. The edge that died gets no credit for the one replacing it.