The Edge I Assume Is Already Decaying

Two headlines from this week, sitting next to each other. One: systematic funds gave back a quarter of their year-to-date returns in a single AI-led selloff, their worst stretch since last August. Two, from the day before the quarter turned: Wall Street is now worried that the AI race is compressing the useful life of a trading edge from something like seven years to something like eighteen months.

I have no way to verify the eighteen-month number, and neither does the person who said it. It is a vibe with a decimal point. But the direction is not controversial to anyone who has watched a good backtest go quietly flat, and the direction is the whole post.

An edge is not a fact, it is a half-life

The mistake that a clean backtest invites is to treat an edge as a discovery: found once, owned forever, like a theorem. It is not. An edge is a mispricing that pays you for as long as not enough other people are taking the same side. The moment the trade is legible enough to be copied, it is being copied, and every dollar that copies it is a dollar of your future return spent early.

So the honest model is not "does this signal work." It is "how fast is this signal dying." Put a crude number on it. Say an edge decays exponentially with some half-life T, so the Sharpe it still delivers at time t is roughly its original Sharpe times two to the power of minus t over T. The model is deliberately too simple. It is still enough to see the consequence.

The consequence lives in the area under that curve, because the area is the total alpha you get to harvest before the edge is gone. For exponential decay that area scales directly with the half-life. Cut the half-life from seven years to eighteen months and you have cut it by a factor of about four and a half, which means the same discovery, the same clever signal, the same afternoon of insight, is worth a bit under a quarter of what it was worth in the slower world. Not because the idea got worse. Because the clock got faster.

That is the part I want to sit with. Nothing about your research changed. The value of research changed, and it changed underneath you, and the backtest cannot show it to you because the backtest runs on the slow-world tape where the edge was still fat.

You cannot defend an edge, you can only outrun the decay

If the value of a signal is set by its half-life, then the thing that produces long-run return is not the quality of any one signal. It is the rate at which you find new ones. Two shops with identical research skill but different discovery cadences do not converge. The faster one compounds; the slower one is always harvesting the flat tail of ideas that already paid out.

This reframes a pile of housekeeping I used to file under overfitting. Walk-forward testing, strict out-of-sample holdout, re-fitting on a rolling window: I had them in my head as guards against fooling myself with a curve I over-tuned. They are that. But they are also the only instruments I have that can measure decay, because they are the only ones that ask how a rule behaves on tape it was not born on. An out-of-sample window that is quietly worse than the in-sample one is not always an overfit. Sometimes it is a real edge, correctly found, that is simply four months further into its own half-life.

The default I actually run

So I do not get to assume that a rule which backtested well last quarter is still alive this quarter. The safe default is the opposite: assume it is already decaying, and make the system prove it is not. That is a colder posture than "I found something that works," and it is the correct one, because the alternative is to keep sizing into a signal on the strength of a return it earned in a regime that has already ended.

The week's headlines are two readings off the same instrument. Crowding is decay you can see from orbit. The quant drawdown is what the flat tail feels like from inside a book that was still betting on the fat part of the curve. Neither is a reason to stop looking for edges. Both are a reason to assume the last one is already going, and to have the next one most of the way found before you need it.