What We Kill

There's a Zen concept that writers know well. Kill your darlings. The idea is simple: the sentence you love most, the one you're most proud of, is probably the one hurting the piece. Your attachment to it is blinding you. Cut it, and the work gets better.

The same thing is true in quantitative research. Except the darling isn't a sentence. It's a strategy you spent three weeks building. One that looks beautiful in backtest. One that tells a story so clean you can already see it in the pitch deck.

We kill those. That's the job.

Most of what we build doesn't survive. Not because we're bad at building. Because survival is the test, and the test is designed to be hard. Build the thing. Train it on historical data. Let it learn every pattern it can find. Then show it data it has never seen. Data from a different period, a different regime, a different world than the one it was trained on.

If the unseen data says no, the answer is no.

No exceptions. No "let me adjust one parameter." No "it would have worked if we exclude that one bad month." No.

This sounds obvious. It is not obvious. It is, in practice, one of the hardest disciplines in quantitative finance. Because the strategy sitting in front of you looks good. The backtest curve goes up and to the right. The numbers are strong. You built it. You understand why it should work. Every instinct says deploy it.

And the holdout test says it doesn't work.

What happens next is the entire game.

At most firms, what happens next is negotiation. The researcher tweaks a parameter. Adjusts the lookback window. Removes an outlier month. Re-runs the test. The curve looks a little better now. They tweak again. Run again. Better still. After enough iterations, the holdout test passes, and the strategy goes live.

This is not research. This is curve-fitting with extra steps. The researcher has, without realizing it, trained the strategy on the holdout data. The whole point of the holdout was to be untouched, and now it's been touched a dozen times. The lock has been picked. The test means nothing.

We don't allow that negotiation. The holdout is sacred. You get one shot. Pass or die.

It feels brutal. It is brutal. We've killed strategies that took weeks to develop. Strategies built on sound economic logic. Strategies where the in-sample performance was extraordinary. The holdout said no, and we walked away.

But here's the thing most people miss about this process. The edge isn't in the strategies that survive.

Every fund has strategies that work in backtest. Every fund has a handful of ideas that look good on paper. The difference between a fund that compounds and a fund that blows up is not the quality of the winners. It's the rigor of the filter that keeps the losers out.

Think about it from the other side. You're an investor evaluating two funds. Fund A shows you their five best strategies. Impressive returns. Clean equity curves. Fund B shows you their five best strategies, and also shows you the graveyard. Every strategy they built and killed, with full documentation of why it died.

Fund A is showing you a highlight reel. Fund B is showing you a process.

Which one do you trust with your money when markets stop cooperating?

The graveyard is not a failure log. It's evidence that the filter works. Every dead strategy is proof that something with a good story and good numbers was stopped before it could lose real capital. The longer the list of dead strategies, the more you should trust the ones that survived.

There's a deeper principle here, and it has to do with attachment.

When a researcher builds a strategy, they develop a relationship with it. They understand its logic. They've debugged its code at midnight. They've watched its equity curve climb during backtest and felt the quiet satisfaction of having found something. That attachment is natural. It's also dangerous.

Attachment creates bias. The researcher who loves their strategy will interpret ambiguous evidence in its favor. They'll see a mediocre holdout result as "promising" rather than "insufficient." They'll find reasons to keep going when the honest answer is to stop.

The system has no darlings. It has no attachment.

When reality says something doesn't work, there's nothing to hold on to. No ego invested in the outcome. No weeks of effort creating a sunk cost. Just a clear signal: this one didn't make it. Document the death, record what it taught us, move on.

This is harder than it sounds. Not technically. Emotionally. There's a real cost to killing something you built. But the cost of deploying something that doesn't work is measured in capital, not feelings. And capital is what we're paid to protect.

We've been asked why we share the graveyard during diligence. Wouldn't it be smarter to just show the winners? Wouldn't investors be more impressed by a clean story?

Maybe. But we've noticed something. The investors who are most serious, the ones who have been doing this long enough to know what matters, they don't spend much time on the winning strategies. They spend their time in the graveyard. They want to know what died and why. They want to see the evaluation methodology. They want to understand the filter.

Because they know something from experience. Returns come from the strategies you run. But survival comes from the strategies you don't.

The Zen masters had it right. The quality of the work is determined by what you're willing to remove. The garden isn't defined by the flowers. It's defined by the weeding.

We weed constantly. We weed aggressively. We weed things that look beautiful.

What's left is what's real.