Post-mortem: why we killed a strategy that "worked"

In early June we retired one of our mean-reversion strategies — a multi-leg momentum system we'd been testing for weeks. Its backtest was technically positive. We killed it anyway. This is the post-mortem, published in full because the reasoning matters more than the result.

What the strategy did

It traded short-term momentum continuation in multiple legs — entering on a trigger, then scaling as the move confirmed. On paper, the appeal was a high reward when a move ran. In practice, most moves didn't run.

The three findings that killed it

1. The backtest edge was too thin to survive reality. A profit factor of 1.05 means gross wins barely exceeded gross losses — before slippage, fees and imperfect fills. Once realistic trading costs were applied, even the strategy's best market regime barely broke even. A thin edge in simulation is usually no edge in production.

2. Live trading confirmed the doubt, fast. Thirteen paper trades produced a profit factor of 0.67 — losing 33 cents for every dollar of gross wins it needed. Small sample, yes. But when a thin backtest edge and a negative live sample agree, you don't wait for a bigger sample. You stop.

3. It never earned a slot. Our capital allocation works like a promotion ladder: research → paper → testnet → live. This strategy had been running long enough to earn promotion and hadn't. Capital and attention are finite; a strategy that can't prove itself is taking both from strategies that can.

The discipline part: kill criteria are written in advance

Every Glasshouse strategy ships with pre-committed kill criteria — the conditions under which we stop it, agreed before it trades. This matters because the alternative is what kills most traders: improvising reasons to keep a losing system alive ("it's just a bad regime", "one more month", "let me tweak a parameter").

If you decide what failure looks like before you start, you can't negotiate with it later.

The kill is reversible in one case only: a genuinely new edge thesis — not a parameter tweak on the same idea. Re-fitting parameters until a backtest looks good again isn't research, it's curve-fitting with extra steps.

What it cost — and what it didn't

Because of the ladder, this strategy never touched real money. The cost was research time. That's the system working as designed: strategies are cheap to kill on paper and expensive to kill in a live account. We aim to do all of our dying in the first category.

Lessons we're carrying forward

• A profit factor near 1.0 in backtest is a "no", not a "maybe" — costs eat thin edges first.
• Paper trading isn't a formality; it's where simulation meets order flow, and it disagrees with backtests more often than you'd think.
• Killing fast preserves the thing that actually compounds: the discipline to only run what's earned it.

The full strategy book — including this entry in the kill log — is on the research page.