Defining Kill Criteria Before Launch

Ask a team that just launched a promising new strategy what would make them shut it down, and the honest answer from most is that they haven't thought about it yet — they're focused on it succeeding. That's exactly backwards. The moment before launch, with no capital at risk and no emotional attachment to a live track record, is the cheapest and clearest-headed moment a team will ever have to decide what failure looks like. Wait until the strategy is actually losing money, and every decision gets contaminated by sunk cost, by hope that the next week will turn it around, and by the natural reluctance to admit a project you championed didn't work.

What a good kill criterion looks like

A kill criterion is a specific, pre-agreed, and ideally quantitative condition that triggers a mandatory review — not necessarily an automatic shutdown, but a forced conversation that can't be waved away with "let's give it more time." Good ones share a few properties: they're set before the strategy has live performance to react to, they're specific enough that reasonable people can agree whether they've been triggered, and they cover more than one failure mode, because strategies fail in different ways.

Type of criterion	Example	What it catches
Drawdown limit	Loss exceeds a set multiple of expected volatility, or a set fraction of allocated capital	A strategy behaving far outside its modeled risk envelope
Time without profit	No net gain after a pre-agreed number of months at target size	A strategy quietly failing to deliver, without an alarming single event
Signal degradation	A rolling measure of predictive power (like a signal's IC) falls below a floor for a sustained period	Genuine decay rather than a bad month
Correlation breach	Realised correlation to other live strategies exceeds a set threshold	A strategy that stops diversifying the broader book
Cost overrun	Realised transaction costs exceed backtested assumptions by a set margin	A strategy whose edge was smaller than modeled once real execution is included
Operational trigger	A data feed becomes unreliable, a venue relationship ends, a key assumption about market structure changes	External conditions invalidating the strategy's premise, independent of its P&L

Worked example

A team is about to allocate capital to a new statistical arbitrage strategy with a backtested annualised Sharpe ratio around 1.2. Before launch, they agree on four kill criteria: a drawdown exceeding twice the strategy's expected monthly volatility triggers an immediate review; six months at target size with no net profit triggers a mandatory reassessment of the thesis; realised transaction costs exceeding backtested assumptions by more than 50% over any rolling quarter triggers a capacity and sizing review; and if the strategy's correlation to the desk's existing book exceeds 0.4 for two consecutive months, it gets reviewed for redundancy regardless of its own standalone performance. Four months in, the strategy hasn't lost money, but transaction costs have run 60% above the backtest's assumption because the actual order sizes are moving prices more than modeled. The cost-overrun criterion triggers automatically. Because the team pre-committed to a review rather than an automatic shutdown, they investigate, find the strategy's true capacity is roughly a third of what was originally planned, and resize it accordingly rather than either shutting it down unnecessarily or continuing to run it at a size the market can't absorb without eating the edge.

In practice

• Set kill criteria as a condition of launch, not an afterthought. A strategy that can't get through a five-minute conversation about "what would make us stop" before launch isn't ready to launch.
• Prefer a mandatory review over an automatic shutdown for most criteria. The goal is to force a conversation immune to sunk-cost reasoning, not necessarily to remove human judgment entirely — some triggers deserve investigation rather than a reflexive kill.
• Revisit criteria as the strategy matures, particularly as more live data becomes available to recalibrate what a "normal" bad stretch looks like; see How Long to Incubate a Strategy for how much data you need before those numbers are trustworthy.
• Write the criteria down somewhere everyone can see them, not just in one researcher's head — a criterion that only the strategy's champion remembers tends to get quietly renegotiated under pressure.
• Kill criteria and decay detection are related but distinct disciplines. A kill criterion is a pre-agreed trigger; Detecting Decay in a Live Strategy is the ongoing diagnostic work of figuring out why performance is sliding once a trigger fires.