You put a strategy live, and after three red weeks you wonder: normal drawdown, or is the edge gone? Deciding that in writing beforehand keeps the decision out of your gut. That part is right. The thresholds now being shared as the professional standard, however, barely tell the two cases apart. They brake a healthy strategy almost as often as a dead one.
To be clear: kill rules are not a stop-loss for single trades. They judge the strategy as a whole and decide whether you trade it with less money or switch it off.
What Horizon proposes
On 11 June 2026 Horizon, a vendor of backtesting and trading software, published an article on X titled "How Top Quants Catch a Dying Strategy Before the Drawdown (Full Framework)". By early October it had close to 169,000 views.
The core rules, checked weekly against the backtest:
- Soft breach: the rolling 90-day Sharpe falls below the backtest's 5th percentile, or the drawdown exceeds 1.0x the backtest maximum. Action: halve position size.
- Hard breach: the drawdown exceeds 1.5x the backtest maximum, or time underwater exceeds 1.5x the longest backtest stretch. Action: pause the strategy.
Much of it is right. The kill rules should be in writing before the strategy goes live, because "the version of you that is down 12% will negotiate". The question is not only whether your rules are good, but whether you still follow them when you are 12 percent down. Measure time underwater, not only depth. And a strategy with a true Sharpe of 1.0 really does post a losing year about one time in six.
One sentence in the article reads: "The exact multipliers matter less than the fact that they exist in writing before launch." That is something you can calculate.
Healthy strategies get braked almost as often as dead ones
The simulation: 5,000 strategies, each with a three-year backtest and a true Sharpe of 1.0. Then they run live, either unchanged, with half the edge, or with no edge at all. The count is how often Horizon's rules fire.
| Live state | Period | position halved (soft) | paused (hard) |
|---|---|---|---|
| intact, Sharpe 1.0 | 1 year | 48% | 6% |
| intact, Sharpe 1.0 | 3 years | 81% | 33% |
| edge halved, Sharpe 0.5 | 1 year | 61% | 11% |
| edge halved, Sharpe 0.5 | 3 years | 90% | 51% |
| no edge, Sharpe 0 | 1 year | 72% | 18% |
| no edge, Sharpe 0 | 3 years | 95% | 70% |
A strategy where nothing changed gets its position halved 81 percent of the time within three years and paused a third of the time. A strategy with no edge at all runs through its first year unpaused 82 percent of the time.
The rules react to everything a little and to nothing in particular.
Why is the gap between healthy and dead so small?
Two thresholds that are bound to break
The 5th percentile breaks by definition. Five percent of all 90-day windows of a healthy strategy sit below it. Check every week and sooner or later you land in one.
The backtest maximum is the extreme of a single path. A second path of the same strategy and the same length exceeds it in 51 percent of the simulated cases. The longer a strategy runs live, the deeper its worst drawdown gets.
Horizon knows the problem. The article recommends resampling the trades thousands of times to build a "performance cone". The breach rules then measure against the one maximum after all.
A threshold a healthy strategy is bound to break is not a warning. It is a calendar.
And the dead strategy? Horizon writes about overfitted strategies: "These die fast, usually within weeks. Let them."
Dead strategies do not die within weeks
A strategy without an edge does not lose fast. It loses slowly and unevenly, and in its first year its drawdowns look much like those of a healthy one. That is why the rules pause only 18 percent of them in year one.
How hard the distinction is in principle shows in a clean test. Instead of checking four metrics every week, test once at a fixed point whether the live Sharpe is significantly below the backtest Sharpe, at a 5 percent false-alarm rate:
| Live state | after 1 year | after 3 years |
|---|---|---|
| intact | 5% false alarm | 5% false alarm |
| edge halved | 11% detected | 15% detected |
| no edge | 22% detected | 34% detected |
The false alarm is now under control. In exchange you see what the kill rules cover up: even a completely dead strategy is distinguishable from a healthy one only a third of the time after three years.
At a Sharpe around 1, one year of live data cannot separate bad luck from decay, whatever rule you use.
If monitoring cannot do that, where do you start?
The more important number comes before launch
The simulation is generous. It assumes the backtest Sharpe is true. In practice it is usually too high, because you did not test one variant but many and kept the best.
How strong that effect is: test several variants of a strategy with no edge over three years and keep the best.
| variants tested | expected best Sharpe with no edge |
|---|---|
| 1 | 0.0 |
| 5 | 0.7 |
| 20 | 1.1 |
| 100 | 1.5 |
Try 20 parameter combinations, land on a Sharpe of 1.0, and you hold exactly what chance delivers. Any monitoring that uses that 1.0 as its baseline is guarding a number that never existed.
Horizon cites McLean and Pontiff (Journal of Finance, 2016): 97 published return predictors earned 58 percent less after publication. The authors split the figure themselves. 26 percent was already missing out of sample, before anyone knew the study, an upper bound for data mining. The remaining 32 points they attribute to trading after publication. Horizon's reading that most of it is crowding holds. Part of the edge was traded away, and part of it was never there.
The Deflated Sharpe Ratio (Bailey and López de Prado, 2014) addresses the second part. It does not ask how high the Sharpe is, but whether it beats what chance alone delivers after that many attempts.
The objection
"Real markets are not normally distributed. Maybe the rules work better there."
Recalculated with fat tails (a t-distribution with four degrees of freedom, same volatility), almost nothing changes: healthy strategies get their position halved 82 percent of the time within three years and paused 33 percent. With alternating calm and wild stretches, the pause rate for healthy strategies rises slightly, to 36 percent. Real markets do not make the rules more accurate, if anything a little more jumpy.
What to do instead
- Check the baseline before you monitor it. Count how many variants you tested and deflate the Sharpe for it. A backtest Sharpe that does not clearly beat chance needs no kill rule, because there is nothing to defend.
- Know how much the metric wobbles. In Backtesting Arena, nudging neighbouring parameters by one or two steps moves the CAGR of a daily backtest so much that only gaps above 17.9 percentage points count as a finding. In the worst case the swing was 45.8 points on a base of about 33 percent. A gap below that is not a finding.
- Few checks, planned in advance. A test at one fixed point has a known false-alarm rate. If you want to check continuously, use methods built for it: a sequential test (SPRT) or a pre-planned testing schedule keeps the false-alarm rate fixed across all checks, and a CUSUM sets how long a healthy strategy runs on average before its first false alarm.
- Measure against the market. A long strategy loses in a bear market without its edge being gone. Whether it has decayed shows in the return against buy and hold, not in the raw curve.
- Validate new parameters only on new data. Everyone agrees on this one, Horizon included.
Tested: you can skip the 5th-percentile, 1.0x and 1.5x thresholds as an early-warning system. That protects you from halving a healthy strategy in a normal drawdown and from trusting a dead one because it has not yet broken a threshold.
Method
Daily returns, normally distributed and independent, 2 percent daily volatility, annualised over 365 days, 5,000 runs per row. Backtest three years at Sharpe 1.0, live one or three years at Sharpe 1.0, 0.5 or 0. Weekly checks; the 90-day window and the drawdown start at the live start. Checking daily instead gives 87 rather than 81 percent halvings for healthy strategies over three years; the pattern is the same in every variant tested. Fixed checkpoint: one-sided test on the Sharpe difference, standard error from both sample lengths. Selection: expected highest Sharpe of K zero-edge variants on three years of data, 20,000 runs. Noise floor: Arena tool arena_is_distinguishable, probe of 31 July 2026, retrieved 8 October 2026.
FAQ
Are kill rules the same as a stop-loss? No. A stop-loss closes a single trade once it runs too far into the red. Kill rules judge the whole strategy across many trades and decide whether it keeps running with less money or gets switched off.
Did Horizon test the rules? The article reports no test of the thresholds. It says the exact multipliers matter less than having them in writing.
Does this hold at a higher Sharpe? The higher the true Sharpe, the more clearly a dead strategy stands out. The figures here apply to a Sharpe around 1, a typical size for strategies that go live.
Should I have no kill rules at all? You should, but with thresholds whose false-alarm rate you know, against a baseline you checked against chance first.
Not investment advice, not a recommendation, not a forecast — historical patterns are no promise.
Study the Past — Improve your Future 🥋