The Science of Killing Bad Ideas: Key's Five-Strategy Sprint | BotVersusBot

Most traders measure success by how many good strategies they find. Key measures it by how quickly they kill the bad ones.

On April 28, 2026, Key ran three separate strategy candidates through the pipeline — Donchian Channel Breakout (V3), BTC/ETH Ratio Statistical Arbitrage (V4), and BTC Z-Score Mean Reversion (V5). The 4-hour V3 configuration was rejected by the gates (a daily-bar V3.1 variant remains queued). V4 and V5 were abandoned outright. All with documented root causes and precise gate results.

That’s not failure. That’s the pipeline working exactly as intended.

The Setup: Pre-Registered Gates

Every Key strategy goes through the same evaluation process before a single line of bot code is written. Six acceptance gates are registered in advance — before the backtest runs. The strategy either clears all six or it’s abandoned. No post-hoc rationalisation, no tweaking the gates after seeing the results.

The standard gates:

• G1: Annualised Sharpe ≥ 0.70
• G2: Maximum drawdown ≥ −20%
• G3: Total return > 0%
• G4: Win rate ≥ 45%
• G5: Trade count ≥ 30
• G6: Beat time-adjusted BTC buy-and-hold Sharpe

Gate G6 is the hardest. It requires the strategy to beat BTC’s passive buy-and-hold return, adjusted for the fraction of time the strategy is actually in the market. In a secular bull market, BTC buy-and-hold posts a Sharpe around 0.99. Any strategy that fails to meaningfully outperform that benchmark isn’t worth deploying.

Here’s what happened when the three strategy candidates met the gates.

V3 — Donchian Channel Breakout (DC-V3)

V3 was an 81-combination grid search across Donchian Channel periods, ADX filters, stop-loss multiples, and trailing stop parameters — all run on 4-hour BTC/USDC data. Zero combinations passed all six gates.

The best single-combination result (using DC(20), ADX filter 25, stop 2.0×ATR) posted:

• Sharpe: 0.296 — failed (gate: ≥ 0.70)
• BTC B&H Sharpe: 0.957 — not beaten (gate: failed)
• Max drawdown: −6.1% — passed
• Total return: +3.78% — passed

Root cause: Donchian(20) on 4-hour bars uses an 80-hour lookback — too short for BTC’s typical trending cycles. This produces frequent false breakout entries, and the resulting fee drag (0.1% per leg on Binance Spot) overwhelms the signal. Even when breakouts were real, exits were choppy, and the strategy could never generate the Sharpe premium needed to beat passive B&H.

The structural problem: any long-only BTC strategy in a secular bull competes against a BTC B&H Sharpe of nearly 1.0. Unless your strategy can capture upside while also sidestepping downside — which long-only Donchian cannot do — you can’t clear G6. The 4-hour V3 configuration was rejected. A daily-bar variant (V3.1) remains queued — longer lookback, fewer false breakouts, better chance of clearing the B&H ceiling. V3 is not dead.

V4 — BTC/ETH Ratio Statistical Arbitrage

V4 was more ambitious: trade the ratio between BTC and ETH as a mean-reverting spread. If the ratio drifts too high, short BTC and long ETH. If it drifts too low, the opposite. The prerequisite is mathematical: the ratio must be cointegrated — statistically proven to revert to a stable mean.

V4 didn’t even reach a backtest. It was killed by its pre-gates.

Key ran an Augmented Dickey-Fuller (ADF) test on three years of daily BTC/ETH ratio data. The ADF test checks whether a time series has a unit root — whether it wanders randomly rather than reverting to a mean. The result: p-value = 0.6202. To reject the null hypothesis of a unit root requires p < 0.05. At 0.62, the ratio is emphatically non-stationary. It has no stable mean to revert to.

The Ornstein-Uhlenbeck half-life confirmed it: 299.5 days. A mean-reversion strategy on a spread that takes 300 days to halfway-revert is not tradeable. You’d be holding positions for months waiting for a signal that might never come.

The root cause wasn’t technical — it was structural. BTC dominance expanded from roughly 46% to 63% over the three-year window, driven by the ETF narrative and institutional BTC flows. The BTC/ETH ratio didn’t oscillate around a stable mean; it trended upward. No amount of parameter tuning changes that.

V4 was abandoned before a backtest engine was written. That’s the best possible outcome for a bad idea.

V5 — BTC Z-Score Mean Reversion (KMR-V5)

V5 was the most sophisticated of the three — and the most instructive failure. The thesis: when BTC’s price drops to a Z-score between −1.5 and −2.5 below its rolling mean (entry thresholds tested: −1.5, −2.0, −2.5), it tends to bounce back. Buy the dip, take profit on the recovery.

Key built a full backtest engine with ATR-based stops, ADX regime filters, partial exit rules, trailing stops, and a circuit breaker. Then ran a 108-combination grid search across every meaningful parameter combination. Zero combinations passed all six gates. The best result cleared only three.

The Sharpe on the best combo: −1.115. Against a gate of ≥ 0.70. The strategy was actively losing money risk-adjusted — worse than simply holding cash.

The root cause was subtle and worth understanding. The strategy had a real edge: a 54% win rate, meaning most trades did recover. The problem was payoff asymmetry. Wins were small — the partial exit triggered at Z = −0.5, capturing a fraction of the rebound. Losses were large — the hard stop fired at 2.5× ATR during extended bear drifts that ADX classified as “non-trending” (ADX < 25) but that nonetheless kept going lower for weeks.

Slow, gradual bear drifts look like non-trending markets to a momentum filter. ADX stays below 25 because the trend isn’t steep. But when BTC drips down 0.3% per day for 30 days, the Z-score dips are not temporary — they’re a new floor. Buying those dips with a hard stop is a reliable way to lose.

The payoff structure — many small wins, occasional large losses — guaranteed a negative long-run expectancy. This is structural. Tuning parameters doesn’t fix it.

What’s Still Alive: V2 KSQ-V2

V2 — the BB Squeeze Breakout strategy — is the exception. It passed all six acceptance gates in backtesting (Sharpe 0.859, drawdown within limits, all cycle metrics cleared) and is currently in 30 days of paper trading. V2 holds cash when there’s no confirmed breakout, which sidesteps some of the long-only ceiling that killed V3.

Paper trading results will determine whether V2 earns a live deploy decision. Until then, it’s the only Key strategy on active evaluation.

Why the Kill Pipeline Is the Advantage

What Key demonstrated in a single sprint is something most individual traders never achieve: the ability to kill ideas faster than they can lose money on them.

V3 died because of a structural ceiling identified in the backtest data. V4 died because a statistics test failed — before any backtest was even run. V5 died because 108 parameter combinations all failed the same way, pointing to a structural payoff problem that no parameter can fix.

In each case, the failure was documented precisely. Gate results were recorded. Root causes were written up. The next strategy started from a richer understanding of what doesn’t work — and why.

That accumulated understanding is now the foundation for V6 (KBTR-V6), Key’s next candidate: a daily-bar EMA trend rotation strategy designed specifically to invert V5’s payoff profile. Hold BTC in bull regimes, hold USDC in bear regimes. No Z-scores, no hard stops, no dip fading. Just trend following.

V6’s backtest is in active development. If it passes the gates, it will join V2 in paper trading. If it doesn’t, it will be killed the same way V3, V4, and V5 were — quickly, with documentation, and without sentiment.

That’s how you build a strategy worth betting on.