In the second week of April, the Quant Crypto Alpha programme reduced its gross digital-asset exposure by approximately forty per cent. Two and a half weeks later the spot market for BTC dropped seventeen per cent over three sessions.

We were not predicting a selloff. We do not predict selloffs. The volatility-regime filter inside the algorithm transitioned from a "normal" state to an "elevated risk" state, and the position-sizing logic responded mechanically. The April outcome was favourable; the rule that produced it has fired four times since 2022, and the outcome distribution is mixed.

What the filter measures

The regime filter is a simple ensemble. It takes four inputs:

Realised volatility over rolling five and twenty-session windows. When the five-day reading exceeds the twenty-day reading by more than a defined threshold, the short-term vol regime has shifted faster than the long-term distribution would predict, and the filter weights this as evidence of stress.

The ratio of realised vol to one-month implied vol. When realised begins to track or exceed implied, the options market is under-pricing the volatility actually being delivered. Historically a leading indicator for further realised-vol expansion.

The skew of recent return distributions. A negative skew that exceeds its trailing-year median by more than half a standard deviation has, in our live data, preceded most material drawdowns by between three and ten sessions.

The correlation of digital-asset returns to the dollar. Digital assets typically display a weakly negative or zero correlation to DXY in normal regimes. During stress, that correlation flips positive. The flip is one of the cleanest tells we have.

What triggered April

On 8 and 9 April, three of the four indicators moved beyond their respective thresholds within a thirty-six-hour window. The algorithm transitioned to the elevated-risk state on 9 April at 14:00 UTC. Target volatility was reduced from twelve per cent annualised to seven per cent annualised, which translated into the roughly forty per cent reduction in gross exposure.

Trading continued at the lower target through the subsequent drawdown. The smaller positions limited losses but, more importantly, kept the system inside its risk budget so that it could reposition cleanly when the regime indicator normalised on 7 May.

The lesson is not that we have a crystal ball. The lesson is that disciplined risk-on and risk-off rules will sometimes look smart, sometimes look conservative-to-the-point-of-missing-out, and over a full cycle preserve capital.

Three lessons we would generalise

First, single-indicator regime filters fail constantly. A four-indicator ensemble with a "two-of-four" trigger has a much better false-positive rate without sacrificing speed.

Second, the cost of a false positive, briefly under-positioned, is small. The cost of a false negative, fully levered into a cascade, is unacceptable. We accept the trade-off on those terms.

Third, the rule must be mechanical. A discretionary version of the same rule, applied by a human watching the same screens, would have hesitated. We know this because we have tested it.

The April episode worked out well. The next one might not. The system will respond identically either way.