The One-Way Alarm: Why Our 'Critical' Signal Has Never Fired in a Bull Market

The Caution Score on the Alphamancy dashboard is meant to do one job: warn you when the current market regime is about to flip. It blends three statistical inputs into a single 0-100 number, and we label the top tier "Critical — 98% flip within 20 days." That label is technically accurate. It is also, in a way you should know about, deeply misleading.

Thirty years of data hide a one-sided alarm.

What we built

The Caution Score combines three independent statistical edges into a single number using inverse-variance weighting:

Combo probability — which exact set of macro indicators is diverging from the prevailing regime, and how often that specific combination has historically resolved into a flip within 20 days.

Score momentum — how far the composite Alphameter has dropped from its peak inside the current episode.

Episode clock — how deep into the current regime we are, ranked against the historical survival curve for that regime.

These three estimates get fused into a single value from 0 to 100. The thresholds on the dashboard — 40 (Moderate), 50 (Elevated), 60 (Critical) — were chosen because, in 30 years of data, every time the score crossed them, the regime flipped within 20 days 67%, 83%, and 98% of the time respectively. We re-validated those numbers across three rolling decades. They held.

Then we split the data by direction.

What the data actually says

Critical (≥60) instances in 30 years of cross-asset data:

During risk-on episodes (warning of an imminent flip to risk-off): zero.

During risk-off episodes (warning of an imminent flip to risk-on): fifty-five.

Not "rare." Zero. In thirty years, the Caution Score has never reached its top tier while the market was in a risk-on regime. The "98% flip" alarm is a one-way alarm. It only screams when fear is breaking.

The ≥50 "Elevated" tier is far more symmetric — 136 risk-on instances vs 169 risk-off, with similar 20-day flip precision (87% from risk-on, 79% from risk-off). The ≥40 tier splits 829 / 456 with comparable accuracy on both sides. The asymmetry breaks sharply at the top.

Why this happens

The mechanism lives in the survival curves of the two regimes.

Risk-off episodes are short and violent. The median lasts under 30 days. By day 40, more than 70% have resolved. Indicators tend to diverge late, all together, because panic compresses the timeline — and when they do diverge while the episode clock is already deep, every input to the score lights up simultaneously. The combined number tips past 60 with ease.

Risk-on episodes are long and grinding. The median runs well over 70 days; many last more than a year. Indicators diverge slowly, one at a time, over months. The composite Alphameter score can erode 40-60 points before a flip, but it erodes on a timescale where the episode-clock input never gets extreme enough to push the combined score over 60. By the time the math would say "Critical," the regime has usually already flipped underneath it.

In behavioural shorthand: fear breaks fast. Greed breaks slow.

What to do about it

Three practical implications for anyone using the score live.

If you are positioned for risk-on — long equities, commodities, growth, crypto — and you are waiting for the "Critical" alarm before reducing exposure: stop waiting. It is not coming. Use the ≥50 Elevated tier as your top operational threshold. It has fired 136 times during risk-on episodes since 1996, and the regime flipped within 20 days 87% of those times. That is the real warning bell on the upside.

If you are positioned for risk-off — long bonds, gold, defensives, cash — and you see the Caution Score push toward 60, that is exactly what it was built for. Historical precision at that tier runs 95-100% across every decade we tested. Take the signal.

For everyone: the asymmetry itself is information. A market that signals exhaustion cleanly from one side and ambiguously from the other is a market where downside risk decompresses faster than upside risk does. That fact alone should change how you size into a rally-from-the-lows versus a rally-from-an-already-high.

Closing

The Caution Score was built to be honest about uncertainty. The validation that exposed this asymmetry was a face-off against three different versions of the formula, run on the entire 1996-2026 cross-asset dataset, with the methodology published openly on the learn page. We did not expect to find a one-sided alarm. We found one.

The more interesting question is no longer "does the score work?" — it does, in both directions, at the lower tiers, with stable calibration across three decades — but "what does it tell us about markets that it only screams from one side?"

We will update the dashboard labels in the next release to surface this distinction. The score will keep doing its job. Just know which job that is.