VIX State vs VIX Weight: The Signal the Composite Throws Away

Alphameter v3 downweighted VIX from 2.0× to 0.5×. That was correct — VIX is a lagging indicator, and adding more of it makes the composite a worse predictor, not a better one. But the binary VIX = − state still carries information about forward returns. The engine is doing the right thing for regime classification; it just isn't doing the only thing the data supports.

The first part of that argument is the boring part. The interesting part is what comes after.

The lagging-indicator problem

VIX is the implied volatility of one-month S&P 500 options. It is not a forecast — it is a measurement of how much hedging demand exists right now. When the market is already falling, hedges get bid, and VIX goes up. When the dust settles and traders stop paying for protection, VIX falls.

That makes it useful for describing the present and useless for predicting the future. By the time VIX has confirmed a risk-off environment, the move that caused the confirmation has already happened. Building a composite score around a lagging indicator is the canonical way to fit yesterday's regime instead of tomorrow's.

The grid search behind Alphameter v3 caught this. We tested 30 years of cross-asset data (1996-2026), with the pre-2020 period as training and 2020-2026 as holdout, against a basket of regime-sensitive pairs: SPY/TLT, QQQ/GLD, IWM/SHY, CPER/SLV, AUD/CHF. The VIX weight that maximised cross-asset signal separation was 0.5×. The forward-looking indicators — AUD/JPY, Copper/Gold, Bond Yields, DXY, Sector Rotation — all ended up at 1.5×.

VIX did not get downweighted because it is noisy. It got downweighted because, even when correctly read, it tells you what already happened.

The dropout test

Once we had v3 in production, the natural follow-up question was: if VIX is contributing so little, what happens if we drop it entirely?

The test is straightforward. For each indicator, recompute a leave-one-out composite score across all 7,617 daily snapshots, and measure the Pearson correlation between that LOO score and the next 20-day forward return for each tradeable asset. The change in information coefficient tells you whether the dropped indicator was adding predictive signal or diluting it.

The result is almost embarrassingly clean. With every indicator present, the Nasdaq IC is 0.055. Drop VIX, it falls to 0.047 — a shift of 0.008 in the third decimal. Bitcoin: 0.133 with VIX, 0.129 without — a shift of 0.004. EURUSD: 0.067 → 0.065. USOIL: 0.128 → 0.126.

For comparison, dropping DXY costs USOIL twice as much IC (-0.029). Dropping Sector Rotation costs Bitcoin -0.028. Dropping Copper/Gold costs USOIL -0.028. Those are real losses. VIX is not.

So the question that comes next is the obvious one. If dropping VIX barely changes the predictive power of the composite, why keep it at all?

The filter that the composite throws away

The answer is that the score is not the only thing you can build out of an indicator. The score asks one question — "is the cross-asset environment in risk-on, neutral, or risk-off territory?" — and answers it by averaging six weighted contributions. That averaging is where the information loss happens.

VIX has only three states: -1 (low / complacent), 0 (mixed), +1 (high / panicked). The composite combines all three with the other five indicators and produces a single number. But the state itself — particularly VIX = -1 — turns out to be a strong filter on forward returns in a way the averaged score cannot express.

Here is the gating test. Take each composite-score bucket — xRO (≥50), RO (≥25), NEU, RF (≤-25), xRF (≤-50) — and split it on VIX state. For Nasdaq:

In the xRF bucket (composite ≤ -50, the deepest risk-off readings), the all-state mean 20d return is +0.28%, annualised Sharpe 0.13. That is roughly cash. Now apply a VIX = − filter: 101 of the 795 xRF days had a low VIX reading. Their mean 20d return is +4.32%. Annualised Sharpe: 1.29. The remaining 694 days, with VIX neutral or high, return -0.31% with Sharpe -0.15.

The RF bucket shows the same pattern. All-state: +0.57%, Sharpe 0.27. VIX = − only: +4.11%, Sharpe 1.30. VIX ≠ −: +0.19%, Sharpe 0.10.

This is not just a deep-risk-off phenomenon. Every Nasdaq bucket gets better with VIX = − as a filter. In xRO, the mean goes from +1.23% to +3.23%. In RO, from +1.45% to +4.89%. The Sharpes triple or more.

Bitcoin is even more dramatic. In xRF, the all-state 20d return is +0.46% with Sharpe 0.11. Filter on VIX = −, and you get +10.08% with Sharpe 2.78 — across 17 historical episodes. In xRO, Bitcoin goes from +7.71% to +18.27%.

The composite cannot express this. A bucket-and-filter rule throws away enormous swathes of history (you are conditioning on a specific state, not averaging over states) but the conditional edge is what a discretionary trader actually cares about.

Pairs, where the Sharpes get serious

The single-filter version is already useful. The two-filter version is where the numbers start looking like research-paper results. Fix VIX at -1 and pair it with the state of one other indicator:

VIX = − × DXY = + (a weakening dollar setup) on Nasdaq: 136 days, +4.50% mean 20d, train Sharpe 2.08, test Sharpe 1.91.

VIX = − × Copper/Gold = + (a recovering growth signal) on Nasdaq: 97 days, +4.67% mean, train Sharpe 1.76, test Sharpe 4.04.

VIX = − × Copper/Gold = + on EUR/USD: 72 days, +1.56% mean, train Sharpe 2.82, test Sharpe 2.30.

The test-side Sharpes survive holdout. That is the part that usually doesn't happen. A pair that trains to a Sharpe of 1.76 and tests at 4.04 is either a real edge or a small-sample artefact, but 44 separate cells across the asset universe show statistically distinguishable behaviour, and the surviving sign-agreement filter (requiring the train and test means to point the same direction with at least 5 observations each) eliminates the obvious noise.

The interpretation is straightforward. A low VIX reading paired with one other indicator already pointing constructive is the difference between a setup with edge and a setup without. VIX = − by itself is interesting. VIX = − inside a particular cross-asset configuration is a high-conviction tactical signal.

The asymmetry, and what it tells you

The VIX = − filter does not work on every asset. On Gold, it slightly hurts in every directional bucket. On 10Y Treasuries (TLT), it hurts in risk-on and helps in risk-off, with the magnitudes roughly cancelling. The filter is sharp where it works and quietly negative where it doesn't.

That asymmetry is informative, not a bug. Gold is a structural trade, not a sentiment trade. Its drivers are real rates, central bank flows, and reserve-asset positioning. The presence or absence of equity-option hedging demand has no causal connection to the gold price. When the model has nothing to say, it should say nothing — and the negative Sharpes confirm that the filter has no useful information in that domain, rather than spuriously pretending it does.

The signals where VIX = − rewards exposure — Nasdaq, Bitcoin, oil (in the right setup), EUR/USD — all share the same structure: they are risk assets whose forward returns depend partly on sentiment normalisation. A low VIX inside a stressed macro means the option market is no longer paying for protection. That is consistent with the marginal hedger being tapped out, which is consistent with imminent mean reversion in the underlying.

The honest caveat

There is a trap inside this finding, and it would be dishonest to leave it out.

The signal "low VIX while macro is panicking" shows up in two opposite contexts. In context A — the crisis bottom — VIX has already spiked and is rolling over. The composite is deep risk-off because the macro indicators are slow to recover. Q1 2009, Q2 2020, and Q4 2022 all fit this pattern. Risk assets bounce hard from these setups.

In context B — the complacent pre-crash — VIX never went up while the macro deteriorated. The composite is risk-off because Copper/Gold, AUD/JPY, and bonds are flagging, but the equity option market has not woken up. The September-November 2008 window is the textbook example, as is February 2020. Risk assets get destroyed from these setups.

Both contexts produce the same six-tuple: composite risk-off, VIX = -1. Looking at only the current state, you cannot tell them apart. The headline edge — the +4.32% mean in Nasdaq xRF + VIX = − — is the blend of these two contexts, and the dispersion within that mean is much wider than the dispersion across the bucket as a whole.

The fix is a separator. A 30-day trailing-max VIX feature would do most of the work: if VIX rallied to 35+ in the last month and is now back at 18, you are in context A. If VIX has been pinned below 20 the whole time, you are in context B. We do not run this feature in production today. It is on the roadmap and will eventually be added to the Tactical Overlay as a context disambiguator.

Until then, the honest summary is: the edge exists, the asymmetric drivers are real, but the headline number is a mixture distribution. Trade it knowing that.

Where this lives in the product

The composite still does its job. It is a six-indicator weighted average designed to classify the cross-asset environment as risk-on, neutral, or risk-off. For that classification task, the v3 weights are optimal and VIX at 0.5× is correct.

What the score throws away is what the Tactical Overlay surfaces. The Overlay panel on the dashboard takes the current six-tuple — including the present VIX state — and ranks historical episodes that matched. When the live signature is xRF + VIX = − and you see the top-ranked historical winner is Nasdaq with a 4%+ historical edge, you are looking at the same finding this post just walked through, applied to today's setup.

The composite answers "what regime are we in?" The Overlay answers "given exactly this configuration, what has historically worked?" Those are different questions, and the difference is exactly the information the averaging step throws out.

What to take away

VIX is a lagging indicator. Downweighting it in the composite was correct. Dropping it from the composite would barely change the score's predictive power.

VIX state, considered separately from the score, contains forward-return information that the composite cannot express. The edge is concentrated in risk-on and risk-off conditional buckets, on risk assets, and intensifies when paired with a constructive read on a second indicator.

The right model is not "VIX should be weighted more." The right model is "VIX should be weighted less in the composite and more in the conditional filter." We have the first half running in production. The second half lives in the Tactical Overlay — and the honest caveat about the Type A / Type B trap lives in this post.

That is how we think a quantitative dashboard should evolve. Not by adding to the score, but by adding the right structures around the score.