Why conviction is a risk factor

Classical theory predicts that strong directional consensus stabilizes markets. Twenty years of multi-asset data shows the opposite — and four independent statistical specifications agree.

The classical prediction

Most factor-stability theory carries an implicit assumption: when a market converges on strong directional consensus — everyone leaning the same way, the dominant factor expressing itself cleanly — volatility should fall. The reasoning is intuitive. Strong agreement removes ambiguity. Removed ambiguity reduces the rate at which prices need to adjust. Reduced adjustment rate is, definitionally, lower realized volatility.

This is the prediction that comes out of standard equilibrium models: as the order parameter of factor coherence rises, the residual fluctuation rate around the dominant direction falls toward zero. It is the same shape as the prediction in many adjacent fields. It is also the shape most allocators carry, often without articulating it: we lean into conviction because we expect conviction to be the calmer state.

What we found instead

We tested the classical prediction on twenty years of daily closing prices from five major US-tradeable assets (SPY, QQQ, IWM, TLT, GLD; 2005–2024; N = 4,979 sessions). The setup is direct: estimate a daily measure of directional-consensus magnitude, estimate a daily measure of price-level dispersion at the position-boundary, and regress one on the other in log space. The classical prediction is a negative slope. We found the opposite, robustly, under four independent specifications.

The reading is unambiguous: when directional consensus rises, dispersion at the position-boundary rises with it. The classical prediction was inverted. Under every reasonable robustness check we could think to apply, the relationship stayed positive.

The story underneath the number

The surprise dissolves once the right mechanism is named. Strong consensus is not a calm state. It is a positioning state. When everyone leans the same way, the marginal trader entering the trade is doing so with progressively thinner edge against a progressively crowded book. The boundary between "I am still long" and "I am being squeezed out" gets steeper. The smallest price move toward the unfavored direction triggers cascading position adjustments — not because new information arrived, but because the position structure itself becomes unstable.

Consensus is not a stability signal. It is a measure of how thin the boundary has become between holding and unwinding.

That mechanism predicts exactly what the data shows: as consensus magnitude rises, dispersion rises with it — not because the dominant direction is uncertain, but because the residual position-adjustment rate is louder. The classical prediction missed this because it modelled price as a free particle exploring a smooth landscape, not as an aggregate of crowded positions on a sloped surface.

What this means for portfolio construction

Three concrete implications for desks running a primary quant process:

• Stop reading conviction as a risk-off signal. A high directional-consensus reading historically gets interpreted as "the trade is well-defined, lean in." On this evidence, that interpretation is exactly backwards. Lean in only when conviction is rising from a cold start. Lean out when conviction has been high for a stretch — that is the regime where the position-boundary is most likely to break.
• Treat extreme conviction as a tail-risk regime. The S4 specification — restricted to the top-30% consensus magnitude windows — carries the second-strongest |t|. The effect is not a small-sample artifact. It is concentrated in exactly the regime where most desks reduce their hedge ratios.
• Hedge boundaries, not centres. If position-boundary instability is the proximate cause, the right hedge structure is one that pays off when boundary positions get squeezed — OTM protective options, far-strike collars, vol-of-vol exposures — not flat delta-neutral hedges centred on the dominant direction.

What this is not

• It is not a market-timing signal. The relationship is contemporaneous, not predictive in the look-ahead sense. We are not claiming you can forecast next-week dispersion from this-week consensus; we are claiming you should not interpret high-consensus readings as a calmer regime than they actually produce.
• It is not a trade signal in isolation. Without an entry rule, a sizing rule, and a stop, the empirical relationship is just an observation. Quark's signal stack uses the consensus reading as one of eighteen inputs to a composite risk read; the article is not the trade, the composite is.
• It is not the only feature that matters. The N = 4,979-day sample contains many other regimes that need their own characterizations — we are reporting one robust contemporaneous relationship, not a complete model of price formation.
• It is not, despite the size of the t-statistic, a license to over-fit. The pre-registered hypothesis has been frozen since registration and the four specifications are the four we committed to before the run. Adding a fifth specification post-hoc until something falls out would cross into the multiple-testing zone Harvey 2016 was written to flag.

Why we publish negative-direction findings

This is a finding that contradicts a widespread heuristic. Allocators who have leaned long-conviction during crowded regimes have, on average, paid for it. A research stack that finds something well-known to be wrong is more useful than a research stack that confirms the conventional read — the latter does not change anyone's allocation.

We use the finding internally. The signal contributes to the regime-stress composite that feeds our model-portfolio allocator. It also informs the hedge-sizing rule on the tail-risk module. The fact that it is now a known, named feature of how our system thinks about risk is exactly why we are publishing it: clients evaluating the signal feed should know what we are weighting and why.

Read the full paper

The long-form companion paper — with the four specifications written out, the eigen-decomposition definition of the order parameter, and the theoretical reconciliation between the canonical Goldstone-mode reading and the crowded-trade reading we actually observe — is available here:

Read the working paper →