Research · Essay
The Consensus Gap
Measuring the distance between what everyone expects and what we believe — and why distance is never enough.
By NDI Research ✦ ✦3 minutes’ reading
In Brief
- The Consensus Gap is the difference between our probability for an event and the best available external probability for the same event.
- A large gap is a reason to look, not a reason to act. Every gap must pass five admissibility tests before it influences exposure.
- Most of the gaps we measure are rejected. The discipline lies in the rejection rate.
Note added 14 August 2026: two of the largest admissible gaps in the July outlook, issued two days after this essay, resolved against us. The framework below is unchanged; how much we trust it is discussed in [Calibration Before Conviction](/research/calibration-before-conviction).
Every forecast we produce has a counterpart. For almost any event that matters to markets, someone else has already expressed a probability, either explicitly through a survey or implicitly through the price of an instrument that pays off when the event occurs. We call that probability external, and the difference between it and our own the Consensus Gap.
A Worked Example
Does a structural liquidity threshold break before the end of the quarter?
- Market
- 34%
- Ours
- 61%
- Consensus gap
- +27
A gap of +27 says that if both probabilities were prices, we would consider the market's contract materially cheap. It says nothing about whether we are right. A large gap is equally consistent with genuine insight and with a broken model, and in our experience the second explanation is more common than the first.
§ IWhy distance is not enough
A system that acted on every large gap would be a contrarian with extra steps. It would systematically overweight the situations in which its models are most wrong, because model failures produce large, confident disagreements with markets. The size of the gap would select for error.
We therefore treat the raw gap as a candidate, and require it to pass five tests before it is considered admissible. Each test can only reduce the gap, never increase it.
| Test | Question asked | Typical rejection share |
|---|---|---|
| Confidence | Is agent dispersion low enough that the pooled probability means something? | 31% |
| Diversity | Is the gap supported by more than one independent model family? | 22% |
| Liquidity | Can a position expressing the view be entered and exited at realistic size? | 14% |
| Downside | Is the loss in the state where we are wrong within the risk budget? | 9% |
| Execution | Does the expected edge survive estimated costs and market impact? | 11% |
§ IIThe admissible gap
Gaps that survive are scaled rather than simply passed through. The admissible gap shrinks the raw gap by a confidence factor derived from agent dispersion and a diversity factor derived from how many model families independently support it:
The effect is that a +27 gap supported by a single model with high dispersion may enter the allocator as +6, while a +15 gap supported by three model families with low dispersion may enter as +13. We would rather act on a modest disagreement we understand than a dramatic one we do not.
§ IIIGaps in both directions
Negative gaps are as useful as positive ones and are treated identically. A gap of −20 on a stress event means we consider the market's protection expensive. In practice, negative gaps on tail events are where we have been most cautious, because the cost of being wrong about a tail is asymmetric in a way that a probability alone does not capture.
§ IVSignal
Across all active questions, the sum of admissible gaps weighted by each question's relevance to the current portfolio is published on the system page as the non-default signal. It is a summary of how much, and in which direction, the system currently disagrees with the world. On most days it is small. It is supposed to be.