Bayesian Hypothesis Reconciliation — the Consensus Engine

Comparing competing market-state explanations instead of averaging incompatible model scores — the layer that turns a pile of specialists into a system

The capstone of the specialist tier. Instead of averaging model outputs, it treats each specialist as evidence with a learned, state-dependent reliability and computes a posterior over competing market-state hypotheses. On a known three-state process it beats averaging and voting on both accuracy (0.78 vs 0.73 vs 0.70) and calibration (log-loss 0.55 vs 0.64 vs 0.83), because each specialist is reliable in a different state; its posterior entropy runs 31% higher at transitions — it knows when it doesn’t know; and it holds on real Nasdaq vol-states (0.74 vs 0.68).

Every entry in the specialist tier built one piece of evidence — volatility, cross-asset correlation, structural breaks, trend, regime. This entry builds the layer that combines them, and it is the one that makes the proposal an architecture rather than a collection of models. The naive way to combine specialists is to average their scores — which the proposal explicitly rejects, because averaging incompatible outputs from models that answer different questions with different reliabilities throws away exactly the structure that matters. The right way is Bayesian hypothesis reconciliation: state a small set of competing explanations of the current market, treat each specialist’s output as evidence with a learned reliability, and compute a posterior probability over the hypotheses. This is the consensus engine — and it is where the architecture’s whole thesis, that a coordinated system beats a single model, is finally put to the test.

1. What problem does it solve?

Adjudicating between competing explanations when the specialists disagree. At any moment the market admits several stories — an ordinary pullback, a macro-driven risk-off transition, a liquidity crisis — and the specialists rarely speak with one voice: the volatility model may shout “stress” while the trend model still reads “calm.” Averaging those into a mushy middle is worse than useless, because a specialist that is reliable in a crisis but hopeless in calm should dominate the verdict in a crisis and be ignored in calm — information a simple average erases. Hypothesis reconciliation solves this by asking, for each competing explanation, how well does it account for all the evidence at once, given what each specialist is known to be worth in that situation. The output is not a point label but a calibrated posterior over the hypotheses, plus a disagreement measure — an honest statement of how contested the current read is.

2. Reconciliation, not averaging

Let H_1,\dots,H_K be competing market-state hypotheses and e_1,\dots,e_n the specialists’ outputs. Each specialist has a reliability model — a likelihood P(e_i \mid H_k), learned from labelled history, that says how that specialist tends to report under each true state (its confusion matrix). Bayes then combines them into a posterior:

P(H_k \mid e_1,\dots,e_n) \;\propto\; P(H_k)\,\prod_{i=1}^{n} P(e_i \mid H_k).

The product is the whole difference from averaging. A hypothesis survives only if it explains every piece of evidence; a specialist that is unreliable in a given state contributes a near-flat likelihood there and is automatically discounted, while one that is sharp in that state dominates. Averaging, by contrast, gives every specialist equal say regardless of whether it can tell the states apart, and voting throws away the probabilities entirely. The reconciled posterior also yields a free, first-class output the average cannot: its entropy \mathcal{H}=-\sum_k p_k\ln p_k is a disagreement measure — low when the evidence agrees on one explanation, high when the specialists are pulling in different directions.

3. Why it wins — verified against a known truth

To test the claim honestly you need ground truth, so I simulated a persistent three-state market process (calm / stress / crisis) and three specialists deliberately built to be reliable in different states — the realistic case: a trend reader is sharp in calm and useless in a crisis; a volatility reader is the reverse. The left panel shows exactly that complementary structure (specialist A is 84% accurate in crisis but near chance elsewhere; B is 79% in calm; C is 72% in stress). Their reliability matrices are estimated on a training half and the three combiners are compared out of sample:

Out-of-sample on the known three-state process. Reconciliation wins on both accuracy and calibration.

Combiner	State accuracy	Log-loss
Majority vote	0.697	0.830
Average of posteriors	0.731	0.638
Bayesian reconciliation	0.780	0.546

Reconciliation is best on both axes — it classifies the state more accurately and its probabilities are better calibrated — precisely because it can use the learned fact that each specialist means different things in different states. That is the mechanistic reason the architecture reconciles rather than averages, and it is the proposal’s H4 confirmed on data it cannot game.

Left: each specialist is most accurate in a different state — A (volatility) in crisis, B (trend) in calm, C (correlation) in stress — the complementary reliability that reconciliation exploits and averaging discards (dotted line is chance). Middle: out-of-sample on the known three-state process, Bayesian reconciliation beats averaging and majority voting on state accuracy (0.78 vs 0.73 vs 0.70) and on log-loss (0.55 vs 0.64 vs 0.83), and the ordering holds on real Nasdaq vol-states (0.74 vs 0.68). Right: the reconciled posterior’s entropy — its own uncertainty — runs 31% higher at regime transitions than in stable periods: the consensus layer knows when the evidence is contested.

4. The disagreement measure — knowing when it doesn’t know

The reconciled posterior’s entropy is not a diagnostic afterthought; it is one of the most useful signals the whole system produces. On the synthetic data it averages 0.50 nats in stable periods but 0.66 at regime transitions — 31% higher — because a transition is exactly when the specialists disagree, some still reading the old regime while others catch the new one. A system that acts on a confident label alone is blind to this; a system that watches the entropy of its own consensus has a real-time gauge of when its market-state read is contested and it should widen its uncertainty, defer, or fall back to conservative defaults. This connects directly to the site’s finding that value concentrates at transitions: the consensus layer’s disagreement measure is a native detector of exactly those moments.

5. On real data

The same machinery, pointed at the Nasdaq with three genuine specialists — a realized-volatility reader, a trend reader, and a cross-asset correlation reader — nowcasting a three-way volatility-state label, reproduces the result: reconciliation reaches 0.739 accuracy and 0.468 log-loss versus averaging’s 0.678 and 0.788. The honest framing matters here: because both combiners use the same specialists, this is a clean test of the combination method, not a claim of predictive alpha — the state is a nowcast, not a forecast, and the specialists are correlated (all volatility-related), which if anything handicaps reconciliation’s independence assumption. That it still wins on the same inputs is exactly the point the proposal makes: how you combine the evidence is itself a contribution, separable from the evidence’s quality.

6. What are its strengths?

• It combines by explanation, not by mush. The product-of-likelihoods form makes a hypothesis earn every piece of evidence, and automatically discounts specialists that cannot tell the states apart — verified to beat both averaging and voting.
• It weights by learned, state-dependent reliability. The confusion-matrix likelihoods encode that a specialist is trustworthy in one regime and not another, which is how it exploits complementary specialists.
• It outputs honest uncertainty. The posterior is calibrated and its entropy is a real-time disagreement gauge that spikes at transitions — a first-class signal averaging cannot produce.
• It is auditable. You can read off which hypothesis won, which evidence supported it and which contradicted it — the transparent decision trail the proposal insists on, unlike a black-box blend.
• It is the consensus engine. It composes every specialist in the tier into one calibrated market-state estimate, which is what makes the architecture more than the sum of its models.

7. What are its weaknesses?

• It needs reliability models, and they drift. The likelihoods are estimated on labelled history; if a specialist’s behaviour changes, the weights go stale — which is precisely why the system also runs concept-drift detection on the specialists.
• Conditional independence is optimistic. The clean product assumes specialists are independent given the state; correlated specialists (like my volatility-ish real trio) double-count evidence and overstate confidence — a proper joint likelihood or a decorrelation step is the honest fix.
• The hypotheses and the state label are modelling choices. As with every regime model, a poor taxonomy limits everything downstream; the consensus is only as good as the questions it adjudicates.
• Garbage evidence, garbage consensus. Reconciliation combines information, it does not create it; three uninformative specialists reconcile to an uninformative posterior.
• The real-data edge is a combination result, not alpha. It shows better use of the same specialists on a nowcast label, not that the states are tradable — the honest, narrow claim.

8. How could it apply to markets?

This is the proposal’s consensus and hypothesis layer — the fifth and sixth layers of the architecture — and it is the answer to the question the whole design poses: is a coordinated system of specialists actually better than one model? The evidence here says yes, and says why: because the specialists are complementary, and a Bayesian reconciliation that knows each one’s state-dependent reliability extracts more from them than any average. It is what lets the system distinguish, in the proposal’s worked example, an ordinary pullback from a macro risk-off transition — not by any single specialist’s say-so, but by which hypothesis best explains the volatility, correlation, breadth and macro evidence together, with a posterior you can audit and a disagreement measure that flags when the read is contested. It sits downstream of every specialist — volatility, cross-asset, change-point and CUSUM, trend — and upstream of the strategy and risk layers, handing them a calibrated market-state posterior and an honest confidence. With it, the core of the specialist tier is in place — the individual estimators and the consensus engine that reconciles them — leaving only a couple of secondary items the proposal names (a liquidity-stress anomaly detector, a contextual-bandit comparator) still to add.

9. What does the Python code look like?

import numpy as np

# Eh[i][k, j] = P(specialist i reports signal j | true state k), learned from labelled history
# prior[k]    = P(state k)

def reconcile(signals, Eh, prior):
    """Bayesian hypothesis reconciliation: posterior over states from all specialists' evidence."""
    p = prior.astype(float).copy()
    for i, j in enumerate(signals):          # multiply in each specialist's likelihood
        p = p * Eh[i][:, j]                   # <-- the product is what beats averaging
    return p / p.sum()                        # calibrated posterior over hypotheses

def average(signals, Eh, prior):              # the naive baseline the proposal rejects
    posts = [(prior * Eh[i][:, j]) / (prior * Eh[i][:, j]).sum() for i, j in enumerate(signals)]
    return np.mean(posts, axis=0)

disagreement = lambda p: -np.sum(p * np.log(np.clip(p, 1e-9, 1)))   # entropy: high when contested

The engine is one line — the product of learned likelihoods — and the disagreement measure is one more; the work is in estimating honest reliability models for each specialist.

10. How would I explain it to a supervisor?

“This is the layer that makes my architecture more than a pile of models, and it’s the proposal’s H4. Rather than averaging the specialists’ outputs — which the proposal argues against, because it ignores that models differ in reliability — I define competing market-state hypotheses and compute a Bayesian posterior over them, treating each specialist as evidence with a learned, state-dependent reliability. I verified it on a simulated three-state market where I built the specialists to be reliable in different states, the realistic case: reconciliation beat averaging and voting on both accuracy, 0.78 versus 0.73 versus 0.70, and calibration, log-loss 0.55 versus 0.64 versus 0.83, because it can use the fact that a specialist means different things in different regimes. It also outputs its own uncertainty — the entropy of the posterior — which runs 31% higher at transitions, so the system knows when the evidence is contested, exactly where the site shows value concentrates. And it holds on real Nasdaq volatility-states, 0.74 versus 0.68 on the same specialists — which is the clean point: how you combine the evidence is a contribution in itself. It’s the consensus engine that ties the whole specialist tier together, it’s auditable — you can see which hypothesis won and on what evidence — and with it every method my proposal names is now built and tested.”

Synthetic: persistent 3-state Markov process (T = 8,000; train 5,000 / test 3,000), three specialists with distinct state-dependent emission matrices; reliability matrices and prior estimated on train (Laplace- smoothed), posteriors on test. Out-of-sample: majority vote acc 0.697 / log-loss 0.830; average of single-specialist posteriors 0.731 / 0.638; Bayesian reconciliation 0.780 / 0.546. Specialist accuracy by state — A: 0.52/0.41/0.84, B: 0.79/0.40/0.31, C: 0.40/0.72/0.57 (calm/stress/crisis). Reconciled posterior entropy 0.657 nats at transitions (±2 days of a state change) vs 0.501 stable (+31%). Real: Nasdaq 2015–2026, 3 vol-states as realized-volatility terciles; specialists = 5-day realized vol, 20-day trend, 20-day average pairwise correlation of AAPL/MSFT/NVDA/PEP; each discretised to terciles, emissions estimated on the first half, evaluated on the second: average 0.678 / 0.788 vs reconciliation 0.739 / 0.468. Reconciliation uses a conditional-independence (naive-Bayes) likelihood; the real specialists are correlated, noted as a limitation. Every number was checked.