Your model reports the signal. Bearing reads the substrate the inputs cannot.
The model is accurate within its substrate. Higher accuracy starts where the substrate widens, where the next compound configuration enters from outside the inputs.
Pod diversification compressed exactly when the diversification assumption needed to hold. The correlation surge is what the model's input substrate could not see.
In March 2026 the "March Malaise" hit the most dominant multistrategy platforms, Citadel, Millennium, Point72, Balyasny, D.E. Shaw, Marshall Wace, with synchronized drawdown despite pod-based diversification. Many pods carried exposure to the same macro factors even where their specific trades differed; under systemic stress, the correlations converged. Two months earlier the trend architecture had worked exactly as designed: CTAs delivered Q1 2026 returns led by energy at +6.25% on a clean Hormuz pivot, 91% short to 100% long crude through the late-February onset. The price regime, once it appeared in the substrate, was captured. The cross-pod correlation buildup, which never landed in the price-derived inputs, was not. That divergence is the whole read.
Source · HedgeCo March 2026 multistrategy correlation breakdown · Kpler CTA Q1 2026 performance analysis · SG Trend Index 2026 constituents
The executive read
Your model is right within its window. Widen the window, widen the accuracy.
Three accuracy layers
Empirically validated
Your model-architecture review is already holding this as the open question: not whether to replace the model, but whether the substrate it trains on extends to the compound configurations entering against your positions, and whether the accuracy gain from that extension reads as material at your governance threshold.
Your model's accuracy is conditional on the input substrate it consumes, price, vol, factor exposure, alternative-data feeds. Within the configuration regimes your training substrate captured, the model performs at the accuracy that substrate validated against. The question is not whether the model is right within its substrate; it is whether widening the substrate widens the accuracy.
Your Q1 2026 print already names where the substrate sits relative to the active configuration cycle. CTAs captured the Hormuz energy regime cleanly once it surfaced into the price-derived inputs. Then March 2026 hit the multistrategy pod platforms with synchronized drawdown, pod diversification compressed exactly when the diversification assumption needed to hold. The model's input substrate did not detect the cross-pod correlation buildup that produced the drawdown; the substrate widening would have.
Three layers of accuracy gain, each anchored against resolved long-record precedent. Layer 1, the model is already at full accuracy within its substrate; the methodology adds zero. Layer 2, accuracy degrades through cross-position correlation surge when compound configurations enter; the methodology delivers substrate the model consumes to detect the surge before it lands in the price-derived inputs (August 2007 quant crisis and the August 2024 yen carry unwind as resolved precedents). Layer 3, accuracy degrades through regime change the inputs structurally cannot detect; the methodology produces regime-detection lead time (1987 portfolio insurance and the March 2026 Malaise as the resolved and active demonstrations).
Four composable integration paths, feature engineering, regime detection, Bayesian prior, ensemble overlay, combine; institutional implementations layer two or three. Six named comparators across systematic CTAs (Man AHL, Winton, Aspect, Transtrend) and multistrategy quant (Citadel, Millennium) name who is building configuration-substrate-aware architecture and who is operating on pure price-derived signal. The gain is bounded, conditional, and grounded in the integration literature, not in PHM-owned backtests.
The accuracy ladder
Three layers of accuracy gain, and the honest zero at the bottom.
Magnitude bands
Conditional, anchored
The ladder is the whole methodology in one frame. Each rung names where the model stands, what the substrate adds, the resolved precedent it is anchored to, and the magnitude band the gain falls in, bounded against integration literature, not asserted against owned backtests.
Where the model is already right.
The configuration regimes the training substrate captured, trending price, stationary cross-sectional factors, bounded vol, stable carry. Backtest and live performance track. The architecture's accuracy is the architecture's contribution; the methodology has nothing to add here.
Demonstration · Q1 2026 CTA energy +6.25%The correlation the inputs report as coincidence.
Compound configurations drive cross-position correlation surge before it lands in the price-derived inputs, the diversification assumption fails before the model has the substrate to explain why. Substrate widening delivers the early warning 2–6 weeks ahead of the price-derived signal.
Resolved · Aug 2007 Khandani-Lo · Aug 2024 yen carryThe regime the substrate structurally cannot detect.
When the configuration substrate itself shifts, supply-chain geometry, monetary-system structure, geopolitical-financial coupling, the inputs read "normal" until the regime has already landed. Substrate at the platform-level risk altitude fires the regime-change signal 2–4 weeks before the synchronized drawdown.
Resolved · Oct 1987 portfolio insurance · Mar 2026 MalaiseLayer 1 is the honest zero: where the substrate already captures the regime, the methodology adds nothing. The gain begins where the next compound configuration enters from outside the inputs, and it is bounded, conditional, and anchored to resolved precedent rather than to PHM-owned backtests.
Where the model is already at full accuracy. The regime the training substrate captured.
The model performs at the accuracy the training substrate validated against, within the configuration regimes that substrate captured. Established factor structures, mean-reverting periods, well-priced vol, stable carry, trending price action that matches the signal definition. Your backtests cover this terrain; your live performance through it matches the backtest's expectation. The model is right here, and the methodology has nothing to add.
Q1 2026 carried clean substrate-validated terrain across most CTA exposures. The Hormuz compound produced a directional move in energy and base metals the trend architecture captured cleanly, positioning pivoted from 91% short to 100% long in crude between January and early February, with energy returning +6.25% (metals +1.28%, agriculture +1.01%). The substrate could see the price regime once it appeared in the substrate; the architecture worked exactly as designed.
This is where the methodology acknowledges zero contribution. If the architecture is sound and the input substrate captures the configuration regime, no additional substrate is needed. The value begins where the next compound configuration enters from outside the substrate, where the input architecture cannot detect what is moving against the positions until after the position has moved. The Q1 2026 print is the model in its full-accuracy regime; the next two layers are where the substrate gap appears.
(a) Hold the existing architecture and position-sizing discipline within configurations the training substrate validated against; the model's accuracy is the architecture's contribution and the methodology does not displace it. (b) Mark the boundary between substrate-validated regime and outside-substrate explicitly, per strategy, the boundary is named (trend-following: directional price moves in well-priced commodities; stat arb: stationary cross-sectional factor structures; vol-targeting: bounded vol regime). (c) Resist extending model output past the substrate boundary by inflating confidence.
Confirming the regime is operative, live performance tracks within ±0.3 Sharpe of in-sample; cross-strategy correlation stays inside rolling 60-day bounds; factor stability holds inside anticipated ranges. Confirming the regime is shifting outside substrate, correlation rises above rolling bounds without an identified driver in the inputs; live Sharpe degrades faster than transaction-cost drag explains; factor returns show autocorrelation the architecture treats as stationary. The gap between baseline and live performance is where Layer 2's substrate addition operates.
Source · Kpler CTA Q1 2026 performance analysis · SG Trend Index 2026 constituents · Man Group on trend-following market universe
The cross-position correlation the input substrate cannot detect.
The diversification assumption holds within the configuration regime the training substrate captured. Positions calibrated against per-position risk, factor exposure scaled against per-factor expectation, pod allocation built against inter-pod independence. Compound configurations entering against the positions drive cross-position correlation surge the input substrate reports as coincidence, because the underlying compound is not yet in the price-derived inputs. The surge lands, the diversification fails, the drawdown happens before the input architecture has the substrate to explain why.
This is the failure mode Khandani-Lo documented in the academic literature. Their core finding: crowding risk is regime-dependent, historical data from the normal period carries no information about the crisis correlation, because the crisis correlation does not exist in normal periods. The input substrate cannot detect what does not yet exist in it; the substrate widening is what closes the gap.
During the week of 6–10 August 2007, multiple quantitative long/short equity funds took unprecedented simultaneous losses despite ostensibly diversified portfolios. Khandani-Lo's NBER analysis documented two unwinds driven by coordinated deleveraging of similarly-constructed portfolios. The compound configuration, subprime stress + forced leverage unwind at one large fund + crowded factor exposure, drove cross-fund correlation surge the per-fund risk models read as coincidence until the unwind was already in motion. The input substrate could not detect crowding risk because crowding risk only exists in the regime where it has already activated. The funds that survived with minimal damage ran lower leverage and held genuine cash reserves, both substrate-independent defences.
The 5 August 2024 yen carry unwind carries the same architecture at scale. BIS analysis documented that hedge-fund sensitivity to carry-trade returns had turned positive and kept rising through that July, a buildup invisible to per-fund risk models in real time, visible only retrospectively in aggregated industry data. The unwind erased roughly $6.4 trillion in market value in a compressed window. Wellington's post-mortem named the cascade: investors short the yen sold their most appreciated asset, US momentum stocks, coupling USDJPY and US momentum at a correlation the historical substrate had read as low.
The March 2026 multistrategy correlation breakdown carries the same shape at the current cycle. The "March Malaise" hit the dominant platforms with synchronized drawdown despite pod-based diversification: under systemic stress correlations converge, and many pods carried exposure to the same macro factors even where their specific trades differed. The institutional response is the validation, risk-model recalibration across multi-strats as firms update correlation assumptions and stress-test scenarios.
(a) Integrate compound configuration substrate as an additional input feed in the factor model or risk-attribution architecture, the configuration vectors deliver cross-position correlation signal before the price-derived inputs surface the surge. (b) Use the substrate as cross-strategy correlation early warning at the firm-level risk function, for pod platforms at the Portfolio Construction / Risk Group altitude; for CTAs at the head-of-research altitude. (c) Test through the resolved precedent windows, replay August 2007 and August 2024 with the substrate present; the claim is detection 2–6 weeks before the unwind lands in the price-derived signal.
Producing accuracy gain, early warnings fire 2–6 weeks before live surges; reserve / leverage adjustments in the pre-warning window track the post-event price action; position sizing through compound regime change degrades less than the un-integrated comparator. Not landing, false-positive rate exceeds what the architecture can absorb without churn; the integration path does not match the architecture and produces inconsistent signal; the governance threshold on substrate-driven decisions fails to land at the PM seat. Magnitude, Sharpe improvements in the 0.1–0.3 range for well-integrated cross-asset substrate, conditional on approach and architecture fit.
Source · Khandani-Lo NBER WP 14465 · BIS Quarterly Review Sept 2024 · Wellington yen carry unwind analysis · HedgeCo March 2026 correlation breakdown
The regime change your input substrate cannot detect.
The input substrate reports the configuration regime through what those inputs measure. When the regime changes through a mechanism the inputs do not measure, supply-chain geometry shift, monetary-system structural change, geopolitical-financial re-pricing, the inputs continue to report "normal" until the regime has already shifted. The model keeps sizing positions and deploying leverage per its calibration. The regime lands; accuracy degrades through the transition. This is structurally different from Layer 2: at Layer 3 the configuration substrate itself has shifted, not just the correlation across positions within the existing substrate.
The dynamic-hedging strategies prevalent in 1987 replicated a put option through systematic selling into declines. The input substrate (price, realized vol) read the early-stage decline as a normal vol regime and executed the prescribed selling. The selling produced additional decline; the decline triggered additional model-driven selling. On 19 October 1987 the Dow fell 22.6% in a single day, the largest one-day percentage decline in US equity history. The dynamic-hedging models' own substrate compounded the regime change those models were architected to defend against. The substrate gap did not just fail to detect the regime change, it produced it. When a model's input substrate is determined by the model's own actions, the substrate cannot detect the regime the actions are creating; substrate widening at the configuration altitude is the structural fix.
The March 2026 Malaise is the current cycle's Layer 3 demonstration at scale. The pod-based architecture was built on an inter-pod independence assumption, different pods, different strategies, different markets, uncorrelated streams aggregating into stable platform performance. It had delivered for the better part of two decades. The compound configuration cluster of early 2026, Hormuz shifting energy and FX regimes, tariff changes shifting cross-border equity flows, AI-infrastructure dispersion shifting sector dynamics, drove cross-pod correlation surge against the platform-level assumption simultaneously. Citadel, Millennium, Point72, D.E. Shaw, Marshall Wace, Balyasny all hit synchronously. The risk-model recalibration that followed is the institutional acknowledgement that the substrate the platform-level risk function consumed was insufficient against compound configuration regime change.
The CTA-versus-multistrategy divergence is the methodology's empirical signal. CTAs that captured the Hormuz pivot did so on substrate that surfaced into their inputs within timeframes the trend architecture could act on. The platforms that absorbed the March drawdown did so on substrate that did not surface to the platform-level risk function until the synchronized surge had already landed in pod-level P&L. Substrate widening at the platform-level risk function would have detected the cross-pod correlation buildup before the synchronized drawdown landed.
The fiscal-monetary and industrial-reconfiguration substrates are the canonical examples of the configuration-substrate class the price-derived inputs cannot read. The post-COVID fiscal-monetary configuration decomposes into named structural shocks, patience, earnings, low-earnings-measure, fiscal-support composition, expectations anchoring, knowable from fiscal and monetary regime structure, upstream of the price-derived inputs by construction. The post-2022 industrial reshoring cycle decomposes in parallel, exposure taxonomy, ramp-up factor distribution, capex sector concentration, FDI mega-deal pattern, critical-minerals concentration. A multistrategy platform's pod-level alpha is exposed to all of these through its sector positioning, yet the risk function reads none of them from the price-derived stream. They enter only through configuration-substrate-aware architecture at the platform level.
(a) Integrate the substrate at the platform-level risk function or head-of-research altitude where decisions about leverage, factor allocation and pod-level capital land, the regime-change signal must enter where pre-positioning can be executed. (b) Calibrate regime-change thresholds against the configuration substrate, not the price-derived signal, the price-derived signal is downstream, so using it as a threshold means the regime has already landed. (c) Replay the March 2026 window with the substrate present at the platform-level risk function, the claim is detection 2–4 weeks before the synchronized drawdown, enabling capital reduction before pod-level stop-outs cascade.
Producing accuracy gain, the platform-level risk function fires regime-change signals 2–4 weeks before live surges; leverage reduction in the pre-warning window prevents the synchronized drawdown; peer platforms without integration absorb the drawdown yours pre-positioned against. Not landing, signals fire but the governance threshold does not act; the regime-detection path (HMM, Kalman, changepoint) does not interface cleanly with the existing risk architecture; signal latency exceeds the 2–4 week lead-time window. Magnitude, 20–40% drawdown reduction through compound regime change in well-integrated implementations, conditional on governance responsiveness.
Source · Brady Commission report + 1987 portfolio-insurance microstructure literature · HedgeCo March 2026 · Andolfatto & Martin (FRB St. Louis, May 2026) · McKinsey Global Institute (May 2026)
Resolved precedents
The accuracy claim, anchored to the long record.
Four cycles · two classes
Cross-position differential
The methodology's claim does not stand on owned backtests; it stands on resolved precedents the long record carries, plus the active demonstration the present cycle is composing. Below: the four resolved/active cycles that anchor Layers 2 and 3, and the two configuration-substrate classes that name what the price-derived inputs are upstream of.
The Layer 2 and Layer 3 cycle precedents (A1–A3) carry strong long-record anchoring; the March 2026 Malaise (A0) is the active demonstration still composing its resolved substrate. The fiscal-monetary and industrial-reconfiguration substrates (S1, S2) are named as configuration-substrate classes, not resolved cycles, their signal is upstream of any price-derived validation by construction. The engine operates more strongly where resolved-cycle compare is structurally valid, and surfaces honest gap acknowledgment where the substrate is upstream of price-derived confirmation.
Integration paths
How the substrate enters your model. Four composable paths, not four alternatives.
Compose two or three
Not a menu
Your model architecture determines which paths the configuration substrate can enter through. A factor model consumes substrate as features; an HMM as observable for regime inference; a Bayesian framework as prior on factor dynamics; an ensemble as model-selection or weight-allocation signal. The paths compose, most institutional implementations layer two or three. Pure single-path implementation is rare in sophisticated quant institutions; the composition is where the leverage lives.
Feature engineering
Configuration vectors enter as additional features in the factor-model architecture alongside existing price-derived and alt-data features. Lowest integration complexity; substrate consumed through retraining or online learning.
Fit, factor models · regression stat-arb · NN feature concatenation.
Regime detection layer
Configuration regime enters as latent-state input to an HMM, Kalman filter, or changepoint-detection layer. Produces explicit regime-shift lead time for the model. Substrate consumed through the state-space architecture.
Fit, state-space models · HMM-based trend · regime-switching.
Bayesian prior
Configuration substrate enters as a prior on factor dynamics, tightening or re-weighting the posterior through compound regime change before the price-derived likelihood updates. Substrate consumed through the hierarchical structure.
Fit, hierarchical Bayesian model houses · dynamic-factor architectures.
Ensemble overlay
Configuration substrate enters as a model-selection or weight-allocation signal across an ensemble, reallocating weight toward the sub-models calibrated for the entering regime. Substrate consumed at the allocation layer.
Fit, platform-level risk functions · ensemble architectures.
Integration path matched to architecture produces the full magnitude band; a mismatched path produces zero accuracy gain regardless of substrate quality. The pilot establishes the substrate-architecture baseline through the single cleanest fit; the second path layers once the first is producing measurable improvement.
Comparator implementations
Six implementations, two registers, one substrate gap.
CTAs + multistrategy
Your integration calibration
You are reading your peer implementations' published performance and architectural commentary alongside your own. Four systematic CTAs (Man AHL, Winton, Aspect, Transtrend, all SG Trend Index 2026 constituents) plus two multistrategy quant platforms (Citadel, Millennium). The CTA divergence through Q1 2026 names which implementations capture compound regime changes cleanly; the multistrategy commentary names which platforms are independently building toward configuration-substrate-aware architecture. The maturity columns read where each implementation sits across the three layers.
| Implementation | Layer 1 | Layer 2 | Layer 3 | Methodology read |
|---|---|---|---|---|
Systematic CTA |
●mature | ◐emerging | ○opportunity | Q1 2026 trend programs captured the Hormuz energy pivot cleanly; strong R&D ethos, 50–150 market universe. Strong Layer 1. Methodology contribution at Layer 2/3: configuration substrate as feature or regime layer extends reach into compounds the price-derived substrate cannot detect early. |
Systematic CTA |
●mature | ◐emerging | ○opportunity | Published signal-architecture commentary suggests awareness of input-substrate limits; 2026 portable-alpha UCITS structure. Layer 1 + early Layer 2. Configuration substrate delivers correlation signal the signal architecture consumes through feature integration or a regime layer. |
Systematic CTA |
●mature | ◐emerging | ○opportunity | Core Diversified, long-duration trend with explicit R&D into substrate extension. Layer 1 + 2 with R&D. Bearing substrate compresses the R&D timeline for compound configuration substrate Aspect would otherwise build internally; feature or regime path fits the trend architecture. |
Systematic CTA |
●mature | ○opportunity | ○opportunity | DTP / Enhanced Risk program, diversified trend across instruments, breadth-based substrate. Layer 1, breadth. Configuration substrate adds depth (compound resolution) where Transtrend adds breadth (instrument universe); the two compose naturally through feature integration. |
Multistrategy quant |
●mature | ◐emerging | ◐building | Five strategy businesses; PCRG (Portfolio Construction and Risk Group) reports directly to the CEO. Explicit Layer 3 architecture. Configuration substrate enters at PCRG altitude as ensemble-allocation input or regime-detection signal; the central infrastructure is the natural integration point. |
Multistrategy quant |
●mature | ◐emerging | ◑methodology locus | 330+ independent pods; documented 5% / 7.5% drawdown thresholds at pod level. Mechanical risk infrastructure but limited cross-pod substrate. March 2026 exposed the limit, pod thresholds fire after the cross-pod surge has landed. Configuration substrate at the cross-pod altitude would surface the buildup before stop-outs cascade. |
Composite accuracy gain
Your composite gain across the three layers.
Literature-anchored
Not owned backtests
The three-layer gain composes through the integration path matched to your architecture. Layer 1 contributes zero, the model is already at full accuracy within its substrate. Layer 2 contributes factor stability and cross-position correlation early warning at the factor-model or stat-arb altitude. Layer 3 contributes regime-detection lead time at the platform-level risk function. The bands below are plausible improvement ranges anchored against alt-data integration literature and the resolved precedent windows, conditional on integration path, architecture fit, and governance responsiveness.
(a) Pilot through the single integration path with the cleanest architectural fit, feature engineering for factor models, regime detection for state-space, Bayesian for hierarchical houses, ensemble for platform-level risk functions; the pilot establishes the substrate-architecture baseline. (b) Track improvement against named compound regime windows, the active Hormuz compound is the empirical validation window; CTA performance through Q3 2026 and platform-level risk evolution are the comparator anchors. (c) Layer the second path once the first produces measurable improvement; most institutional implementations converge on two-path composition.
Producing gain, the pilot path delivers Sharpe improvement in the 0.1–0.3 band at multi-quarter horizon; Layer 3 substrate fires 2–4 weeks before live surges; peer comparators diverge in the direction the read named; governance acts on substrate-driven signals within the lead-time window. Not landing, pilot Sharpe gain below 0.05 over a multi-quarter window; signals fire but governance does not act; integration latency exceeds the resolved-precedent lead-time window. Net, +0.1–0.3 Sharpe and 20–40% drawdown reduction through compound regime windows, conditional and literature-anchored.
Scope limits
What this read is not.
Five limits
Not a model replacement. Your architecture and quant discipline remain the canonical substrate. The methodology delivers configuration substrate as additional input the existing architecture consumes through whichever path matches, not a substitute for the model.
Not a backtested Sharpe-improvement claim. The magnitude bands are plausible ranges anchored against alt-data integration literature and resolved precedent windows. PHM does not own claims of the form "fund X integrated Bearing and improved Sharpe by Y%"; the bands are conditional, literature-anchored, and path-dependent.
Not an integration-engineering recommendation. The four paths describe substrate-ingestion architecture at altitude. The specific implementation, codebase, vendor, timeline, composes against your quant team's architecture, not against this read.
Not a governance / risk-infrastructure replacement. The methodology integrates into your existing risk governance, position-sizing and capital allocation. Substrate-driven signals require your governance threshold to act on them; the methodology produces the substrate, your governance produces the action.
Not a substitute for the CIO's or CRO's judgement. The read produces three-layer accuracy-improvement substrate. Integration into model architecture, capital allocation, leverage discipline and the platform-level risk function composes against the CIO/CRO's judgement and the firm's capital strategy.
Colophon
A Bearing Model Accuracy Read.
institutional product
This artefact is a Bearing Hedge Fund Model Accuracy Read, the deployed institutional product, composed at systematic CTA and multistrategy quant altitude for CIOs, CROs and heads of research evaluating Bearing's analytical contribution alongside the model architecture their quant team already runs. The operator's question is substrate-anchored: does the substrate the model trains on extend to the compound configurations entering against my positions, and is the accuracy gain material at my governance threshold. The three-layer ladder is the explanation layer for that question; configuration substrate integration is the canonical analytical move. Bearing is the engine; BearingA is the company; PHM, the Predictive History Method, is the canonical methodology origin invoked only at methodology-claim altitude.
Every load-bearing claim drills to source. The resolved precedents, the August 2007 quant crisis (Khandani-Lo NBER), the August 2024 yen carry unwind (BIS, Wellington), October 1987 portfolio insurance (Brady Commission), are publicly documented; the comparator implementations' architectural commentary is publicly disclosed and citable; the March 2026 multistrategy correlation breakdown is the active demonstration. The configuration-substrate classes (fiscal-monetary, industrial reconfiguration) are named honestly as upstream of any price-derived validation.
The read does not replace the model. It surfaces the three layers of accuracy gain compound configuration substrate produces when integrated through the architecture-matched path, and grounds the gain against resolved precedents the long record carries.
Available for model-architecture integration discussion at the CIO / CRO / head-of-research's discretion · Bearing · BearingA · 23 May 2026
BearingA
The standing configuration-substrate-integration substrate. Bearing reads the three layers of accuracy gain compound configurations produce when integrated through your model's architecture-matched path, the substrate your price-derived inputs are upstream of and cannot natively read.
