A short memory buys the wrong risk

ARCANE Research

Abstract

We specify and evaluate Orpheus, an allocation method for a core-plus ETF sleeve whose mandate is more yield and total return than the Bloomberg US Aggregate without exceeding its volatility of 4.64%. Orpheus maximises trailing distribution yield subject to a volatility ceiling on a full-history covariance, the mandate’s credit ceilings and sector floors and a 20% position cap, and adds a one-signal duration guard. Three findings. First, the volatility-only optimiser’s preference for senior loans over agency mortgage-backed securities is an artefact of the three-year estimation window: it is absent from every 36-month window ending between 2016 and 2022 and present in every window ending in 2023 or later, on conditional value-at-risk as well as on volatility. Second, on the platform’s eligible menu the frontier under the mandate’s limits is flat: the Orpheus book yields 5.07% at 3.80% volatility, against 4.85% at 5.64% for the incumbent book and 4.16% at 4.69% for the Aggregate, with a worst drawdown of −12.7% against −18.9% and −17.1%. Third, in a walk-forward race of published construction methods, construction alone moves the Sharpe ratio little; the duration guard, a standard time-series momentum rule, raises it from −0.14 for the Aggregate to 0.40 (Sortino −0.18 to 0.64), with a stationary-bootstrap interval that excludes zero against both the Aggregate and the incumbent book. Every figure is a hypothetical backtest on public data.

Keywords core-plus fixed income · volatility ceiling · estimation window · credit limits · time-series momentum · walk-forward evaluation

1Introduction

The mandate is one line: hold more yield and earn more total return than the Bloomberg US Aggregate, with no more volatility than the Aggregate’s 4.64%. The obvious implementation, maximise yield under a volatility ceiling, was built in July 2026 and reviewed by the client’s portfolio managers. The review found the defect that motivates this paper: the optimiser’s lower-volatility books held more leveraged loans and less agency MBS, and would have lost more in a credit event than the incumbent book. The optimiser was not wrong at arithmetic. On the trailing data it was given, government-guaranteed MBS measured more than twice as volatile as B-rated loans, so a volatility ceiling instructed it to sell the one and buy the other and to record the trade as a reduction in risk.

This paper makes three contributions. It locates the defect in the estimation window rather than the risk measure (Section 4.1). It rebuilds the allocation on the platform’s actual eligible menu under the mandate’s own limits and reports the book, the frontier it sits on and the statistics a risk committee asks for (Section 4.2 to Section 4.4). And it races the construction methods the literature credits with better risk-adjusted returns on the same menu, out of sample, and shows that the one lever that moves the Sharpe ratio is a duration trend rule, then tests that rule for robustness (Section 4.5, Section 4.6). The method is disclosed in full; what ARCANE retains is the implementation and its calibration to the client’s limits.

The evaluation takes the sceptical literature seriously. Estimation error in expected returns swamps the gains of optimisation out of sample [2], which is why the objective is an observable yield rather than a forecast and why the covariance is shrunk in every out-of-sample test [6]. Risk-based construction avoids expected returns altogether [8, 1, 7] and is raced here on equal terms. Volatility targeting raises Sharpe ratios in equities and credit but not in bonds [9, 4], and the data agree. Trend rules on duration are old and public [3, 10]; the guard is one of them.

Restricted from here

The rest of this paper opens with an access code.

This paper was written for one portfolio team and its firm’s risk reviewers. The abstract and the introduction are public and may be cited. The method, the results, the references and the appendices open with the access code the team was given.

Held 2. Data · 3. Method · 4. Results · 5. Discussion · 6. Conclusion and recommended policy · References · Appendices

Cite as

ARCANE Research (2026). A short memory buys the wrong risk. ARCANE Research Technical Report 06, v1, 27 September 2026. https://arcaneintel.net/research/orpheus

@techreport{arcane2026orpheus,
  title       = {A short memory buys the wrong risk},
  author      = {{ARCANE Research}},
  institution = {ARCANE},
  type        = {Technical Report},
  number      = {06},
  year        = {2026},
  month       = sep,
  note        = {Version 1},
  url         = {https://arcaneintel.net/research/orpheus}
}