Alvermont Research

The Risk Framework

Paper Portfolio
Est. Annualised Vol.
9.5%
Band 8–12% · ceiling 14%
Equity Share of Risk
~86%
On 50% of capital
Max Single Name
6.0%
Or 180 ÷ σ, whichever binds
Vol. Ceiling
72%
Above this, uninvestable

A framework that sized positions in dollars would systematically mis-state what the book is actually exposed to. Every limit in this document is therefore derived from a measurement, and every measurement is one we publish.

This document sets out how positions are sized, how exposure limits are set, and what the firm measures to know whether either is working. It governs the paper book as it stands today — nine ETF lines, four sleeves, $100,000 of notional capital — and it binds from the first trade placed after its issue.

The order of what follows is deliberate. Measurement comes first, because in this framework the limits are not policy preferences that happen to be expressed as numbers; they are consequences of what the covariance matrix says about the book. Change the measurement and the limits move. That is the intended behaviour.

Capital weight and risk weight are not the same number

Equities account for 50% of the book's capital but an estimated 86% of its risk, because equity volatility and intercorrelation dwarf those of the defensive sleeves. The cash sleeve adds essentially no risk. The Treasury sleeve contributes about 2%. In risk terms the book is not the conservative portfolio its 35% cash-and-bonds weight implies.

That is intentional for a book built to showcase equity research. The defensive sleeves diversify and provide dry powder; they do not neutralise equity risk, and were never intended to. But it means a dollar-denominated limit framework would describe a portfolio that does not exist.

Figure 1 — Capital weight versus risk contribution, by position
Full covariance matrix · % of total
Hover any position to compare its capital weight with its share of portfolio variance.

Interpretation. BIL is a fifth of the book and none of its risk. SPY is a fifth of the book and roughly a third of its risk. Any rule expressed as "no more than X% of the portfolio" is, for a position like BIL, not a risk limit at all — and for a volatile single name, a far looser one than it appears. The sizing rules below resolve this by making the binding constraint a risk constraint, with the capital cap retained only as a backstop.

What we measure, and how it is estimated

Every limit downstream depends on the covariance matrix, so the estimation procedure is specified rather than assumed. The inputs below are recomputed monthly and on any change to holdings.

Price historyDaily adjusted close on a rolling ~3-year window — 752 observations at the 18 June 2026 build (21 Jun 2023 – 18 Jun 2026).
SourceYahoo Finance, pasted into the Price Data sheet of the HRP workbook.
EstimatorLedoit–Wolf shrinkage applied to the covariance matrix before any risk-parity step. Sample covariance across a handful of single stocks is too noisy to cluster on.
RecomputeMonthly, and on any change to holdings.
Minimum history250 trading days before any single name may be sized. Without it the covariance input does not exist.

Reference volatilities from the 18 June build, used throughout the sizing rules below:

Figure 2 — Annualised volatility by holding
Rolling 3-year daily returns · 752 observations

Definitions as published

These are the definitions used wherever a figure appears on the site or in a report. No metric is published that cannot be reconstructed from a stated source and a stated method.

Exposure% of net asset value
BetaAgainst S&P 500 total-return index
VolatilityAnnualised σ of daily returns, rolling ~3y
Risk contributionShare of total variance, full covariance matrix
Maximum drawdownPeak-to-trough on stitched daily + 15-min NAV
Sharpe ratioStated with the risk-free rate used — currently 4%
ReturnsPaper account · no fees · no slippage assumption
Stated limitation — carried forward from the model workbook

True hierarchical risk parity — the dendrogram and recursive bisection — cannot be performed in native Excel formulas. Sleeve tagging supplies the hierarchy and the sheet performs the inverse-variance allocation within and across sleeves. This is a transparent, macro-aware risk-parity weighting, not textbook HRP, and it is described that way in anything published. When the researched sleeve reaches eight or more names, the clustering step moves to Python where the dendrogram can actually be computed.

Risk contribution is measured from the full covariance matrix rather than standalone volatility. The distinction matters: a name's own volatility says what it does in isolation, while its marginal contribution says what it does to this book, given everything already in it. Only the second is a risk number.

Five sleeves, each with a defined sizing treatment

Assets are tagged to exactly one of five sleeves. The sleeve is the hierarchy — it is supplied by judgement rather than discovered by an algorithm, and that is stated openly rather than presented as an output of the model.

CASH

Dry powder; principal preservation

Held at a fixed weight, outside risk parity, so a near-zero-volatility instrument cannot distort the optimisation.

DEBT

Ballast; convexity in a growth scare

Inverse-variance within sleeve, inverse cluster-variance across sleeves.

REAL ASSETS

Inflation hedge; tail-risk diversification

As above. Gold is sized at twice commodities, being the lower-correlation and more reliable diversifier.

BASE EQUITY

Core and factor beta; benchmark anchor

As above, with the SPY floor binding before the risk-parity weight does.

RESEARCHED

Published single-name theses — the firm's actual product

As above, subject to every single-name limit in the sizing section. Currently empty; the book is 100% ETFs at the issue date.

Capital balances risk within each sleeve by inverse variance, then across sleeves by inverse cluster-variance, with cash held at a fixed weight. Holding cash outside the optimisation is not a convenience: an instrument with near-zero volatility attracts an unbounded inverse-variance weight, and would otherwise consume the book.

Two limits always apply, and the smaller one binds

Four principles govern sizing. A position is sized by what it contributes to portfolio variance, not by what it costs. Every single name faces both a capital cap and a risk cap, and the binding limit is whichever is smaller — there is no averaging between them. Floors are as binding as ceilings, because the framework fails as easily by abandoning the defensive base as by over-concentrating. And limits are checked before the order, not after the fill.

Capital limits

Limit, as % of NAVFloorTargetCeilingType
Single researched name, at cost2.5%3.5–4.5%6.0%Hard
Single researched name, at market7.5%Hard
Single ETF line20.0%Hard
Researched sleeve, total50%55%Hard
SPY anchor10%12%20%Hard
Defensive base (IEF + GLD + DBC)25%30%35%Hard
Cash sleeve (BIL)8%8–10%no capHard
Total equity (base + researched)62%70%Hard
Gross exposure100%100%Hard
Number of researched names810–1416Soft
The 6% single-name cap is set so that the complete failure of any one thesis — a 50% drawdown in the name, which is what a broken thesis looks like — costs the book 3% of NAV. That is survivable, publishable and explicable. A 10% cap would make a single mistake a 5% hit and would put the −12% drawdown protocol in play on one name. The 16-name ceiling is a coverage constraint, not a risk one: stale coverage is worse than none, because it is indistinguishable from current coverage to a reader.

The risk cap — the binding constraint for volatile names

The governing rule is that no single researched name may contribute more risk than the SPY sleeve. SPY is the benchmark; a name the firm has researched may be sized to matter, but not to matter more than the thing it is measured against.

At the SPY anchor of 12% and SPY volatility of 15%, the sleeve's volatility-weighted contribution is 12 × 15 = 180. A candidate name is therefore capped at:

max weight (%)  =  min ( 6.0 , 180 ÷ σname )   where σ is annualised volatility in percent
Figure 3 — Maximum position size as a function of name volatility
At $100,000 NAV · capital cap versus risk cap
Below 30% volatility the 6% capital cap binds. Above it, the risk cap binds and falls hyperbolically. Hover the curve to read the maximum position at any volatility.
Rule S-1 — The volatility ceiling Hard

The risk cap and the 2.5% minimum position cross at approximately 72% annualised volatility. A name more volatile than that cannot be sized large enough to be worth holding without breaching the risk cap, and is therefore outside the investable universe regardless of conviction. This is the framework's answer to the pre-revenue biotech question, and it is an answer given in advance.

Proxy versus true contribution. The w × σ formulation above is correlation-blind, and deliberately so: it is the pre-trade screen. It is computable in seconds, it is conservative for names correlated to the book, and it never requires a model run to answer "can I own this at all". The binding measure at monthly review is the true marginal risk contribution from the full shrunk covariance matrix. Where the two disagree, the covariance figure governs and the position is adjusted at the next review. A name that passes the proxy but fails on true contribution is almost always a name highly correlated to something already held — which is information about the book, not just about the name.

Rule S-2 — The untouchable base Hard

IEF, GLD and DBC are off-limits to the deployment programme and may not be sold to fund a researched position under any circumstance. SPY may be trimmed but never zeroed, with a floor of 10% of NAV.

These are the only holdings in the book that are not for sale, and the reason is structural: the defensive base is what allows a concentrated single-name book to be run at all, and the SPY anchor is what makes benchmark attribution honest. A firm that sells its benchmark to buy its own ideas has stopped measuring itself.

Exposure is counted through the ETFs, not around them

Sector limits apply on a look-through basis. A portfolio can hold no technology stocks and still be at a technology ceiling, and a framework that only counted direct holdings would not notice.

ExposureCeilingTypeBasis
Any one GICS sector, incl. ETF look-through20%HardLook-through weights from published fund holdings
Any one GICS industry group12%SoftSingle names only
Any one country outside the US25%SoftDomicile of primary listing
Unhedged non-USD revenue exposure30%SoftReported revenue mix, not listing venue
Top five positions combined35%SoftExcludes BIL
Any one theme with a shared driver25%SoftDeclared in the note; e.g. AI capex, China consumer
Figure 4 — Technology look-through against the 20% sector ceiling
Estimated tech content by line · % of NAV
Estimate. Sector content is taken from typical index composition, not from current fund fact sheets, and must be confirmed against published holdings at the next monthly review before it is quoted externally.

Two consequences follow directly. First, the book sits at its technology ceiling before a single technology stock has been researched — which is the quantitative justification for XLK sitting second in the recycle queue. Second, no researched technology name may be added until XLK is recycled. Neither of those is a judgement call made at the time; both fall out of the measurement.

Where the book sits against its own ceilings

MetricCurrentTarget bandCeilingType
Annualised portfolio volatility9.5%8–12%14%Hard
Equity share of total risk~86%80–90%92%Soft
Beta to SPY~0.60.5–0.91.0Soft
Effective number of bets≥ 5≥ 4 floorSoft
Net exposure100%100%100%Hard
Figure 5 — Current readings against target bands and ceilings
Band shown shaded · ceiling marked

The volatility ceiling will bind before the capital limits do. Replacing ETFs at 15–24% volatility with single names at 25–40% raises book volatility even after diversification across ten to fourteen names. A reasonable expectation for a fully deployed researched book is 11–13% against a 14% ceiling. If the projected figure exceeds 14%, the response is to hold fewer, lower-volatility names, or to run the researched sleeve below its 50% target — not to raise the ceiling. This is modelled in the workbook before deployment begins, not discovered after it.

Liquidity and eligibility

A name must clear all of the following before it can be sized at all: 250 trading days of price history Hard; $2bn minimum market capitalisation Soft; $10m median daily traded value over 60 days Soft; and a position no greater than 1% of median daily volume Soft. Non-US businesses are held via ADR or US listing. At a $6,000 maximum position the volume constraint is non-binding for any name meeting the others, and is stated so that it scales with the book rather than needing to be invented later.

Two kinds of position, two entirely different disciplines

Rule E-1 — The governing distinction Hard

An ETF is exposure you rebalance. A single stock is a bet you exit at a target.

ETF sleeves — including the equity sleeves — are broad exposure vehicles and are managed back to target weight. They are never "taken profit on": selling SPY because the market has risen is market timing wearing a risk-management costume. Individual company positions are bought against a fair-value target derived from a published thesis, and therefore have a defined exit. Applying take-profit logic to ETFs, or rebalancing logic to a single name whose thesis has broken, are the two symmetrical errors this rule exists to prevent.

Rebalancing — ETF sleeves

TriggerThresholdAction
Absolute drift from target± 3.0 ppRebalance the line to target at the next monthly review
Relative drift from target± 25%As above — whichever trigger fires first, for small target weights
Sleeve-level drift± 4.0 ppRebalance within the sleeve before rebalancing across
Floor or ceiling breachanyPer breach policy below
CalendarquarterlyFull rebalance to target regardless of drift
Current status: maximum line drift is SPY at +0.54pp; maximum sleeve drift is Base Equity at +0.23pp. Both are an order of magnitude inside the bands. No rebalancing action is required or permitted at the issue date — trading inside the band is cost without benefit, and in a published book it is also noise in the track record.

Exits — researched single names

These rules become operative on the first researched fill. As at the issue date the book is 100% ETFs, so no take-profit rule is currently live.

Exit typeTriggerAction
Target reachedPrice meets the published 12-month targetRe-underwrite or exit — never drift. Holding past a target without re-underwriting converts a researched position into an unresearched one.
Thesis stopA published falsifier occursExit within five trading days. Publish a note explaining what was wrong.
Risk reviewPosition −25% from costMandatory written review, not an automatic exit. A price stop on a fundamental thesis would systematically sell exactly the mispricing the thesis predicted.
Drift trimPosition exceeds 7.5% of NAV at marketTrim back to 6.0% at the next review.
Time stop18 months with no measurable thesis progressExit or re-underwrite. Capital and, more importantly, coverage capacity are finite.
Coverage lossThe analyst can no longer maintain the name to standardExit. An unmaintained published position is worse than no position.

Drawdown protocol

Measured on NAV from the prior peak, using the same equity curve published on the website.

Figure 6 — Drawdown escalation ladder
Current maximum drawdown since inception · −2.04%
LevelResponseDetail
−5%NoteRecorded and commented on in the monthly risk note. No action.
−8%Written risk reviewFull attribution: what lost the money, whether it was the theses or the beta, whether any limit is implicated. Published.
−12%Mandatory de-riskRestore defensive floors to target. No new single-name positions until the review is complete.
−15%Halt and re-underwriteAll new positions halted. Every open thesis re-underwritten from scratch and the outcome published. The framework is reviewed — but may only be tightened, never loosened, while in drawdown.
Current maximum drawdown since inception: −2.04%. No level is engaged.

Breach policy

Breach typeDefinitionRequired action
ActiveCaused by a trade the firm placedMust not occur. If it does: unwind to compliance same session, log it, and disclose it in the next published risk note.
PassiveCaused by market movement, or by a new limit applying to a pre-existing positionCorrect at the next monthly review, or sooner if the excess exceeds 1.5× the limit. No forced intra-month trading.
SoftAny limit tagged SoftWritten explanation in the monthly review. May be carried deliberately if the reason is stated and dated.
Every breach, of any type, is recorded in the breach log with the date, the limit, the magnitude, the cause and the resolution. The log is reviewed quarterly. A limit breached three times in four quarters is either wrong or not being taken seriously, and the review must decide which.

Monitoring cadence

DailyConfirm the feed is live and NAV is plausible. No trading decisions are made daily.
WeeklyDrift check against the rebalance bands; research pipeline stand-up.
MonthlyFull covariance recompute with shrinkage; risk-parity refresh; risk-contribution check against the risk cap; rebalance if outside band; publish the risk snapshot; review the breach log.
QuarterlyFull rebalance; thesis-status review of every open name; performance attribution against SPY; review of this framework.
Event-drivenEarnings for any held name; any published falsifier; any drawdown level.

What is published, and what the firm commits to disclosing

Published continuously

Net asset value and indexed equity curve; open positions with weight, entry and return; net, gross, long and short exposure; largest position; top sector weight; maximum drawdown; cash weight; recent trade activity. These populate from the paper account and are recomputed daily.

Published monthly

Estimated portfolio volatility; risk contribution by position and by sleeve; correlation matrix; effective number of bets; benchmark-relative return with a plain-language attribution; and any breach together with its resolution.

Honesty commitments

Benchmark-relative underperformance is reported in the same place, and with the same prominence, as outperformance1
The paper-account caveat appears wherever a return is quoted, without exception2
Estimates are labelled as estimates, with the sample window stated3
Failed theses get a published post-mortem, not silent removal from the site4
Methodological limitations are disclosed in the document that relies on them5
No metric is published that cannot be reconstructed from a stated source and a stated method6

The framework is reviewed quarterly. Review is a scheduled task, not a discretionary one: an unreviewed manual past its review date is treated as a soft breach and disclosed in the next risk note. Any amendment that loosens a limit, widens the universe, or changes the mandate requires written agreement and a version increment. Any amendment that tightens a limit may be made unilaterally by the risk owner and ratified at the next review.

The one rule that has no exception

A limit is not a target and it is not a suggestion. If a limit is inconvenient, the correct response is to amend it deliberately at a review, in writing, before the trade — never to exceed it and document the exception afterwards. The credibility of published research rests entirely on the claim that the sizing framework was applied as written.

Alvermont Research publishes independent research for educational and informational purposes only. Nothing in this document constitutes financial, investment, legal or tax advice, nor a recommendation or solicitation to buy or sell any security. The portfolio described is a paper account and does not represent real capital or actual trading. Volatility, correlation, return, risk-contribution, beta and Sharpe figures are estimates computed from historical daily data over a finite sample and will differ from future realised values. Sector look-through figures are estimates based on typical index composition rather than current fund fact sheets. Past performance, whether actual or hypothetical, is not indicative of future results. © 2026 Alvermont Research.