Multi-Strategy Portfolio Construction Under EVAR Constraints

1. Introduction

A client who runs several systematic strategies on one set of exchange accounts holds one book. Its strategies draw on the same collateral, pass through the same margin engine and, in stress, load on the same market factors. This paper sets out a framework for sizing such a book: how to combine strategies, how to set allocation limits that respond to the market regime, and how to hold a loss limit on the whole book rather than on each strategy alone.

The setting is Hedgehog Capital's. Hedgehog is an AI-first global macro research-and-risk service for professional and high-net-worth investors. Clients keep their capital on their own exchange accounts at Kraken, Coinbase, Deribit, Binance or Hyperliquid, and each client's dedicated Hedgehog AI trades there through an API key with read and trade permissions only. Hedgehog never holds client assets and is not a fund.

In this paper a segregated account means a sub-account on the client's own exchange account that is dedicated to the client's Hedgehog AI, as Hedgehog recommends. It is an ordinary trading sub-account under the client's own exchange agreement. The term does not mean legal segregation, custody protection or insolvency protection, which depend on each exchange's terms. Nor is it Deribit's "segregated" margin setting, which keeps each currency's collateral separate. Deribit margins each sub-account entirely separately, whatever the margin account type (Deribit). The sub-account is the unit of margin and liquidation, because every strategy inside it shares its collateral and liquidation rules. A client whose Hedgehog AI trades on more than one exchange has one book for loss limits but several margin units, and each needs its own check.

Hedgehog has published the outline of its risk architecture. Size has two limits, and the smaller wins: the client's allocation, and the most that strategy can take. EVAR limits then tighten or widen with the regime. EVAR is calculated for each strategy and for the book as a whole, and when several strategies approach their individual EVAR ceilings at the same time, the overall book reduces further. Leverage is the client's choice for each strategy, from 1x (none), and the exchange's own margin rules always apply. This paper does not extend that description or state EVAR's formulas or parameters, which are not published.

Section 2 reviews the academic and regulatory evidence. Sections 3 and 4 measure, from public data through 25 October 2026, the market behaviour a book-level limit in crypto must cope with. Section 5 sets out a general framework, and Sections 6 and 7 cover implications and limitations. The paper reports no returns or performance of any strategy, actual or hypothetical. Every figure is either a market statistic or a labelled hypothetical calculation that uses no return inputs. The generic loss models tested are not EVAR.

2. Background and evidence

Allocate risk rather than forecast returns. Expected returns are the weakest input to any allocation. DeMiguel, Garlappi and Uppal compared 14 optimised models with an equally weighted portfolio across seven datasets and found none reliably better out of sample, because estimation error offset the gain from optimisation. Their calibration suggested that a sample-based mean-variance model would need an estimation window of around 3,000 months for 25 assets before it could be expected to do better (DeMiguel, Garlappi and Uppal). MacLean, Thorp and Ziemba, citing Chopra and Ziemba, rank errors in means, variances and covariances at about 20:2:1 in importance, a ratio that depends on risk aversion (MacLean, Thorp and Ziemba). Equal risk contribution portfolios avoid the weakest input: each component contributes the same share of risk, no expected returns are needed, and volatility lies between the minimum-variance and equally weighted portfolios (Maillard, Roncalli and Teiletche). Risk parity holds more of the lower-risk assets and applies leverage to reach a target volatility (Asness, Frazzini and Pedersen). On a client's exchange accounts that leverage is bounded by the client's choice and by venue margin rules.

Use a tail measure that can be added up. Value-at-risk (VaR) is not subadditive in general: the VaR of a combined book can exceed the sum of its parts. Acerbi and Tasche gave a definition of expected shortfall (ES), the average loss beyond the VaR, that is subadditive and stays coherent when the distribution has jumps, as scenario sets do (Acerbi and Tasche). Rockafellar and Uryasev showed that conditional value-at-risk, the same tail average for continuous distributions, is never smaller than VaR and can be minimised by linear programming over scenarios (Rockafellar and Uryasev). For a risk measure that scales in proportion to position size and is differentiable, such as volatility, or expected shortfall when losses have a continuous distribution, Euler contributions add up exactly to the risk of the whole (Tasche), so a book limit can be divided into sleeve limits without double counting.

Bank regulation took the same route. Basel's market risk standard uses a 97.5% one-tailed ES, scales a 10-day base figure measured directly over 10 days to liquidity horizons of 10 to 120 days, and calibrates the measure to the most severe 12-month stress period in data back to 2007, with equal weights (Basel Committee, MAR33; Basel Committee, 2019). These are bank capital rules, used here only as design context. The design lesson is a tail measure, a horizon matched to exit, and a stressed figure kept beside the current one.

Condition on the regime, with care. Volatility clusters, and RiskMetrics warned that an equally weighted window changes abruptly when a large shock drops out of it (J.P. Morgan and Reuters). Moreira and Muir scaled factor exposure by the inverse of the previous month's realised variance; taking less risk when volatility was high improved the risk-return trade-off in their sample, because volatility changes were not matched by proportional changes in expected returns (Moreira and Muir). Harvey and co-authors found across more than 60 assets that volatility targeting reduced the likelihood of extreme outcomes and the volatility of volatility (Harvey et al.). Cederburg, O'Doherty, Wang and Yan supply the counterweight: across 103 equity strategies, volatility-managed versions were not systematically better than the originals in direct comparisons, and weak out-of-sample results came mainly from structural instability (Cederburg et al.). Volatility scaling holds risk roughly steady. It is not a forecast of return.

Correlation also depends on the regime. Longin and Solnik found that large negative returns in the five largest equity markets from 1959 to 1996 were much more correlated than a normal model predicts, while large positive returns were not (Longin and Solnik). Forbes and Rigobon showed that part of any measured rise in crisis correlation is a statistical effect of higher volatility (Forbes and Rigobon). Crypto's link to other markets has moved too. Liu and Tsyvinski found that cryptocurrencies had no exposure to most common stock market and macroeconomic factors in their sample (Liu and Tsyvinski), while IMF staff later reported that the bitcoin-S&P 500 correlation rose from 0.01 in 2017-2019 to 0.36 in 2020-2021 (Adrian, Iyer and Qureshi).

Hedgehog's published description of EVAR is aimed at this problem. EVAR starts from the structure of a standard loss estimate at a given confidence level but conditions the distribution on the current regime rather than a fixed window. The regime is a continuous estimate refreshed every day from cross-asset signals, realised volatility and spread behaviour, and EVAR updates in real time as prices move. The EVAR note states that EVAR does not predict direction, is not a guarantee and is not a replacement for judgement.

Constrain the path and the aggressiveness. Grossman and Zhou showed that when wealth must stay above a fixed fraction of its running maximum, the risky position is proportional to the surplus over that floor (Grossman and Zhou). Chekhlov, Uryasev and Zabarankin defined conditional drawdown-at-risk, usable as an optimisation constraint (Chekhlov, Uryasev and Zabarankin). Such rules act after losses; the EVAR note contrasts stop-losses, which exit after the loss, with EVAR-based sizing, which reduces exposure while losses are still possible. The Kelly criterion bounds aggressiveness from above, but its positions can be very large, and because inputs are uncertain MacLean, Thorp and Ziemba argue for smaller, fractional positions (MacLean, Thorp and Ziemba). Busseti, Ryu and Boyd add an explicit limit on drawdown probability (Busseti, Ryu and Boyd). In the standard approximation the Kelly position scales with the inverse of variance: if volatility doubles and nothing else changes, it falls to a quarter.

Liquidity, crowding and the venue. Bangia and co-authors showed that ignoring exit costs through the bid-ask spread can understate market risk in emerging markets by as much as 25 to 30% (Bangia et al.). Brunnermeier and Pedersen showed that margins can be destabilising and that market and funding liquidity reinforce each other in spirals (Brunnermeier and Pedersen). Pedersen analysed why investors crowd into trades and run together (Pedersen), and Khandani and Lo hypothesised that the August 2007 quant losses began with the rapid unwinding of one or more large quantitative equity market-neutral portfolios, and that the resulting losses triggered stop-loss and de-leveraging policies at other funds (Khandani and Lo). A rule that cuts risk when measured risk rises can add to the selling when many books share it.

Crypto adds risks from the venues themselves. The IMF judged that heavy leverage likely deepened the 40% fall in crypto market capitalisation in May 2021, and recorded exchange outages on 19 May, when liquidations peaked (IMF). On 10 October 2025 more than $19 billion of crypto leverage was liquidated in roughly a day, top-of-book bitcoin depth fell by more than 90% on key venues, and auto-deleveraging reduced or closed some of the best-hedged short positions (FTI Consulting). On Kraken, isolated margin limits the funds at risk to the margin set aside for a position, but if the equity of the whole wallet falls below the maintenance margin for all positions, the account is liquidated, isolated positions included (Kraken).

Model risk and validation. Kupiec's test checks how often a VaR estimate is breached (Kupiec), and Christoffersen showed that breaches should also be independent of each other (Christoffersen). Danielsson, James, Valenzuela and Zer found that risk models agree in calm periods and disagree much more in distress (Danielsson et al.). Danielsson and Zigrand showed that under jumps, square-root-of-time scaling understates risk (Danielsson and Zigrand). Federal Reserve staff found that banks' VaR models breached together in March 2020 (Abboud et al.). Section 4 tests these points on crypto data.

3. Data and method

Data. Daily UTC candles from Coinbase Exchange for BTC-USD (from 20 July 2015, 4,116 days, none missing), ETH-USD (from 18 May 2016) and SOL-USD (from 17 June 2021), cross-checked against Binance spot (daily return correlation 0.9913 for BTC and 0.994 for ETH over 3,356 days). Binance spot hourly candles from 2019 for intraday paths. Deribit's DVOL 30-day implied volatility index (Deribit), with statistics from 1 April 2021. Eight-hourly funding rates on Binance and Bybit USDT-margined perpetuals. The S&P 500 and VIX from FRED (SP500; VIXCLS). Margin tiers and order books from public exchange APIs, captured on 26 October 2026 at 14:38 and 14:37 UTC.

Conventions. Daily log returns of UTC closes; losses as simple-return losses; volatility annualised by the square root of 365. A forecast for day t uses data to the close of day t-1. rv30 is the annualised standard deviation of the previous 30 daily returns. Volatility regimes are quintiles or terciles of BTC rv30 (or of DVOL in Section 4.5) known at the start of each day, with full-sample thresholds unless stated.

Generic loss models. Four textbook 99% one-day VaR estimates: historical simulation over 180 and 365 days (HS180, HS365), a normal estimate with RiskMetrics EWMA volatility (decay 0.94), and filtered historical simulation (FHS365). None is EVAR, and the results say nothing about EVAR or any strategy.

Episodes. Nine stress episodes since 2020, each dated by BTC's worst UTC close-to-close loss in its window. For February 2026 we rely on Nasdaq, which describes sharp declines with spot bitcoin ETF outflows against geopolitical tensions, US tariff uncertainty and AI-driven fears, and total digital asset market capitalisation falling from about $4.2 trillion in October 2025 to a low of about $2.17 trillion (Nasdaq Global Indexes).

Hypothetical illustrations. Three exercises use assumed inputs and are labelled: the exposure a volatility-targeting rule would hold for a hypothetical 50% target (exposures only, no returns), a three-sleeve book used to show the arithmetic of a book-level ES limit (no return inputs), and approximate liquidation distances for hypothetical isolated positions, compared with BTC's market moves. None represents a Hedgehog strategy, parameter or EVAR. Raw data and scripts are retained, and every Section 4 statistic not attributed to a cited source is reproducible from them.

4. Findings

4.1 Volatility moves a lot, and it persists. BTC's 30-day realised volatility since August 2015 has ranged from 14.0% (11 October 2016) to 209.4% (7 April 2020), with a 5th percentile of 25.9%, a median of 55.3% and a 95th percentile of 110.6%, 4.28 times the 5th. For ETH the 95th percentile is 4.07 times the 5th, and for SOL 3.01 times. Calendar-year BTC volatility fell from 93.1% in 2017 to 41.9% in 2025 and stands at 44.9% for 2026 to 25 October. Price moves are large: BTC fell 53.1% from $124,720 on 6 October 2025 to $58,524 on 30 June 2026, and closed at $80,053 on 25 October 2026, 35.8% below that peak.

Volatility persists. The autocorrelation of BTC daily returns at one day is -0.048, but of absolute returns 0.203 at one day and still 0.083 at 60 days. With days in terciles of rv30, a high-tercile day was followed by another 95.3% of the time, and 0.08% of high-tercile days were followed by a low-tercile day. High-tercile spells had a median of 8 days and lasted up to 195 days. Because consecutive 30-day windows share 29 returns, much of this day-to-day persistence is mechanical; the autocorrelation of absolute returns at 60 days shows that regimes outlast the window.

4.2 Most shocks began from below-median volatility. Persistence did not give warning. In seven of the nine named episodes, BTC's 30-day volatility on the eve of the worst day was below its historical median to that date (Table 1), and in two it was in the lowest fifth. Before 10 October 2025 it was 30.3%, at the 7.9th percentile. Only the May 2021 unwind and the Celsius and Three Arrows failures had above-median volatility on the eve of their worst day.

The worst BTC day breached the generic HS365 estimate in eight of nine episodes and the EWMA estimate in all nine. The models responded afterwards. On 10 October 2025 the EWMA estimate was 3.80%; five days later it was 5.46%. A daily-updated estimate sees a shock only at the next close: the EWMA volatility used for BTC on 12 March 2020 was 64.5%, and on 13 March 238.2%. Episodes also ran over days: BTC fell 38.75% from the close before 12 May 2021 to the close on 23 May, and 37.0% over 10 to 18 June 2022. Reversals were sharp too. BTC rose 12.32% on 6 February 2026, the day after its 13.98% fall, so a rule keyed to the previous close would have cut exposure into that rise.

A regime estimate therefore cannot rely on realised volatility alone. Hedgehog's published description lists cross-asset signals and spread behaviour alongside it. This paper does not test whether any such inputs gave warning before these episodes. Whatever the inputs, a book also needs limits that do not depend on the regime estimate being right on the eve of a shock.

4.3 Fixed windows and static calibrations misfire. Over 3,585 days, a correct 99% estimate for BTC would be breached about 35.9 times (Table 2). The normal EWMA model was breached 77 times: crypto tails are too fat for a normal distribution. HS365 came close on count, with 42. Two of its breaches fell on consecutive days, so a breach was 4.8% likely the day after a breach against 1.1% otherwise, but the independence test does not reject at the 5% level (p = 0.100). The filtered model, which rescales history by current volatility, combined a near-correct count with no significant clustering. Clustering is clearer for ETH. Over 3,415 days from 20 June 2017, its HS180 estimate was breached 54 times against 34.1 expected, and a breach followed a breach 9.3% of the time against 1.5% otherwise (independence test p = 0.002).

Fixed windows also cut risk for no market reason. All 20 of the largest one-day falls in BTC's HS365 estimate since 2017 came when one of the five largest losses in the window dropped out. The estimate for 20 August 2023, set at the close of 19 August, fell 1.67 percentage points after a day on which BTC rose 0.18%, because the 10.15% loss of 19 August 2022 had left the window. The estimate for 20 May 2022, eleven days after the worst day of the Terra-Luna collapse, fell 1.12 points because the 14.29% loss of 19 May 2021 had left. The median absolute BTC move on the day each new observation entered the window was 1.22%, against 1.51% on all days. The estimate for 11 October 2026 fell 0.51 points as the 10 October 2025 loss left.

Static calibrations miss in both directions (Table 3). The 2022 calibration of 10.23% for BTC was never breached in 2023, when the realised 99% loss was 5.13%. The 2023 calibrations were then breached 10 times by BTC and 18 times by ETH in 2024, against about 3.7 expected. The 2021 calibration, by contrast, was breached by BTC in 2022 about as often as expected, three times against about 3.6. After calm 2025, misses grew with the horizon. In 2026 the 2025 calibrations for BTC were exceeded on 1.0% of days at one day, by 5.9% of 10-day windows and by 11.2% of 30-day windows (30-day threshold 21.37%), against 1% expected. The windows overlap: 24 of the 30 exceeding 30-day windows ended between 4 and 27 February 2026, during and after the February sell-off.

Model disagreement widens after shocks. The median spread between the highest and lowest of the four BTC estimates was 2.47 percentage points in the calmest rv30 quintile and 4.97 in the most volatile. The ratio of highest to lowest did not rise across volatile periods (1.67 calm, 1.50 volatile), because a one-year window remembers old stress. Disagreement did widen straight after shocks: on the day after the worst day, the spread exceeded its all-days median of 3.35 points in seven of nine episodes, and it reached 20.7 points after 12 March 2020, when the ratio was 3.04.

4.4 Diversification across coins shrinks on the way down. From 2019, the BTC-ETH correlation was between 0.801 and 0.845 in every BTC-volatility quintile. Volatility level alone moves it little. Direction does (Table 4). On days when both fell more than 1.5 standard deviations, their correlation was 0.899 against 0.462 for a normal benchmark; when both rose that far it was 0.294, below the benchmark of 0.464. A normal benchmark at the same threshold removes the mechanical effect of conditioning on large moves. It does not remove the effect of fat tails or changing volatility that Forbes and Rigobon describe, but in a symmetric distribution with constant correlation those effects would raise both tails together. Here the upside correlation sits below its benchmark and the downside far above it, and that asymmetry is the stronger evidence.

Other measures agree. From 2019, ETH's sensitivity to BTC was 1.183 on days when BTC fell more than two standard deviations and 0.665 on days when it rose that far. The correlation was 0.693 with BTC within 10% of its peak and 0.869 when it was more than 30% below. On BTC's worst 5% of days since June 2021 (98 days), ETH and SOL also fell 93.9% of the time; the average falls were 6.35% for BTC, 7.53% for ETH and 8.40% for SOL. A multi-coin book holds one dominant factor, and it strengthens in sell-offs.

4.5 Strategy types meet through market factors. Different strategies lose together when they load on the same factors in stress. From April 2021, the correlation between BTC's daily return and the daily change in Deribit's DVOL was -0.12 when DVOL had closed in its lowest quintile the day before and -0.49 when it had closed in its highest. On BTC's worst 5% of days in that period DVOL rose 4.75 points on average. Short-volatility and long-directional sleeves therefore tend to lose on the same days, in line with Hedgehog's published note that volatility harvesting losses often coincide with losses across other strategy types. Implied volatility also jumped from low levels. BTC DVOL closed at 39.7 on 9 October 2025 and reached an intraday high of 64.9 on 10 October (close 44.1). It closed at 38.0 on 28 January 2026, the day before the February 2026 episode window opens, and at 53.5 on 4 February, then reached an intraday high of 90.3 on 5 February (close 82.6).

Crypto's link to equities also rises with equity stress. The weekly BTC-S&P 500 correlation from November 2016 to 16 October 2026 was 0.064 in weeks that began with the VIX below 15, 0.199 with the VIX from 15 to 25 and 0.365 with the VIX at 25 or above. By full calendar year it ranged from -0.093 in 2019 to 0.429 in 2022, and the rolling 52-week figure reached 0.540 on 10 July 2026.

4.6 The arithmetic of a book-level limit (hypothetical). Table 5 shows how rising volatility and converging correlations compound, in a hypothetical book of three sleeves weighted for equal risk contribution. For stress, each volatility is multiplied by 2.14, the measured ratio of BTC's volatility in its most volatile rv30 quintile to its calmest, and every correlation rises to an assumed 0.6. To hold book ES at its calm level, gross exposure must fall to 0.381 of its calm level, a 62% cut. Volatility does most of the work: with correlations unchanged, the volatility rise alone requires a factor of 0.467, while correlation alone, rising to 0.6, requires 0.816.

The fifth row carries the main lesson. In calm conditions the stand-alone ES of the three sleeves sums to 3.73%, 43% more than the book's 2.61%. In stress the sum is 7.99% against 6.84%, only 17% more. Sleeve limits that are set separately and summed rely on a diversification credit that shrinks when it is most needed. A book-level limit, with sleeve limits set from Euler contributions under stressed correlations, closes that gap. Hedgehog's published architecture states a related rule: when several strategies approach their individual EVAR ceilings together, the overall book reduces further. How book EVAR treats correlation is not published, and this paper does not describe it.

4.7 Volatility scaling moves exposure late and both ways (hypothetical). With a hypothetical 50% target and an EWMA forecast with a 20-day half-life, BTC exposure from 2017 ranged sevenfold, from 0.27x to 1.88x, and turned over 3.72 times a year (Table 6). A 1x cap bound on 32.7% of days. The rule holds its pre-shock exposure on the shock day. On 10 October 2025 its exposure stood at the 97th percentile of its distribution since 2017, because volatility had been low, and it fell by almost a quarter the next day. It rebuilds slowly, taking 57 days after May 2021 and 33 after FTX to regain the pre-shock level. It cuts on rallies too: the largest five-day cut, of 0.79x, ended on 22 August 2026, after BTC rose 7.14%, 5.36% and 7.28% on 19, 20 and 21 August.

The forecasts are noisy. Realised volatility over the 30 days starting on the forecast day exceeded the EWMA forecast by more than half on 9.9% of days and fell short by a third or more on 17.5%, with a worst ratio of 3.88 for the forecast made at the close of 7 March 2020. Volatility scaling is a useful constraint layer that should sit inside harder limits, not replace them.

4.8 Margin engines see the intraday path. On Binance hourly data, BTC's intraday low on 10 October 2025 was 16.16% below the previous UTC close, against a Binance close-to-close fall of 7.31% (7.18% on Coinbase, Table 1). On 19 May 2021 the figures were 29.99% and 14.38%. Since 2019, BTC's 99th percentile intraday fall from the previous close was 12.70%, against 8.55% close to close. Table 7 counts the days on which BTC moved far enough from the previous close to reach the approximate liquidation distance of a hypothetical isolated position. At 10x, BTC's intraday fall reached it on 57 days, against 23 days on closing prices. Leverage judged against closes understates how often a margin engine acts. The counts describe market moves, not outcomes for any position, and do not show that any level is safe outside the sample.

4.9 Venues price size into margin and exit. Tiered margin is a liquidity-adjusted size limit set by the exchange. On 26 October 2026 the maximum leverage on a single BTC perpetual position fell from between 50x and 100x at $1m to between 3.3x and 20x at $100m (Table 8). Bybit subtracts a fixed deduction from each tier's maintenance requirement (Bybit), so Table 8 shows its effective rate. Deribit's BTC perpetual reports a maximum leverage of 50. Order books set the cost of exit (Table 9). Selling a hypothetical $20m at market would have cost 19 to 44 basis points on the four BTC books in Table 9, 107 to 413 on the three ETH books and 390 to 1,626 on the two SOL books. These are calm-market figures from one snapshot of displayed depth. FTI Consulting reports that BTC's top-of-book depth fell by more than 90% on key venues on 10 October 2025, and suggests stress scenarios in which depth shrinks by 90% for short periods (FTI Consulting). A stressed exit should therefore assume far less displayed depth, and far higher costs, than these figures show. Depth of this kind is one input to the most a strategy can take. How Hedgehog sets that limit is not published.

4.10 Funding gave little warning. On Binance's BTC perpetual from September 2019, 34.8% of eight-hour funding prints sat exactly at 0.01%, and the median print was 0.0094%. In the 30 days before the COVID-19 crash and the May 2021 unwind, mean funding stood at the 80.0th and 76.4th percentiles of its history (against only six months of history for COVID-19). Before the other seven episodes it was below its median, for example at the 35.4th percentile before 10 October 2025. A crowding signal built on funding alone would have missed most shocks.

4.11 Horizon. Multi-day losses are large: BTC's 99th percentile 10-day loss since July 2016 was 26.29% and its 30-day loss 39.27%. In this sample, square-root-of-time scaling of the one-day figure overstated both: the measured 10-day and 30-day figures were 0.837 and 0.722 times the scaled ones. That runs against Danielsson and Zigrand's theoretical result, probably because some of the largest one-day losses partly reversed within days and BTC rose over the sample. We do not read this as evidence that square-root scaling is safe. A limit's horizon should match the time needed to exit, measured directly where possible, as Basel requires for its 10-day base. Timing matters too. Of BTC's worst 1% of UTC days since 2017, 13.9% (5 of 36) fell on a Saturday or Sunday, against a calendar share of 28.6%. In this sample the largest losses fell mostly on weekdays.

5. A framework for the full book

The evidence points to a layered design in which each layer answers a different failure from Section 4. Some layers match Hedgehog's published architecture; others are general recommendations. We mark which is which and do not describe Hedgehog's formulas, parameters or thresholds.

1. Allocation and the most each strategy can take. Published: size is the smaller of the client's allocation and the most the strategy can take. General: a limit of this kind works best when it reflects the depth needed to exit within the position's horizon in stressed conditions.
2. A regime-conditioned tail limit per strategy. Published: EVAR is calculated for each strategy, each has an EVAR-based limit that tightens and widens with the regime, and the volatility book uses a separate regime signal. General: use an ES-type measure for anything added up or optimised, match the horizon to exit, and draw regime inputs from beyond a strategy's own recent volatility.
3. A book-level overlay. Published: EVAR is calculated for the book, accounts for cross-strategy correlation, and the book reduces further when several strategies near their ceilings together. General: divide the book limit by Euler contributions computed with stressed correlations.
4. Hard limits outside the regime estimate. Published: client-chosen leverage from 1x, with exchange margin rules always applying. General: judge leverage against intraday excursions; choose margin mode deliberately; set venue concentration limits, since price-based estimates do not see a venue failing; and keep a stressed figure beside the regime-conditioned one, as the EVAR note recommends.
5. Validation and governance. Published: the team reviews the limits, and no manual step is needed for them to apply. General: count breaches and test for clustering, track model disagreement, damp turnover with bands, and watch for crowding.

The order in which layers apply matters. Hedgehog's published rule for the first layer is that the smaller limit wins. As a general design choice, and not a description of Hedgehog's implementation, no later layer allows more than the layers before it. Within that cap, regime limits can tighten or widen. The binding constraint then shifts with conditions: allocation or depth in calm markets, regime limits after a shock, the book overlay in a correlated sell-off, and leverage and venue limits at all times.

The evidence pulls two ways on timing. Volatility persists and diversification shrinks in falls, so a limit that waits for a lookback window to fill is late. Yet regime estimates err in real time, volatility rules cut on rallies, and shared signals can add to selling. A framework should respond early to rising risk, in measured steps, with a stressed figure in view so that sizing does not depend on current conditions lasting.

6. Implications for allocators

The findings have general implications for professional allocators who run several systematic strategies on their own exchange accounts. They describe market behaviour and risk-measurement methods. They are not advice on any allocation, venue or leverage level.

• Each sub-account is one margin unit. Its strategies share collateral and liquidation rules, so a loss limit on the whole is needed as well as limits on each strategy. Deribit margins each sub-account separately, and on Kraken an account-wide liquidation also closes isolated positions.
• Coins diversify little in sell-offs. ETH and SOL also fell on 93.9% of BTC's worst 5% of days since June 2021, and from 2019 the BTC-ETH correlation on joint falls beyond 1.5 standard deviations was 0.899.
• Regime-conditioned limits move both ways. They cut after shocks, rebuild slowly and can cut after sharp rallies. In the hypothetical exercise, BTC exposure ranged sevenfold. These moves are a property of the method.
• Intraday moves set liquidation risk. Since 2019, BTC moved far enough to reach the approximate liquidation distance of a hypothetical 10x isolated long position on 57 days intraday, against 23 days on closing prices. Venue tiers also reduce the leverage available as positions grow.
• A stressed figure shows what current sizing assumes. BTC's 2022 calibration overstated its 2023 tail by about two times (10.23% against 5.13%), and its 2023 calibration was breached 10 times in 2024.
• When breaches come matters as well as how many. A correct count can hide clustering (Section 4.3).
• Price-based estimates do not see venue risk. Outages, auto-deleveraging and venue failure sit outside them. The EVAR note states that venue, custody and exchange-rule risks need their own limits, and Hedgehog's FAQ notes that many clients spread capital across more than one supported exchange.

Hedgehog's site describes daily reports that carry the relevant readings: the CIO report includes EVAR for the book and for each major strategy type and flags any strategy approaching its EVAR ceiling, and the trader report includes suggested sizes based on current EVAR and the allocation target.

7. Risks and limitations

Limits reduce exposure; they do not remove it. Trading digital assets and derivatives carries significant risk, and investors can lose some or all of the capital allocated. A single session can move beyond any reasonable estimate. Nothing here implies that a regime-conditioned framework avoids losses.

Model and regime risk. Model disagreement widened just after shocks, regime estimates can miss changes or raise false alarms in real time, and volatility forecasts missed by more than half on about one day in ten. The generic models tested here are not EVAR, and their results are not evidence about EVAR either way.

Crowding and venues. Shared limits can create the stress they respond to, as Khandani and Lo argue happened in August 2007. Price-based measures miss outages, auto-deleveraging, local price dislocations, stablecoin de-pegs and venue failure, and margin rules differ by exchange and change without notice.

Data. Regime thresholds use the full sample unless stated. Extreme exceedance correlations rest on 68 to 125 observations. Overlapping multi-day windows are serially dependent. Order books and margin tiers are single snapshots taken on 26 October 2026, after the data cut-off, with truncated books. Intraday data are from one venue, funding covers two venues, DVOL starts in 2021, and BTC's daily close falls three to four hours after the US equity close. The February 2026 drivers rest on one source. All price, funding, order-book and margin data come from venues still operating, and the coins studied are large survivors. Failed venues such as FTX and collapsed tokens such as LUNA are absent, so the sample understates losses from venue and token failure. The hypothetical examples use assumed inputs and contain no returns.

8. Conclusion

A multi-strategy crypto book is one book. The evidence supports four design rules. Allocate risk rather than forecast returns. Condition limits on the regime with inputs beyond recent volatility, because volatility persists but most shocks began when volatility was below its median. Hold a book-level limit built on stressed correlations, because diversification between sleeves shrinks when it is needed. Set leverage, depth and venue limits outside the regime estimate, because margin engines act on intraday paths and order books thin in stress.

Hedgehog's published architecture covers parts of this outline: two size limits of which the smaller wins, EVAR-based limits for each strategy and for the book that move with the regime, a further reduction when several strategies near their ceilings together, and client-chosen leverage within each exchange's margin rules. Its published description does not set out how book EVAR treats correlation in stress, or any venue limits. This paper sets out the evidence and a general framework. It does not describe EVAR's internal method, and it does not show that any framework, Hedgehog's included, prevents losses.

About the data

Hedgehog Capital Research. Market data run from 20 July 2015 to 25 October 2026 (UTC daily closes) unless stated. Daily candles for BTC-USD, ETH-USD and SOL-USD come from the Coinbase Exchange public API, cross-checked against Binance spot daily candles; Coinbase ETH-USD is missing 21 and 22 May 2016, and the return across the gap is assigned to the next day. Intraday paths use Binance spot hourly candles for BTCUSDT and ETHUSDT from 1 January 2019; these files are missing 60 hours across 20 days, none of them a Table 1 episode day. Funding rates are eight-hourly prints on Binance (BTCUSDT, ETHUSDT) and Bybit (BTCUSDT) USDT-margined perpetuals. Implied volatility is Deribit's DVOL index (history from 24 March 2021; statistics from 1 April 2021). The S&P 500 and VIX series come from FRED (S&P 500 to 21 October 2026, so weekly comparisons end on 16 October 2026; VIX to 23 October 2026). Exchange margin tiers (OKX, Bybit, Kraken Futures, Deribit) and order books (Coinbase, Kraken, Binance, Deribit) are single snapshots taken from public APIs on 26 October 2026 at 14:38 and 14:37 UTC. All endpoints are public and keyless. All price, funding, order-book and margin data come from venues still operating, and the coins studied are large survivors; failed venues such as FTX and collapsed tokens such as LUNA are absent, so the sample understates losses from venue and token failure. Returns are daily log returns of UTC closes, losses are simple-return losses, and volatility is annualised by the square root of 365; forecasts for day t use data to the close of day t-1. Loss-model breach counts apply generic textbook models (historical simulation, RiskMetrics EWMA, filtered historical simulation) to market prices; none is EVAR. The volatility-targeting exposures, the three-sleeve book and the leverage thresholds are hypothetical and use no return inputs. No strategy returns, performance or trading backtests are computed or shown; the breach counts in Section 4.3 are standard backtests of generic VaR models on market prices. The fetch script, raw data files and analysis script are retained by Hedgehog Capital Research, and every statistic not attributed to a cited source is reproducible from them.