Restrict the basket to the 25 cards that reprice each month and Sharpe drops from 1.60 to 1.11. CAGR scarcely changes, at 20.00% against 20.82%. This filter is the authors' own sensitivity test, and it carries more weight than anything else in the paper.
Pompei and Nonino construct a monthly price index for 117 high-value Magic: The Gathering cards covering October 2010 to March 2025. Their aggregated marketplace data come from MTGGoldfish, which draws transactions from TCGPlayer, Card Kingdom and eBay. They treat the cards as an emotional asset and claim diversification benefits within a multi-asset portfolio: Sharpe rises from 0.86 to 1.26, while drawdown improves from 18.13% to 13.29%.
Scarcity and use value supply the economic case. The cards produce no cash flows. Prices instead depend on playability across formats, print runs and survival rates, alongside a rarity premium that the paper bases on rank among close substitutes rather than absolute quantity. The paper says the majority of selected cards are on the Reserved List, the issuer's standing commitment not to reprint specific cards. Their supply is therefore credibly fixed. With no distributions, price appreciation provides the only cash return.
Each card receives an equal weight of 1/117. Monthly average closing prices are converted to log returns, and the index begins at 100. The authors also construct a volatility-weighted version, then put it aside because low measured volatility partly proxies for an absence of trading. They acknowledge that problem and choose the equal-weighted index as the conservative version.
Against five iShares proxies covering the S&P 500, US small-cap, long and short Treasuries, and gold, the index records an annualized mean log return of 18.73% (t = 5.48) with volatility of 13.01%. Sharpe is 1.44 at rf = 0. Max drawdown is minus 28.53%. Correlations with conventional assets range from minus 0.14 to plus 0.16.
Beta to the S&P 500 is 0.07, with R² of 0.01, while annualized alpha reaches 17.84% (t = 5.06). A Fama-French three-factor regression using HAC errors reports market beta of 0.02, R² = 0.05 and alpha of 17.8% (t = 4.09). The paper says the full output was omitted to save space, leaving the SMB and HML loadings unavailable for checking.
Adding the cards to an equal-weight six-asset portfolio reduces volatility from 7.97% to 7.09% and raises mean log return from 6.82% to 8.97%. Max drawdown improves from minus 18.13% to minus 13.29%. Across twelve-month forward rolling returns, the median rises to 9.63% from 6.86%, while the worst window improves from minus 16.87% to minus 12.70%.
Smoothed marks set the risk estimate
The underlying observations are monthly average closing prices aggregated across marketplaces. Volume, turnover and holding periods cannot be observed, as the paper acknowledges. Average monthly coverage across the basket is 87.12%, and average staleness is 14.57%. Roughly one month in seven therefore shows no price change.
The authors fill gaps by linearly interpolating log returns and exclude series missing more than half their months. Creating a continuous series from a discontinuous market can be defended, and they openly concede that residual appraisal smoothing remains possible. The resulting 13.01% volatility and near-zero correlations describe a smoothed valuation series. We did not find a lagged-beta or Dimson-style correction for nonsynchronous pricing. Such a correction is standard when low correlation may reflect stale pricing rather than economic independence.
Then comes the test that matters. Requiring coverage of at least 90% and staleness of at most 3% cuts the universe to 25 cards. CAGR holds at 20.00%, compared with 20.82% for the full 117. Annualized volatility, however, climbs from 13.03% to 17.98%, and Sharpe declines from 1.60 to 1.11.
Smoothing changes the measured risk while leaving the return broadly intact. The paper therefore concludes that performance "is not an artifact of illiquidity-induced smoothing", and I accept that conclusion for the return level. The filter weakens the risk-adjusted headline. A volatility estimate of 13.01% looks like a floor; 17.98% volatility and a Sharpe of 1.11 give the more credible account.
A universe chosen with hindsight
Cards qualified for the sample when their market value exceeded USD 1,000 at any point during the study period, or when the authors judged them historically and competitively relevant. Applying the first criterion across the complete window filters on the outcome variable, so selection enters the index mechanically.
The paper describes the basket openly as a blue-chip proxy and says the findings do not extend down the price distribution. It captures about 92% of the current market value in the relevant sets, USD 552,700 out of roughly USD 600,000. Fine. An index made from assets that became expensive at some point over fourteen years still cannot estimate the return available to a 2010 buyer choosing high-end cards then.
The authors say the split from traditional assets appears during and after COVID, describing a structural break in relative performance. As a result, the 17.84% alpha with a t of 5.06 across 174 months depends heavily on one repricing episode. The 12-month forward distributions also use overlapping windows. In independent, non-overlapping terms, the evidence amounts to roughly fourteen draws rather than 174.
Costs stay outside the reported returns
The paper sets out three bands of round-trip costs. Total frictions run from 40 to 50% on a $10 card before fees, fall to 12 to 20% around $50, and settle near 10 to 12% above $200. It explicitly warns that realized net returns can differ sharply from observed price appreciation. Yet the index and portfolio include none of those costs, and I did not find a net-of-cost simulation for either one.
An equal-weighted 117-card index requires rebalancing by construction. Even the cheapest part of the paper's schedule, 10 to 12% for cards above $200, consumes more than half of one 18.73% gross year.
Sharpe is calculated at rf = 0 over a period that includes 2022-2025 positive real rates. Short Treasuries averaged 1.12% annualized over the full Oct 2010 to Mar 2025 sample. That average differs from the 2022-2025 rate environment implicitly treated as rf = 0.
The tested allocation is 1/6, close to 16.7%, while the abstract extrapolates from that single weight to moderate exposure more generally. I did not find a sweep over allocation sizes. The omission matters because the market lacks a low-cost vehicle and standardized custody, individual-card drawdowns frequently exceed minus 60%, and reprint policy remains at the issuer's discretion. A 16.7% allocation to these cards does not carry the same exposure as 16.7% in gold.
The conclusion concedes much of the investability problem. For purely financial investors seeking passive exposure, the authors write that the absence of low-cost index vehicles, standardized custody and deep institutional liquidity creates a material constraint. They identify the collector-investor as the most natural owner. That narrows the case to buyers who also consume the asset, although the portfolio improvement still relies on gross quoted marks. The friction schedule says those marks cannot be realized.
Where the stronger returns appear
The highest CAGRs in the full dataset come from non-Reserved-List foils and cards with less extreme initial prices. Shivan Dragon [7E] foil records 53% CAGR, 168.7% volatility and an 88.3% drawdown. Its history begins in March 2014 rather than at the start of the full sample.
The canonical trophies all have complete October 2010 histories and sit near the bottom: Time Walk [LEB] at 11.3%, Mox Ruby [LEB] at 11.5%, and Tropical Island [LEB] at 11.1%. The authors themselves describe the pattern as a size or base-price effect, with high starting prices compressing percentage upside. A buyer can act on that observation without accepting the index arithmetic.
A repeat-sales or transaction-level index for the same basket would change my view if it applied realized bid-ask spreads and dealer discounts, then produced a Sharpe above 1 net. Until then, the defensible conclusion remains narrower than the headline. High-end Magic cards appreciated quickly over 2010-2025, with most of the move occurring after 2020. The authors place the result beside LEGO at 10 to 11% and digital collectibles at 18.6 to 41%, rather than art at 1.90 to 4.22%. That comparison fits. The return level survives the liquidity filter. The risk-adjusted figure does not, especially for a universe selected because its members crossed USD 1,000 at some point during the window.
We could not test this ourselves. We do not hold card-level Magic price histories or bid-ask, volume, condition or grading data for the cards. Replacing them with equities or commodity ETFs would remove the illiquid secondary-market structure on which the finding depends. We have flagged the same gross-of-cost pattern before in turnover-limited state-price-density portfolios, where the reported edge was 0.4 to 2.7 points with no transaction costs at all.