Universal Portfolios

Thomas M. Cover, Departments of Statistics and Electrical Engineering, Stanford University. Mathematical Finance 1(1), January 1991, pp. 1–29: the first issue of the journal. Manuscript received January 1990, final revision July 1990; NSF grant NCR 89-14538, with stock-market data credited to Hal Stern. Read from the author's own Stanford copy, a 29-page scan with no text layer (page images read directly).

The paper that gives this spoke a result of a kind it did not hold: a portfolio algorithm with a proof attached instead of a backtest. Everything else here earns its performance number by replaying a history; Cover earns his by refusing to assume the history has any statistical structure at all.

The refusal the paper is built on

"Throughout the paper we are unwilling to make any statistical assumption about the behavior of the market. In particular, we allow for the possibility of market crashes such as those occurring in 1929 and 1987. We seek a robust procedure with respect to the arbitrary market sequences that occur in the real world."

No distribution, no stationarity, no ergodicity, not even randomness. The price-relative sequence x₁, x₂, … is arbitrary, and Cover states plainly that "a malicious nature can structure future xₖ's to take advantage of past beliefs." That premise is what separates this from the model-based tradition it cites — Markowitz (1952) and Sharpe (1963) in the mean–variance plane, Kelly (1956) and Thorp (1971) on rebalancing under a known distribution — and Cover names the departure himself: the goal "depends solely on the data and does not depend upon underlying statistical assumptions."

The benchmark, and why it is the interesting one

A constant-rebalanced-portfolio b holds fixed wealth proportions and rebalances every day, achieving Sₙ(b) = ∏ᵢ bᵗxᵢ. The paper's target is

S*ₙ = maxb Sₙ(b)

— the best constant rebalanced portfolio chosen after the outcomes are revealed. This is a hindsight benchmark and a deliberately hard one: S*ₙ exceeds the best single stock, the Dow Jones average and the value line index, "usually by an exponential factor."

The algorithm

The universal-portfolio is a performance-weighted average over the whole simplex of portfolios: start uniform, b̂₁ = (1/m, …, 1/m), then

b̂ₖ₊₁ = ∫ b Sₖ(b) db / ∫ Sₖ(b) db

Lemma 2.5 telescopes the resulting wealth into Ŝₙ = ∫ Sₙ(b) db / ∫ db. The universal wealth is simply the average of every constant rebalanced portfolio's wealth over the simplex. Cover's own reading (§3): hand db/∫db of your capital to each portfolio manager indexed by b, let each run their rebalancing strategy, and repool at the end. All of it done on paper.

What is proved

• First order, arbitrary sequences. (1/n)ln Ŝₙ − (1/n)ln S*ₙ → 0 for any bounded sequence. The universal portfolio has the same exponential growth rate as the unachievable hindsight optimum.
• A bound for every finite n (Thm 5.1, two stocks), not just in the limit: Ŝₙ/Sₙ ≳ √(2π/nJₙ), where Jₙ is the curvature of ln Sₙ(b) at its maximum and τₙ the relative range of the price relatives. Cover's own gloss: the universal wealth is within a factor of C/√n of the exponentially large Sₙ, and "every additional stock in the universal portfolio costs an additional factor of 1/√n."
• General m (§9): Ŝₙ/S*ₙ ~ ((k−1)!/(2ᵐ−1))(2π/n)^{(k−1)/2}|Jₙ|^{−1/2} over the k active stocks. The cost of universality is polynomial; the wealth being tracked is exponential, so the ratio vanishes in the exponent.
• Two free dominance results. Ŝₙ ≥ (∏ⱼ Sₙ(eⱼ))^{1/m} — the universal wealth beats the geometric mean of the constituent stocks (the value line index) for every sequence (Prop 2.5). And Ŝₙ is invariant under permutations of the sequence (Prop 2.6), which Cover reads as: "the crash of 1929 will have no worse consequences for wealth Ŝₙ than if the bad days of that time had been sprinkled out among the good."
• The i.i.d. special case (Thm 7.1): if the market does happen to be i.i.d. ~ F, then (1/n)ln Ŝₙ → W*(F). The algorithm ends up performing as if F had been known. It costs nothing to drop the assumption.

The examples, and what they are worth

§8 runs the algorithm on daily NYSE closes over a 22-year period ending in 1985 (dividend-adjusted; the figure axes are labelled "20 Year Return", which the text does not reconcile). Integrals are quantized to the 21-point grid b = i/20.

The Iroquois/Kin Ark row is the whole argument in one line: buying and holding the better of the two stocks with perfect foresight returns 791%, and a fixed 55/45 mix rebalanced daily returns 7,262%. The gain does not come from prediction. It comes from rebalancing across volatile, uncorrelated series, which is why IBM and Coca-Cola, moving in lockstep, yield almost nothing.

A fourth example crudely models 50% margin: four assets (Commercial Metals, Kin Ark, and each on margin at 6%/year, r = 0.000233/day). Both stocks bought on full margin are ruinous on their own — Kin Ark on margin returns 0.0000 to four significant digits, yet the best rebalanced mix (0.2, 0.5, 0.1, 0.2) reaches S*ₙ = 262.4021 and the universal portfolio reaches Ŝₙ = 98.4240, above the 78.47 it achieves without leverage available.

These pairs were chosen, and Cover says so — Iroquois Brands and Kin Ark are "two stocks chosen for their volatility," picked in 1990 from a history ending in 1985. The theorem is assumption-free; the demonstration is a hand-picked pair over one fixed window, which is the selection pattern backtest-overfitting is about. Nothing is wrong with the paper here, since a worst-case theorem does not need a representative sample, but the numbers in §8 carry none of the theorem's guarantee, and no count of how many pairs were examined is given.

What the paper admits

§10 is unusually direct about the gap between the theorem and practice:

• Trading costs are ignored outright. "In practice we would not trade daily, but only when the current empirical holdings were far enough from the recommended b̂ₖ." The proposed rule of thumb (trade only if the increase in W exceeds the logarithm of the normalized transaction costs) is offered as a suggestion, not a result. A daily-rebalanced strategy over 5,600 trading days is precisely where costs bite.
• The case rests on accepting the benchmark. "The case for success rests almost entirely on the acceptance of S*ₙ as a natural investment goal." Beating the best hindsight constant rebalanced portfolio says nothing about beating a strategy outside that class.
• The guarantee is relative, not absolute. Ŝₙ tracks S*ₙ; if the best constant rebalanced portfolio loses money, so does the universal one. Nothing here is a claim about making money.

Related

universal-portfolio · constant-rebalanced-portfolio · thomas-cover · mathematical-finance-journal · stanford-university · backtest-overfitting · strategy-optimization · backtesting · algorithmic-trading · black-scholes · synthesis