In a 22-year backtest, the most instructive number is the one that looks like a failure. Cyclical value posted an information ratio of -0.20 from 2004 through May 2026. It sits in the same multifactor portfolio that produced an information ratio of 1.00, the best figure in the study. The intuitive reading of "diversification all the way down" is to collect good things and spread the bets. The mechanism Renshaw, Gu, Candar and Ang describe works differently. Diversification pays through weak correlation between sleeves, so a component's standalone record is a poor guide to whether it belongs. The layered design (signals within factors, factors across the portfolio, timing models across time) applies that one rule three times.
The headline also hides a problem of scale. The combined portfolio beat its best sleeve, enhanced value at 0.94, by only six hundredths of a point in information ratio. If the case for combining rested on beating the winner, it would be a thin case. The case is stronger than that because nobody in 2004 knew which sleeve would win.
Correlation Is the Engine, Not the Sleeves
The authors state the mechanism directly: the factor sleeves were weakly correlated, and that low correlation drove the combined result. The full-sample information ratios were enhanced value 0.94, quality 0.73, momentum 0.65 and cyclical value -0.20. Averaged naively, those figures do not add up to 1.00. The combination can exceed its average, and even its best member, only because the active returns zig and zag at different times. Each sleeve's active risk partly cancels the others', so active return accumulates faster than active volatility.
Here intuition goes wrong. Investors tend to judge each component on its own merits and remove the laggards. In a combined portfolio, a sleeve's contribution depends on its standalone payoff and on its covariance with everything else. Value and momentum are the classic example. Their active returns have often moved in opposite directions, so a value sleeve can dampen momentum's drawdowns even in periods when value earns little on its own. Under the arithmetic of portfolio construction, a sleeve with a negative standalone information ratio can still raise the portfolio's information ratio if its correlation with the rest is negative enough.
The excerpt does not say whether cyclical value cleared that bar in this study. It does not report the combined portfolio's information ratio with cyclical value removed, so the claim that the losing sleeve helped is an inference from the design. The authors' choice to keep it, at a balanced weight and through a twenty-year drought for traditional value, signals that they treat it as a diversifier first and a return source second.
The authors also note that factor leadership rotated across subperiods, and that the weakest factor in one window was often the strongest in another. That rotation is the empirical counterpart of the low-correlation result. If the sleeves did not take turns, there would be nothing to diversify.
Two Kinds of Value, Two Different Bets
The within-factor layer is easiest to see in how the paper splits value into two sleeves. Cyclical value uses the traditional ratios: book-to-price and earnings yield. Enhanced value uses measures tied to economic profit, intangible capital and free cash flow. The authors found the two behaved differently, and enhanced value held up better during the long stretches when traditional value struggled.
The split addresses a measurement problem. Book value records a company's physical and financial assets well but much of its intangible investment poorly. Spending on research, brand and software usually runs through the income statement instead of building book equity. A business that invests heavily in intangibles can therefore look expensive on book-to-price while being reasonably priced against the capital it has actually built. Cyclical value systematically underweights such firms. Enhanced value is designed to see them.
These are two different bets. Cyclical value is largely a wager on mean reversion in cheap, asset-heavy businesses, the kind that rally when the economic cycle turns. Enhanced value leans toward businesses earning returns above their cost of capital that the market still prices modestly. When the authors say the definitions are "not interchangeable," they are describing two distinct risk exposures that share a label.
Momentum gets the same treatment. Instead of a single price-trend signal, the sleeve blends idiosyncratic momentum, short interest, hedge fund positioning, analyst revisions and economic-profit momentum. Price momentum on its own is known for sharp reversals when markets turn violently. Revisions and positioning data reflect different information and need not reverse at the same time, so the blend is meant to stop any one signal's failure from becoming the sleeve's failure. Quality uses a composite as well, the Axioma US5.1 Profit Quality factor, which combines profitability, cash-flow strength, operating efficiency and accruals.
What a 1% Tracking-Error Budget Decides
Most summaries treat the constraints as footnotes. They actually determine which factors the portfolio can hold.
The portfolio targets 1% ex ante tracking error against the STOXX US Universal Index. It rebalances quarterly, caps one-way turnover at 5% per quarter, and applies concentration and industry limits. The authors excluded size and low volatility for a specific reason. Small stocks carry betas well above 1 and low-volatility stocks carry betas well below 1, so either factor would have brought in more unintended market exposure than a 1% budget can absorb. Neither was excluded because its premium looked weak. The risk budget decided the factor menu.
Read that way, "diversification all the way down" has a boundary. The portfolio diversifies across sources of active return that fit inside a tight, roughly beta-neutral envelope. It does not diversify across every documented premium. An investor running a looser mandate faces a different menu.
The budget also gives the headline number a scale. An information ratio of 1.00 at a 1% tracking-error target implies active return of roughly one percentage point a year, assuming realized tracking error stays close to target. The excerpt does not report realized tracking error, so that conversion is approximate. At that scale, implementation decides a lot. The results are hypothetical and exclude management and trading fees. The 5% turnover cap limits trading costs, but a fee of 30 or 40 basis points would take a large share of a 100-basis-point active return before any slippage. The paper's information ratio describes the signal. Whether an investor captures it depends on what they pay to hold it.
Timing as a Fourth Diversifier
The timing layer is the most restrained part of the design, and that restraint is the main finding. Three timing models built on different information sets each propose tilts. Each model can at most double or halve a factor's nominal weight, and none can remove a factor entirely. The portfolio then averages the three proposals.
Averaging bounded tilts compresses them a second time. Unless all three models agree on the extreme tilt in the same quarter, the realized deviation from balanced weights is smaller than any single model's call. The ensemble effectively treats timing as one more diversifiable source of noise, rather than a forecast the portfolio commits to.
The results back that caution. The overlay improved performance over the most recent three years but added little over the full sample. A three-year window holds only twelve quarterly rebalances, too few to separate skill from a favorable stretch. The fair reading is that timing did no harm, sometimes helped, and was never needed for the core result. The authors cite Cliff Asness's advice to "sin a little," and their structure puts it into practice: the overlay is too small to matter much if it is wrong.
This matters for allocators because timing is where most factor programs sometimes lose their diversification. A manager with conviction about value's rebound, or quality's defensiveness, can quietly turn a four-factor portfolio into a one-factor bet. The double-or-halve bound and the three-model average are built to block that drift.
Where the Layering Could Be Overfit
The strongest objection concerns when the research was done. Enhanced value, built on economic profit and intangible capital, belongs to a family of measures that became popular largely because traditional value performed so poorly in the decade after the financial crisis. A definition developed partly in response to that period will, almost by design, look resilient when tested over it. The finding that enhanced value held up while cyclical value struggled is real in the data. It is also partly what the definition was built to produce.
Each extra layer adds choices for the researcher: which signals enter each composite, how they are weighted, which three timing models make up the ensemble, and which bounds apply. Every choice made with the full 2004 to 2026 history available is a possible route for fitting to the past. Layered diversification lowers dependence on any single choice, but the menu of choices is itself a decision made with hindsight.
Correlation stability is the second risk. The whole mechanism depends on weak correlation between sleeves. Factor correlations have shifted before, and in liquidation episodes, when crowded positions unwind together, sleeves that looked independent can fall together. A portfolio built on low correlation is only as robust as the correlation estimate.
The design has an answer to both objections, which is why the thesis survives them. Keeping cyclical value despite its negative record is the opposite of overfitting: a data-mined portfolio would have dropped it. The tight bounds on timing also limit how much any fitted timing model can hurt live results.
The observable test is the correlation between cyclical value and enhanced value out of sample. If the two sleeves start moving together, the within-factor layer was a relabeling exercise, and the information-ratio edge over enhanced value alone should disappear. If they stay distinct through the next value rally, as enhanced value lags while cheap cyclicals lead, the layering holds up in exactly the period it was designed for. For anyone implementing the approach, the second variable is cost. At roughly one percentage point of hypothetical active return, the strategy works only if fees stay well below that edge.
