Returns-Based Style Analysis: The Copycat Portfolio Test

Definition

Alpha analysis — known formally as *returns-based style analysis* — answers one blunt question: was this fund worth owning, compared to just holding what it's secretly made of? It judges a fund by trying to *copy* it. Instead of reading the fund's marketing material, you look only at its day-to-day total returns and ask: could a simple homemade portfolio of cheap, well-known index ETFs have produced almost exactly the same returns?

Here is the whole idea, explained like you're five.

Imagine every fund is a smoothie someone is selling you. The manager says it is a special smoothie, worth paying extra for. But you have a shelf of basic, cheap ingredients at home:

• The whole stock market — an S&P 500 fund like SPY
• Big tech — a Nasdaq-100 fund like QQQ
• Sector funds — banana, strawberry, mango (energy, healthcare, tech...)
• Gold, silver, bitcoin — the exotic ingredients
• Plain water — cash, in the form of a T-bill fund like SGOV: safe, boring, tiny return

The test: keep adjusting a homemade mix — say 70% big tech, 25% semiconductors, 5% water — until it rises and falls, day after day, as closely as possible to the fancy smoothie. That best copycat mix is called the best-fitting benchmark portfolio (or the fund's *style*). The tool that finds it is a regression — an automated taste-tester that tries recipes until the match is as close as it can get.

Once you have the best copy, whatever difference remains is the interesting part:

• Fund earned the same as the copy → the manager added nothing you could not make yourself. Alpha is zero.
• Fund earned more → there is skill (or luck) the cheap ingredients cannot explain. Positive alpha.
• Fund earned less → you paid up for a smoothie that is worse than homemade. Negative alpha.

The method was introduced by Nobel laureate William Sharpe in 1992, and professional analysts have used it ever since precisely because it needs nothing but returns — no holdings files, no trust in the fact sheet. Read Sharpe's original returns-based style analysis paper.

How This Differs from Jensen's Alpha

Jensen's alpha uses a single market-risk model: it asks whether a fund outperformed the return implied by its beta. Dividend Vision's style analysis uses a constrained, multi-asset replication model: it asks whether a blend of stock, sector, bond, commodity, crypto, and cash benchmarks could have reproduced the fund. The first is one-factor attribution; the second is a copycat-portfolio test.

Why It Matters

Fund names and fact sheets describe what a fund *says* it does. Style analysis reveals what the fund actually *behaves like* — and for income investors the two are often different in ways that matter:

• It exposes closet indexing. A fund charging a high expense ratio whose copycat recipe turns out to be "95% SPY, 5% cash" is an expensive way to own the index. The recipe makes that visible in one line.
• It detects hidden leverage. This is the clever twist in the recipe rule: every ingredient amount must be zero or positive — you cannot put negative bananas in a smoothie — except cash, which is allowed to go negative. Negative cash means borrowed money. A fund whose best-fit recipe reads "200% QQQ minus 100% cash" is running two-times leverage, whether or not the word "leveraged" appears in its name. The negative cash weight *is* the loan.
• It gives context to yield. A double-digit distribution rate tells you how a fund pays, not how it performs. Comparing the fund against its own best-fit recipe answers the question that matters: after accounting for what this fund is actually made of, did it add value or subtract it?
• It is honest about what it cannot see. The quality of the copy is measured by R-squared — how much of the fund's movement the recipe explains — and by tracking error, how far the copy drifts from the real thing. A low R² is the method's own way of saying "this fund does something my ingredient shelf cannot replicate, so read the alpha with caution."

Example

All numbers here are illustrative. Suppose the taste-tester runs on three funds using daily total returns over two years:

Reading each row in plain English:

1. Premium Tech Income behaved like a tech-tilted portfolio with a dash of borrowing. After matching that recipe, it *still* earned 2.7% a year more than the homemade copy — unexplained outperformance worth investigating, which could be skill, an ingredient the shelf lacks, or luck.
2. Steady Growth Select is a closet indexer: 95% of it is just the S&P 500. The negative alpha of −0.9% matches its fee almost exactly — consistent with paying a chef to pour you a SPY.
3. Turbo Nasdaq 2x shows the leverage detector at work. The recipe holds 200% QQQ financed by −100% cash: every dollar invested controls two dollars of Nasdaq exposure, with the borrowing cost (about the T-bill rate) showing up as the negative cash leg. Its −1.1% alpha is consistent with the drag of fees and financing — typical for leveraged products, and exactly what the method should find.

Notice what made the leverage visible: the sum of the weights must equal 100%, and only cash may go below zero. Without those two rules, the math could just say "2x QQQ" without admitting that the extra exposure is bought with borrowed money.

What It Catches in the Real World

Five situations where the copycat test tells you something no other number on a fact sheet will. The fund names and figures are illustrative, but each pattern shows up regularly in real income funds:

1. The expensive closet indexer. "Blue Chip Select" charges a 0.85% expense ratio and markets an active stock-picking process. Its recipe: 95% SPY, 5% cash, R² 0.99, alpha −0.9% a year — almost exactly its fee. The manager isn't doing anything an index fund doesn't do; you are paying a chef to pour you a glass of SPY. One line of analysis, and the fee conversation is over.
2. The yield that isn't income. "MegaYield Weekly" advertises a 45% distribution rate and pays like clockwork. Its alpha: −15% a year. The distribution rate tells you how you're *paid*; alpha asks whether there is anything real behind the payments. A giant yield sitting next to deeply negative alpha means the fund is steadily losing ground to its own ingredients — however the payments are sourced, total return is not keeping up. That is the classic yield trap signature, flagged — not proven — in one number. (See also why high yield isn't high income.)
3. The hidden borrower. "Preferred Income Plus" never uses the word "leveraged" in its name or marketing. Its recipe: 60% preferred stocks plus another 85% across credit funds, financed by −45% cash. The negative cash leg flagged the borrowing anyway — a model inference from price behavior, and one that would explain why the fund falls harder than plain preferred-stock funds in bad months. Price behavior raises the flag; page 47 of the prospectus is where you confirm it.
4. Picking within a family of lookalikes. Ten covered-call funds on the same index all advertise similar double-digit yields, and their fact sheets are nearly interchangeable. Measured against the *same* copycat recipe, one shows +2% alpha, most sit near zero, and two show −4%. Same strategy, same benchmark, honest scoreboard — the fair way to choose among covered-call ETFs that all "pay well."
5. Knowing what you actually own. "Dividend Select 80" sounds like a broad market fund. Its recipe: 35% consumer staples, 20% utilities, 15% real estate, and barely any technology. Nothing is wrong with that mix — but it explains in advance why the fund is likely to lag a tech rally, so the holder who checked the recipe isn't surprised (or panicked into selling) when it happens.

On Dividend Vision, this analysis runs nightly for a curated subset of the universe — see the FAQ below for exactly which funds. Look for the Alpha Analysis card on a fund's ticker page, or enable the Alpha column in the screener to sort the analysed funds directly. Read every number on that card as a model estimate: alpha is the residual return the published method below cannot explain, not an audited decomposition of the fund's holdings, borrowings, or the source of its payouts.

The Ingredient Shelf

Transparency matters more than cleverness here, so this is the complete shelf — all 148 benchmark ingredients the taste-tester is allowed to build copies from, grouped the way the model groups them. The philosophy: plain index building blocks only, favoring the biggest fund in each category, so a recipe is always something you could actually buy. Crypto funds get matched against crypto ingredients, sector funds against sector ingredients — and cash (SGOV) is the one ingredient allowed to go negative, which is how borrowing is detected.

The option-overlay row deserves a note: those are mechanical trackers of published option indexes (the CBOE BuyWrite and PutWrite families), with full- and half-coverage versions on the shelf so the taste-tester can express in-between degrees of call-writing by mixing them with the plain index. They exist so a covered-call fund's recipe can *name* its option overlay instead of imitating it with watered-down index positions.

The single-stock row has stricter house rules than any other. Those shares exist for one class of fund only: single-stock option-income wrappers (the YieldMax, Kurv and REX families — MSTY, NVDY, TSLY, CONY and kin), whose one honest benchmark is the underlying share itself. The question their buyers are asking is "did this fund beat just holding the stock?", and no basket of ETFs can stand in for that. So a stock is admitted to a fund's shelf only when the fund tracks it at identity-level correlation, at most one stock may appear in any recipe, and a stock leg still has to earn its keep under the same out-of-sample tests as any narrow ingredient. A diversified fund never sees a stock on its shelf at all — its recipes remain built from index ETFs alone.

Two kinds of funds are deliberately excluded from the shelf. Actively managed strategy funds (JEPI, JEPQ and their kin) stay off it — a fund should be explained by what it is built from and the published indexes it resembles, not by a competitor's manager. And dividend-factor ETFs are excluded because they are *subjects* of this analysis: using them as ingredients would define away the very alpha the test exists to measure. The shelf is versioned and evolves — when a recipe looks wrong, the cause is usually a missing ingredient, and the fix is stocking it.

How the Math Works

For readers who want the machinery, not just the smoothie metaphor. Everything below runs nightly on daily total returns (adjusted close, so every distribution counts, reinvested on its ex-date) over a trailing window of about two trading years.

The fit. Find the benchmark weights *w* that make the recipe track the fund as closely as possible:

minimize   mean over days t of  ( r_fund,t − Σi wi · r_i,t )²
subject to Σi wi = 1
           wi ≥ 0 for every benchmark
           w_cash may be negative (that's borrowing)

This is Sharpe's (1992) constrained returns-based style analysis. The weights-sum-to-one and no-negative-ingredients rules make the recipe a real, buildable portfolio — and force leverage to show itself as a negative cash weight instead of hiding inside inflated weights.

Alpha. With the best-fit weights in hand, alpha is the average daily gap between the fund and its recipe, annualized:

alpha_daily  = mean( r_fund − r_recipe )
alpha_annual = alpha_daily × 252

Fit quality. Two numbers describe how good the copy is:

R²             = 1 − Var(residual) / Var(fund)
tracking error = StdDev( r_fund − r_recipe ) × √252

Is the alpha real? Daily returns are autocorrelated, so the standard error of alpha uses a Newey–West (autocorrelation-robust) estimate, written *se*:

t-statistic = alpha / se          (|t| ≥ 2 ≈ statistically significant)
95% range   = alpha ± 1.96 × se
info ratio  = alpha_annual / tracking_error_annual

Out-of-sample check. The recipe is fitted on the first ~80% of the window only; alpha is then re-measured on the final ~20% the fit never saw. When the two disagree sharply, the in-sample alpha was likely luck.

Rolling persistence. The recipe is held fixed and alpha is re-measured over trailing 6-, 12- and 24-month windows stepped monthly — that is the strip chart on the card, and the "positive in N of M windows" count. For funds older than the two-year fit window, this long lens reaches back up to five years (as far as the fund and its recipe ingredients both have history), so an older fund's persistence record can span dozens of windows. Holding the recipe fixed matters twice over: refitting each window would blur recipe *changes* into alpha *persistence* — and applied to older data, today's recipe makes a strategy shift visible as early windows that disagree with recent ones.

Guardrails. A handful of safeguards keep the recipes honest rather than merely well-fitted: the wide, collinear shelf is solved with a ridge penalty chosen by walk-forward cross-validation, but the penalty is used only to *select* ingredients — published weights come from an unpenalized refit, so real leverage is never shrunk away. Tiny weights are pruned, and any benchmark whose removal barely hurts the fit is eliminated, keeping recipes to a handful of legs. A negative cash leg must *earn* its place by improving the fit in proportion to its size, or the unlevered recipe is published instead. And on short histories the model gets deliberately humbler: until a fund has ~10 months of data, a narrow subsector ingredient stays in its recipe only when it carries a large share of the fit — the signature of a fund that *is* that subsector — because below that history length a noise-chasing side leg and a genuine small holding are statistically indistinguishable.

Common Mistakes

• Treating all leftover return as manager skill. Alpha is the *unexplained* return — which includes luck, and includes strategies the ingredient shelf cannot copy. Covered-call funds are the classic case: selling options produces return patterns no mix of plain index ETFs can replicate, so part of their "alpha" (positive or negative) is really option-strategy exposure, not skill.
• Ignoring the R². A recipe that explains 40% of a fund's movement is not much of a copy, and the alpha computed from it is mostly noise. Trust the alpha in proportion to the R-squared next to it.
• Using price returns instead of total returns. For dividend and income funds, distributions are most of the story. Running the test on price alone manufactures fake negative alpha. Always use total return.
• Reading the recipe as actual holdings. Style analysis describes what a fund *behaved like*, not what it literally owns. A fund can hold no Apple shares and still fit a tech-heavy recipe because its holdings move with tech.
• Judging on a short window. A few months of daily returns can fit a flattering recipe by accident. Longer windows and out-of-sample checks (does the recipe still fit on data it was not tuned on?) separate durable behavior from coincidence.

FAQ

How is this different from just asking "did it beat the S&P 500?"

"Did it beat SPY?" compares every fund against the same yardstick, no matter what the fund actually is — which makes the comparison unfair in both directions. A consumer-staples dividend fund that trails SPY in a tech-led rally isn't failing; it was never built to hold what SPY holds. A 2x leveraged fund that beats SPY in a bull market isn't showing skill; it borrowed money to double an index. And for a bond or preferred-stock fund, the SPY comparison is close to meaningless — they're different asset classes entirely. Alpha analysis fixes this by building each fund a custom yardstick from its own behavior: the copycat recipe. The staples fund gets measured against a staples-heavy mix, the leveraged fund against a borrowed 2x mix, the preferred fund against preferreds and credit. Beating *your own recipe* is the version of "did it beat the market?" that is actually fair — a fund can trail SPY by miles and still show positive alpha (it beat what it really is), or beat SPY handily and show negative alpha (its ingredients did the work, minus a fee). The single-benchmark, risk-scaled version of this idea is Jensen's alpha; this analysis is the many-benchmark generalization.

Is this the same alpha as Jensen's alpha?

Same idea, bigger ingredient shelf. Jensen's alpha compares a fund against a single benchmark scaled by its beta. Returns-based style analysis compares it against the best-fitting *combination* of many benchmarks — market, sectors, commodities, bonds, cash — so less gets misattributed to skill when a fund simply holds an unusual mix of exposures.

Why is cash the only ingredient allowed to be negative?

Because a negative cash position has a real-world meaning: borrowed money. A fund cannot meaningfully hold "negative healthcare sector" in this framework, but it can absolutely borrow at roughly the T-bill rate to buy more of everything else — that is what leverage is. Letting only cash go negative lets the math detect leverage while keeping every other weight interpretable as a simple allocation.

What does the sum-of-weights rule do?

Forcing the weights to add up to 100% makes the recipe a real portfolio — one you could actually build. It is also what makes leverage show up honestly: to hold 200% of QQQ, the recipe *must* book −100% cash to balance, rather than quietly claiming double exposure from nowhere.

Can a fund with a great yield still have negative alpha?

Yes, easily. Yield describes how return is *paid out*, not how much return there is. A high-yield fund whose share price erodes can sit far below its copycat recipe — deeply negative alpha — while a modest-yield fund beats its recipe year after year. That is exactly why this test is useful to income investors.

What do the "95% range" and "Info ratio" numbers mean?

Both answer the same follow-up question: is this alpha real, or a lucky stretch? The 95% range is a confidence interval — given how noisy daily returns are, it is the range where the true alpha plausibly sits. A fund showing +2.7% alpha with a range of +0.5% to +4.9% probably added real value; the same +2.7% with a range of −4% to +10% is a shrug — the data cannot tell skill from luck yet, and the honest reading is "not proven." Whenever the range spans zero, treat the headline alpha as unconfirmed. The information ratio divides alpha by tracking error: how much unexplained return the fund earned per unit of drift from its recipe. It rewards steady, repeatable outperformance over erratic swings — as a rough guide, above +0.5 is strong, and a number near zero means the alpha is small next to the noise, whatever its sign.

How far back does the analysis look?

Two lenses, two answers. The recipe and the headline alpha are fitted on the trailing two years of daily total returns (about 504 trading days — the window dates are printed at the top of the card). Two years is a deliberate choice: funds change strategy and managers over time, so a longer fit increasingly describes a fund that no longer exists, while a shorter one cannot tell skill from luck. Two years of daily data is roughly 500 observations — statistically substantial while still describing the fund you would be buying today. The rolling persistence view then looks back up to five years for funds old enough: the current recipe is held fixed and alpha is re-measured across all of that history, which is what the strip chart and the "positive in N of M windows" count show. The lookback is limited by whichever runs out first — the fund's own trading history or that of an ingredient in its recipe (a fund whose recipe includes a young ETF like a spot-bitcoin fund cannot be checked against dates before that ingredient existed). At the other end, a fund needs about six months of history before an estimate is published at all, and stays subject to the young-fund restrictions until it has roughly a year.

What do the Reliable, Credible and Inconclusive labels mean?

The label is the one-word answer to "should I trust this alpha number?" — and it applies to negative alpha just as much as positive: a reliably *negative* alpha is an equally trustworthy verdict. Behind it are three independent checks. First, the 95% range must exclude zero — the alpha is too large to be noise. Second, the out-of-sample re-measurement must agree in direction with the headline. Third, the fund must have beaten (or lagged) its recipe in at least two-thirds of rolling windows — we hold the recipe fixed, re-measure the alpha over trailing 6- and 12-month windows stepped monthly (reaching back up to five years for funds old enough), and count. Reliable means all three checks pass and the fund has at least a year of history. Credible means two of three. Inconclusive means the data cannot yet separate this alpha from luck — which is the honest reading for most funds most of the time, and always the verdict when the recipe itself explains too little of the fund's movement. The rolling count appears on the card (e.g. "13/17 positive"): a fund that earned its alpha steadily looks very different there from one whose whole headline number rode on a single lucky stretch.

Each label a fund can carry on the rankings page, in one place:

A concrete case of a negative alpha wearing a trust label: CONY, the covered-call fund on Coinbase stock, behaved like roughly three-quarters COIN with the rest split between T-bills and a bitcoin fund — and trailed that simple copy by about 30% a year over the fitted window. Two of the three checks agreed the gap was genuine: the 95% range sat entirely below zero, and *zero* of twenty rolling 12-month windows showed the fund ahead of its copy (the one dissent: the out-of-sample re-measurement flipped positive). So the shortfall earned Credible. Read next to the red number, the label means "this underperformance is probably not a statistical fluke" — the covered-call overlay systematically gave up more upside than its option income recovered. It is a warning wearing a trust badge, not praise.

Why do some funds say they're measured against T-bills instead?

Ultra-short and money-market-style funds (SGOV, BIL, JPST, BOXX and kin) barely move day to day, and that breaks the usual yardstick: R² compares the copy's error to the fund's own variance, and when that variance is nearly zero, even a superb copy scores nonsense. So when a fund's volatility is under about 1% a year, the analysis skips the recipe hunt entirely. The recipe is pinned to 100% cash and the alpha shown is the fund's return spread over T-bills (SGOV) — which is exactly the number a buyer of these funds cares about: what does it pay beyond parking the money in cash? These funds sit out the alpha rankings, since a cash spread and a recipe alpha are different yardsticks.

Which funds does Dividend Vision analyse?

A deliberate subset, not the whole market, driven by two demand signals: the funds Dividend Vision users hold in their portfolios (privacy-filtered — a fund must appear in at least five accounts before it enters the analysed set, and only ticker symbols are ever read, never who holds them), and the funds users search for (a symbol typed into site search often enough over the trailing month joins the set automatically — interest counts before ownership does). Coverage refreshes nightly and follows what readers actually own and look up. If a fund has no Alpha Analysis card on its ticker page, it simply isn't in the analysed set yet — that means "not yet covered," never "zero alpha." A fund also needs at least six months of trading history before we will publish an estimate, and individual companies (stocks, REITs, BDCs) are excluded from the rankings because their "alpha" against a fund benchmark shelf is mostly single-company risk, not manager skill. We run a subset on purpose: it keeps every published number on a fund people actually hold and watch, where a methodology problem would be noticed quickly. Coverage expands over time.

Who invented this?

William F. Sharpe — the same economist behind the Sharpe ratio — published the method in 1992 as a way to determine a fund's effective asset mix from returns alone. It remains a standard tool in professional fund analysis.