Methodology

Welcome to the engine room. Here is how everything on this site is computed: which file each number comes from, the exact formula that turns it into a score, how the portfolio is built, and what the curve actually measures. Every claim on this page carries the source file in brackets, so you can check it rather than believe it. And behind every page there are twenty-nine automated checks — prices cross-checked against a second source, curves reproduced window by window across two independent engines, even the layout reviewed with a real browser. Twenty-one run on every cycle of the day; the other eight (layout, indexability, structured data, accessibility, stable addresses, profile navigation, the company-page figures and the price cross-check) run at deploy time, because they need a browser or a second source. And it is worth being exact about what happens when one fails: the portfolio check reverts on its own — if carteras.json comes out inflated, the last validated copy is restored before anyone sees it — while the rest warn but do not revert: they write the failure into the cycle log by name, and what was generated is published anyway.

The specific figures on this page are measured on the data of October 21, 2026 —the last closed portfolio is the 10/21/2026 session and the last ranking is the 10/21/2026 session— and each one carries its date. A figure without a date means nothing. And you do not have to take any of them on trust: the daily ranking and each portfolio's curve download as CSV, free, with no account and no signup, from the same file that paints the table (§10). That is something few services offer even for money; here it is the norm.

1. What gets published, and what does not

Let us start with the core of it: every return published here has a reproducible walk-forward backtest behind it, and this page tells you exactly what each number is and where it comes from.

And one distinction, the hardest to maintain: today's holdings are not the backtest. A model portfolio's table lists today's best companies according to the algorithm — the model's output today; the curve above it is the result of having repeated that rule every month since December 1998. Both come from the same code and still answer different questions.

2. The universe: who is eligible, and on which date

The universe is the S&P 500. The hard question is not that one, it is when: who was in the index on the day the model made its decision. Rebuilding that is laborious and easy to skip; here it is the foundation of the whole calculation, and this is how it is done. Up to Aug 21, 2026, membership comes from Sharadar's historical S&P 500 table, back to 1998 and including the companies that later disappeared. What follows in this section —the reconstructed composition, renames, recycled tickers, foreign tickers and the Cboe patch— applies to our usual data, which only feeds the months lived since that date: the Sharadar history uses a single symbol space and is computed without patches.

Two more filters, small and with a measurable effect. 20 tickers that were never in the S&P 500 are dropped from the price history (AAP, B, BL, CC, CR, DEC, E, EC, G, HR, MC, ORC, R, RC, SND, SY, TE, TY, UA, XE): 3.8% of foreign names, and not harmless ones, since ORC pays out ~20% a year and EC 8% per quarter — exactly the profile the Value factor grabs. And when both share classes of the same company are present, the secondary one is removed —GOOG with GOOGL, FOX with FOXA, NWS with NWSA— before the portfolio is picked (factores.py, CLASES_DUPLICADAS): without it, an "equally weighted" ten-stock portfolio would hold 20% of the money in a single business in 9 of the months measured.

There is one declared exception, in plain sight and by name: Cboe Global Markets disappears from the source table in February 2019 and in reality stays in the index to this day. That is a source failure, not an index event, and it is patched explicitly and can be switched off (universo_pit.py, PARCHES). Its effect is small: CBOE occupied 14 stock-months of the 50-stock multi-factor portfolio and none of the 20-stock one.

3. The data and its sources

Fundamentals: the SEC, and the filing date

Fundamentals come from what companies file with the U.S. Securities and Exchange Commission (10-K, 10-Q and, for insiders, Form 4). Every SEC data point carries two dates: end, when the period closed, and filed, when it was published. A fiscal year ending in December is not published until February or March. Here everything is indexed by filed (construir_fundamentales_pit.py): using a figure before it existed is information from the future, the classic error that inflates backtests across this industry. Point-in-time SEC data, straight from the primary source: something even most paid platforms do not offer, and it underpins every figure since Aug 21, 2026; before that, accounts come from Sharadar, which also dates them by their SEC filing.

TTM: the last twelve months, not the last fiscal year

Flow items —revenue, net income, EPS, operating cash flow, gross and operating margin— are trailing-twelve-month figures: for the history, Sharadar's as-filed TTM; in the daily calculation, rebuilt quarter by quarter and summing the last four. Stock items —equity, assets, liabilities, debt, shares, cash— are the last quarter's snapshot.

The classic problem with SEC quarterlies is that they come mixed: a 10-Q publishes both the standalone quarter and the year-to-date total, and the 10-K publishes the full year but not the fourth quarter separately. It is solved by subtraction: Q4 = full year − nine-month cumulative. For Apple FY2024, 391,035 − 296,105 = 94,930 million, which is exactly the fiscal fourth quarter Apple reported. The subtraction is always between two periods starting on the same day and under the same accounting tag; if a piece is missing and the chain cannot be built, it falls back to the plain annual figure.

The XBRL tags

The same accounting concept is tagged differently by company and by era. All of them are tried in preference order and the first one with data for that specific quarter wins (a whole tag is never dropped at once). These are the canonical ones (construir_fundamentales_pit.py, CONCEPTOS and BALANCE):

Those choices are not cosmetic. Cost of revenue was added because Netflix abandoned GrossProfit in 2021 and McDonald's, Chevron, Dow, Nucor and Corteva never use it: with no gross margin they were left with two of Quality's five components, below the minimum of three, and dropped out of the factor entirely. Debt tags were added because the first three left banks, insurers, oil companies and REITs with no leverage figure. And SecuredDebt, UnsecuredDebt and SubordinatedDebt are not added even though many companies publish them: they are slices of debt, and using a slice as the total says the company owes less than it does, which is worse than not scoring it.

Prices: what is discarded, and why

Price history up to Aug 21, 2026 comes from Sharadar (split-adjusted, with spin-offs rescaled and every company that left the index). Day to day, prices come from professional financial APIs: live quotes through a keyed, rate-limited API, and a second provider that acts as a cross-check on the first. The close used is split-adjusted but not dividend-adjusted (daily_update.js): dividends are added separately on their ex-date, and using an already dividend-adjusted price would count them twice. The symbol list lives in exactly one place, simbolos_yahoo.json: two copies of the same list always end up drifting apart.

Merger exchanges that the provider encodes as splits. AvalonBay stopped trading on Aug 14, 2026 and each share was exchanged for 2.793 Vivmark shares. The provider declares this as a 2.793:1 split dated 17/8 and serves AVB's entire series divided by 2.793 (65.90 where it closed at 184.06). It is not a split, it is an exchange, and treating it as a split would rescale 1,543 of 1,556 rows and leave a −64% cliff in June 2020 that never happened. AVB's series stays in its own scale and ends on the 14th; mapping it to VMRK would splice two companies together. The distinction is declared, with its evidence, in simbolos_yahoo.json: only pure, verified renames go in renombres —BNY gives exactly the same closes as BK across the 1,044 common dates, and ECHO the same as SATS across 1,045— while mergers go in delistados.

Real splits are applied, and applied backwards, across the full history: the download is compared with the stored data over the overlapping days and, if the median ratio drifts from 1 or the provider declares a split in the window, that ticker's entire series is re-downloaded. Rescaling only a window of days would leave the series split across two scales and manufacture crashes that never happened — KLAC from 1,940.04 to 192.92 (−90.1%), CVNA −80.3%, CRWD −75.4%, DD +209.1% from a reverse split — which is why the scale is always checked against the full series.

The market benchmark

The yardstick for the equal-weight version is the equal-weight S&P 500 computed from the same data as the portfolios: every company that was in the index each month, weighted the same, dividends reinvested, including the ones that later disappeared. It is the fair benchmark for a portfolio that also splits its money evenly: measuring it against a cap-weighted index would mix two different things —which companies were picked, and how the money was split— and hide which of the two is doing the work. The cap-weighted version is measured against SPY, the S&P 500 ETF, which has traded since 1993 and covers the whole window.

Until September 2026 the equal-weight yardstick was RSP, the fund that tracks that index, rather than an average of our own, for a reason that was true: an average of one's own universe only averages companies with available prices, so it inherits the very survivorship bias it is meant to judge (depending on which universe was averaged, it came out at +64.7% or +83.9% cumulative over a window in which RSP did +46.3%). With the long backtest that reason goes away —we now have prices for every index member at every date, including the ones that disappeared— and a new one appears: RSP did not exist before May 2003, and the backtest starts in 1998. So the yardstick is the calculation, and RSP becomes its check: over the 23 years in which both exist, the calculation is ahead by +0.55 points a year, which is the fund's fee plus its trading costs, and month by month they move almost identically (correlation 0.999, measured on Sep 11, 2026). If they ever drift apart by more than a point and a half a year, the site stops publishing the track record rather than show a broken yardstick.

An asymmetry worth declaring: neither the calculated yardstick nor our portfolios deduct costs, while RSP and SPY, being funds, carry their fee inside the price (0.20% RSP and 0.09% SPY). In the equal-weight version the comparison is like for like; in the cap-weighted version, SPY's fee works slightly in our favor.

And if the benchmark does not reach as far as the portfolio, the calendar is cut short and we say so — the gap is not filled. Dragging the benchmark's last close forward when it lags the stock prices would hand the market a rise it never made — four sessions of lag are worth an invented +0.87%. One fewer point on the curve is a visible gap; one point with two dates inside it is a false figure (regenerar_carteras_honesto.py).

4. The four factors, one by one

Here is the complete recipe, thresholds and limits included. There is one single definition of each factor and both halves of the system use it: the ranking you see and the backtest that measures it (factores.py). This is not housekeeping: with two definitions of the same factor —one for the backtest, one for the site's engine— you would be showing the returns of one next to the ranking of the other, which is not the same thing. Here it is one definition, in one file, shared by both halves.

Momentum

Twelve-month return skipping the most recent month, divided by annualized volatility. It comes from prices, not from accounts.

momentum = ( P[t−21] / P[t−252] − 1 ) / ( σ · √252 )
σ = population standard deviation of daily returns over the last 252 sessions

It requires 253 closes, more than 30 valid daily returns and strictly positive volatility; otherwise there is no momentum and the reason is stated (historia_corta, hueco_en_extremos, pocas_sesiones, volatilidad_cero). The 21-session skip is the standard in the literature: the most recent month is excluded because at one month the effect reverses.

Value — three yields on price, minimum 2 of 3

ey = EPS (TTM) / price earnings yield
by = (equity / shares) / price book yield
sy = (revenue TTM / shares) / price sales yield

Quality — five ratios, minimum 3 of 5

roe = net income / equity
roa = net income / assets
gm = gross profit / revenue gross margin
om = operating income / revenue operating margin
lev = debt / equity ENTERS INVERTED: more debt, worse

All five are requested as same-filing pairs (§3). And explicitly declared zero long-term debt is not a gap: it is the best possible leverage. Five index companies declare it today —CMG, FSLR, ODFL, SNDK and WDC— and that declared zero scores as what it is: the best possible leverage mark, never confused with a missing figure.

Growth — three year-on-year growth rates, minimum 2 of 3

g_rev = revenue(today) / revenue(one year ago) − 1
g_eps = EPS(today) / EPS(one year ago) − 1
g_ocf = operating cash flow(today) / same one year ago − 1

"One year ago" is another point-in-time snapshot taken under the same rule: only what was public that day (pit.py, foto_anterior()). The base must be positive: growing a percentage over a loss means nothing — from −16.50 to +2.81 is not "117% more earnings", it is something this ratio cannot measure. That is different from having no accounts for last year, and two different reasons are emitted (base_no_positiva and sin_ano_anterior). Boeing publishes all three of its prior-year figures; the point is that two of them are losses.

Before normalizing: what is thrown away and what is clipped

There are two ranges per component, and the difference matters a lot. POSIBLE is what a real company can actually have: outside it the data point is wrong —not an extraordinary company— and the component is discarded. LIMITES is the statistical clip inside what is possible: real but extreme values that, left unclipped, blow up the standard deviation and compress everyone else's Z-score.

Clipping broken data is worse than dropping it. Equity Residential showed a gross margin of 5,070 because its revenue tag for that period was 216,000 dollars instead of 2,900 million; clipped to 1.0 it did not end up "discarded", it ended up as the best in the index. That is how SBAC finished #4 of 476 in Quality, and how BRK-B was #1 in Value with the maximum mark, on a class-A EPS applied to the class-B share (a 4,253% earnings yield).

The floors are asymmetric for a concrete reason. Revenue cannot fall more than 100%, so g_rev is floored at −1. EPS and cash flow can: they only need to cross into negative territory. A −1.16 there is not broken data, it is a company that went from making money to losing it — exactly what the factor should penalize; with the floor at −1 it was discarded and the company was left with no growth score.

Two sanity checks on the per-share metrics

Two, not one, because they do not depend on the same thing. Book value and sales per share only need the share count to be credible (at least one million; the index company with fewest shares is NVR, at about 2.9 million). Earnings yield additionally needs the EPS to reconcile: EPS × shares / net income must fall between 0.33 and 3.0.

The share minimum exists because several companies publish the count in millions under one tag and in units under another: McDonald's gave both 711.1 and 710,505,859, and at 711 shares McDonald's "capitalized" 195,000 dollars. The EPS band is deliberately wide, so as not to confuse a scale error with the legitimate effect of preferred stock: Albemarle pays around 190 million a year in convertible preferred dividends, its identity gives 0.16, and the data is correct. Each of the three metrics carries its own flag: an odd EPS does not take the other two down with it.

When a cell is blank, the reason is emitted by whoever blanked it

Every component that fails to score writes down its code and the figure behind it on the same line that discards it (factores.py, function _anotar; published in motivos_factores.json). An emitted reason cannot contradict the calculation because it comes out of the calculation.

The easy alternative —letting the browser guess the reason from which fields are missing— was ruled out by measuring it: a reason deduced from symptoms would tell Boeing "no accounts from a year ago" while all three prior-year figures are published, and NVR "banks and insurers, or they owe more than they own" — a homebuilder with 3,391 million of positive equity. Guessing is right one time in three; that is why the reason is issued by the calculation, not by the symptom.

The codes are: falta_en_la_sec, acciones_no_creibles, bpa_no_cuadra, patrimonio_no_positivo, ingresos_no_positivos, base_no_positiva, sin_ano_anterior, fuera_de_rango, sin_cuentas, periodos_distintos; and at factor level, pocos_componentes, sin_normalizar and clase_duplicada. In the Aug 21, 2026 session there were 41 blank cells across 499 companies: 39 for not reaching the minimum number of components and 2 for too short a price history.

5. How they combine: the Z-score

Each component is normalized separately and only then are the Z-scores averaged. In that order, and the order is the whole point.

Z(component) = clip±3 [ sign · (x − mean) / stdev ]
Z(factor) = clip±3 [ mean of the Z of its available components ]
Z(multi) = ( Z_momentum + Z_value + Z_quality + Z_growth ) / 4

Mean and standard deviation are computed over that date's scored universe, not over a historical table: the deviation is the population one, at least 20 companies must carry the component, and the deviation must be strictly positive. The sign is −1 for leverage (more debt, worse) and +1 for everything else. Winsorization at ±3 is applied twice, at component level and at factor level.

The multi-factor score only scores companies that have all four factors. Scoring one on two and comparing it with one scored on four is not a comparison. That is why the multi-factor ranking is shorter than the single-factor ones: in the 10/21/2026 session there were 494 companies with momentum, 491 with value, 491 with growth, 471 with quality and 464 with all four, out of a 503-name constituent list.

6. The walk-forward

Our backtest is built to answer one question only: what would genuinely have happened. The easiest way to make a strategy look brilliant is to pick the winners using today's data and then measure them against past prices —it produces spectacular figures and zero predictive value— so here it is done exactly the other way round, and reproducibly: the history up to Aug 21, 2026 was computed once and frozen, and each new month is added by the daily cycle (regenerar_carteras_honesto.py, and its laboratory twin backtest_honesto.py):

Returns are total returns: each leg's return is (end price − start price + dividends) / start price, counting dividends whose ex-date falls inside the leg. The same treatment is applied to the strategy and to the market.

And a position that stops trading is closed, not evaporated. If the quote is cut mid-leg by a takeover, a merger or a delisting, returning "no data" would erase that company from the average as if it had never been bought, and survivorship bias would walk back in through the side door. It is liquidated at the last known price and that result counts.

The metrics, and how they are measured so they cannot flatter the portfolio

Where each curve starts, and why not at zero

Since December 1998 the point-in-time universe already fills all twenty portfolios, so today they all start on the same day. The rule is still there for when that is not the case: a period with no portfolio is a declared gap, never an invented zero: one flat month that never happened would be enough to make two screens report two different CAGRs for the same cumulative return. The series starts at the first period where the portfolio genuinely fills, and a gap after the start is not filled in: the program stops and says so.

All five strategies of a given size start on the date the universe fills all five, which is the multi-factor date (it needs all four components at once and is the most demanding). With the 10/21/2026 file: the 10- and 20-stock portfolios start on 12/31/1998 and the 50- and 100-stock ones on 12/31/1998. Momentum, value and quality could have started earlier; starting later is not inventing a data point, but hiding it would be.

7. Costs

Very few sites tell you what following their strategy costs; here it is measured and published. And so you can read it with the cards face up, the important part first: no curve on this site includes costs. No commissions, no bid-ask spread, no slippage, no taxes. Your real result will be lower. It is written in the notices attached to every portfolio in carteras.json, and it is repeated here because it is the limitation that weighs most on a strategy that turns over every month.

What is measured, separately, is turnover and what it would cost (medir_senal.py). The assumption is written down: 10 basis points per side (0.10% of the traded amount, on each side), a round and deliberately conservative figure for S&P 500 names — it covers a retail broker's commission, half the spread of a large liquid company, and some slippage. It can be changed with --coste. Annual cost in points is annual_turnover × 2 × 0.0010, because each replacement is a sale and a purchase.

The hysteresis band is the cheap way to cut turnover without changing the signal: a position is held while it stays inside the top 2N and is only sold when it drops below; the gaps are filled with the best names not already held. Same entry rule, looser exit rule. It is not applied to the portfolios this site publishes: it is a measurement, not the strategy.

The half a point a year above is the cost of turnover under an assumption worth saying out loud: 10 basis points per side, over 59 rebalances from 2021-2026 and without a point-in-time universe. The same portfolio, measured with medir_costes.py over the 330 rebalances of the long history, costs 31 bp a year assuming a quoted spread of 3 bp, and 219-274 bp with the spread our two estimators infer from daily highs and lows, which is an upper bound. The three figures are the same calculation under different assumptions, not three results that contradict each other. That is why every cost figure on this site says three things: whether the spread is assumed or estimated, at what commission and what account size, and over which window. Without those three, two numbers from the same calculation read as a contradiction. It does not include currency conversion —which for a euro-based investor can be the largest expense and is the easiest to avoid— nor taxes: a portfolio rebalanced monthly realizes capital gains monthly, while an index fund defers them.

8. What the tests say

Nearly three decades of reproducible walk-forward, file by file: few houses can show you a sample like this, and this one is ours. Every figure in this section comes from carteras.json exactly as published today, over the window 12/31/1998 → 10/21/2026: 27.81 years and 333 measurable months, the same for every portfolio size. (One more rebalance than months: the last one is already bought and its month has not closed.)

And a sample of how nothing gets dressed up here: read the Value and Growth rows carefully: they are the ones with negative alpha, and Value also has the highest beta and the deepest drawdown in the table (carteras.json, Sep 11, 2026). A strategy can finish up and still have beaten nothing, and it is published all the same.

The edge: measured, not proven

Putting your own edge through a statistical test and publishing the result even when it does not flatter you is exceedingly rare in this industry. We do it on every server start, on the published curve:

The signal referee: does the score actually rank?

A portfolio backtest answers "how much would a 20-stock portfolio have made". That is one path: it depends on the cut at 20, on the luck of a handful of names and on the luck of the dates. To choose between two similar definitions of the same factor you have to measure the signal, not the portfolio (medir_senal.py). Four statistics, each with its exact definition written inside the file itself:

That table is worth reading in full, not just its last row. These are small ICs —a hundredth on average— with a between-month deviation of 0.12 to 0.21: the signals rank weakly and noisily. Value is the only factor whose decile ladder runs cleanly from top to bottom (monotonicity +0.90; at twelve months its best decile beats its worst by 3.4 points). Momentum has a positive IC but a flat ladder, quality has it upside down with an IC close to zero, and growth comes out inverted: at twelve months its worst-scored stocks return 2.8 points more than its best-scored. The multi-factor averages all of that: a positive IC and an almost flat ladder. On the short history quality came out inverted and momentum was the cleanest; with almost twice the history the picture changes, which is what you would expect from signals this weak. Everything is published because everything is measured; we do not pick the flattering part.

9. What we do not know

A section like this is one of the rarest things you will find on a financial site, and it is the one we are proudest of: telling you precisely where our knowledge ends is the best proof that everything else is genuinely measured. Nothing below subtracts from the above; it draws its boundary.

There is also a lag that is not a bug and is worth understanding: fundamentals are quarterly and a quarter takes about six weeks to reach the SEC. Until then the model carries on with the previous quarter. That is the same delay any real investor faces, not a bias.

10. How to check it yourself

We are not asking for an act of faith: we hand you the data. Publishing a methodology is worth little if the numbers cannot be downloaded, so everything this site paints downloads free and without an account, from the same file that paints the table: it cannot say anything different from the screen.

Every CSV carries its date in the header and in the filename, so that a month from now you know when it is from.

The twenty-nine guardians

A check that runs only once is not a guarantee: it is a snapshot of one good day. Twenty-nine verifiers run over the data before it is published, from a single command (verificar_todo.py), and they distinguish three states rather than two: fine, failed and not examined — because "I could not check it" is not "it is fine", and a verifier that conflates the two passes things falsely.

One on that list, medir_cobertura_pit.py, watches a risk none of the others do: the backtest silently falling back to a fundamentals file holding only today's companies —507 instead of 880—. That would leave 205 of the 498 index members of February 2011 with no fundamentals, pick the 2011 top 20 from survivors, and inflate the 20-stock multi-factor portfolio by +2.06 CAGR points — measured in August 2026 on the previous history, which started in 2011. And the backtest verifier would not catch it: its poison test looks for look-ahead leakage, and this is not leakage, it is selection bias. That is why coverage of every date has a guardian of its own.

Every daily calculation is redone three times each business day (00:00, 15:31 and 22:20 Madrid time) and the data files are rewritten in full on each cycle; the history up to Aug 21, 2026 is frozen and not recomputed. The live figures —the ones that change from one cycle to the next— are written by the server as it serves the page, straight from the published file. The one-off measurements are typed by hand, which is why each carries the date it was taken: nobody regenerates them, and a measurement without a date would be worth nothing.

The model portfolio, weekly in your inbox

Every Saturday, its week in one minute: how it went, the best and the worst, and which insiders bought with their own money. Plus, the day the model is rebuilt, which stocks come in and which go out. A simulation, not advice. At most one email a week, no ads, no account; one-click unsubscribe. Made by one person: Pol Gracia.

Send it to me

A model portfolio with simulated results, not a personalized recommendation. Your email is used for this only: privacy.