AAQS Backtested: 16.95% a Year Since 1993 — and Just 0.77 Percentage Points Ahead Since 2010

The AAQS — the AlleAktien Quality Score — awards one point for each of ten criteria a company meets, ten in total. We recalculated what two mechanical portfolios built on that score would have returned across the entire US stock market from January 1993 through August 2026 — survivorship-free, including every company that has since vanished from the exchange. Every figure in this study is a simulated result of a mechanical rule — no such portfolio ever existed and no investor earned these returns.

Over the full stretch the result looks spectacular: Portfolio A, buying once a stock reaches 9 out of 10 points and selling the moment the score drops below that, returned 16.95% a year against 10.9% for the S&P 500. That is exactly the number we trust least ourselves. Through the 1990s only a fraction of the market carried a score at all, and from 1997 to 2000 coverage fell to between 24% and 29%. A sampling gate we built decided the calculation only becomes reliable from 2010 onward.

Inside that reliable window the lead shrinks dramatically. Portfolio A returned 15.04% a year against 14.27% for the S&P 500 — a lead of just 0.77 percentage points. Portfolio B, the stricter rule that buys only at a perfect 10, returned 17.08%, a lead of 2.81 points. That inverts the obvious expectation: it is not the looser threshold that carries the result, but the stricter one.

The universe is 22,890 US stocks, and 16,484 of them — 72.01% — are delisted today: bankruptcies, takeovers, quiet exits. All stay in the sample for as long as they actually traded, which is what makes the test survivorship-free; counting only stocks still listed today would flatter the past. Trading happens the next trading day after a signal at the closing price, equal-weighted across every position, on dividend- and split-adjusted prices. The buy filter requires a minimum price of $1 and a minimum market value of $100 million. The base case runs without trading costs; a separate sensitivity run charges 0.1% per side.

Portfolio A turned $100 into $19,019.40 over the full 33.52-year stretch, Portfolio B into $16,041.24. The S&P 500 reached $3,206.33, its equal-weighted counterpart $1,229.99. Portfolio A's worst drawdown, at −48.07%, is milder than the S&P 500's −55.19% and far milder than the equal-weighted index's −59.92%; Portfolio B's −46.18% is milder still. Both portfolios also carry more return per unit of volatility: 0.93 for A and 0.89 for B, against 0.65 for both indices.

Portfolio A held 249.2 positions on average against Portfolio B's 102.9, generated 4,342 orders against 2,003, and turned its capital over 49.0% a year against 55.2%. A win rate of 62.9% for Portfolio A and 63.2% for Portfolio B sounds solid until read the other way round: close to four trades in ten still lose money. And 17.1% of Portfolio A's purchases — 12.9% for Portfolio B — went into companies that later disappeared entirely.

One qualification, not a footnote: the equal-weighted S&P 500 series only begins on 30 April 2003, a 23.27-year span, so its 11.39% is not measured over the same period as the other three columns. One reassurance too, since readers of backtest studies are used to the opposite: both portfolios and both comparison series trade dividend-adjusted prices, so the comparison is fair in that respect.

Why we do not trust our own headline number

A rule that only a quarter of the market could even qualify for is not a rule tested at any real scale. Coverage sank to between 24% and 29% from 1997 to 2000, meaning roughly seven in ten stocks in those years carried no score and could never be bought, whatever their quality. We built a sampling gate to decide, independently of the outcome, from which year the calculation deserves to be called reliable — and it settled on 2010. The table below repeats the headline metrics, normalized to $100 on 4 January 2010.

Over the 16.59 years since 4 January 2010, Portfolio A turned $100 into $1,021.29 and Portfolio B into $1,367.86, against $914.28 for the S&P 500 and $717.63 for its equal-weighted counterpart — a lead of 0.77 points for Portfolio A and 2.81 for Portfolio B. It is the stricter rule, not the looser one, that wins in the stretch we trust. Drawdowns tightened too, to −38.70% and −37.17%, close to the S&P 500's −33.72% but well inside the equal-weighted index's −39.04%.

Worth stating plainly: 16.6 years is essentially one market cycle, and a friendly one — recovery from the financial crisis, a decade-long bull market interrupted mainly by 2022, no extended bear market of the kind the 1990s or 2000s produced. A single friendly cycle is not proof the pattern holds in a hostile one.

What the score checks

The AAQS (AlleAktien Quality Score) is based on the criteria catalogue developed and published by AlleAktien. We calculate it independently from our own business figures; there is no partnership.

Just as clearly, here is what we did NOT test: this is our own rebuild of that criteria catalogue, not the score AlleAktien publishes. Three of the ten criteria had to be replaced by substitutes for the historical run, and we never compared our reconstructed scores against the values AlleAktien actually published. AlleAktien took no part in this study and has neither reviewed nor endorsed it.

The full catalogue, one point for each criterion met:

1. Revenue growth over the last 10 years above 5% a year
2. Expected revenue growth for the next 3 years above 5% a year
3. EBIT growth over the last 10 years above 5% a year
4. Expected EBIT growth for the next 3 years above 5% a year
5. Net financial debt below four times EBIT
6. EBIT positive in every one of the last 10 years
7. Largest EBIT decline under 50%
8. Tangible return on equity above 15%
9. Return on capital employed (ROCE) above 15%
10. Expected return — free-cash-flow yield plus expected EBIT growth — above 10%

From 9 of 10 points, a stock counts as a quality stock under this catalogue (AlleAktien's own description of the score). We also run it live: our AAQS scanner recomputes the same criteria catalogue daily; the backtest results do not carry over to it. The table below breaks both portfolios and both benchmarks down year by year.

The dot-com bust is where the score's discipline shows most clearly: Portfolio A gained 23.59% in 2000 and 22.83% in 2001, essentially flat at 0.46% in 2002, while the S&P 500 lost 9.74%, 11.76% and 21.58% in those same three years. A score built on ten years of actual profitability kept earnings-free story stocks out of the portfolio in exactly the years that punished them hardest.

2008 tells a different story: Portfolio A fell 27.58% and Portfolio B 25.58%, sharply better than the S&P 500's −36.79% and the equal-weighted index's −40.06%, but still a serious drawdown — a quality filter reduces damage in a systemic crash, it does not prevent it. 2022 splits the two rules: Portfolio A lost 17.07%, a touch better than the S&P 500’s −18.18%, while the stricter Portfolio B lost 20.43% and finished behind the index.

2024 is the counter-example worth stating plainly: Portfolio A returned 11.29% against 24.89% for the S&P 500. A score built on profitability and moderate leverage does not buy the handful of mega-cap technology stocks that carried the index that year — the rule falls behind in exactly the concentrated rally it was never designed to chase. 1993 and 2026 are partial years in the table below, starting late January 1993 and ending early August 2026.

None of the year-by-year swings above mean much without knowing how much of the market actually carried a score at the time. In 1993, 1,119 of 1,555 tradable stocks — 72.0% — already had one. That share collapsed as the decade went on: to 29.0% in 1997 (1,703 of 5,863 tradable stocks), and to just 23.9% in 1999 (1,994 of 8,354), the weakest year in the entire study. It climbed back only slowly — 28.5% in 2000, 50.7% in 2005, 65.7% in 2010, 80.1% in 2015, 85.1% in 2020, 92.4% by 2026.

The collapse from 1997 has a specific cause: from that point on, the price series fill up with thousands of small companies that later vanished from the market entirely and for which usable financial statements were never available. In 1999, 6,654 of the 8,354 tradable stocks are not listed today. A stock with no usable filing gets no score, and no score means it is never bought, however promising it looked at the time — exactly why the earliest years carry a thinner slice of the market, and exactly why we built the separate "window from 2010" section rather than let the full-period headline stand alone.

The next question is what happens when the underlying assumptions change — trading costs, the size filter, the sell rule for stale filings, whether the portfolio keeps its weights balanced. The table below runs six variants against the base case.

Trading costs, at 0.1% per side, cost roughly 0.4 points a year — 16.95% falls to 16.56% for Portfolio A, 16.36% to 16.00% for Portfolio B. Dropping the size filter and letting the smallest, least liquid stocks back in actually improves the result slightly, to 17.32% and 16.49%: the rule is not carried by a handful of illiquid micro-caps. Turning off the sell trigger for stale filings makes essentially no difference — 16.95% and 16.37%. The most expensive assumption is not rebalancing: without continually restoring equal weights, the result drops to 15.84% and 15.52%, a loss of 1.1 points — part of this study's return comes directly from re-equalizing position sizes, not stock selection alone.

The counter-check: the same rule without forward-looking criteria

We also ran the rule with three criteria removed. Criteria 2, 4 and 10 depend on future expectations no longer available as history; the main run substitutes trailing three-year growth, the counter-check drops all three, scoring to a maximum of 7 with thresholds of 6 of 7 and 7 of 7. Full period: $13,723.81 for Portfolio A — 15.82% at −49.75% — and $12,119.24 for Portfolio B, 15.39% at −48.51%. From 2010: 14.12% and 15.72%. Main run and counter-check are not directly comparable, since they score a different number of criteria — but the counter-check moves the same direction, surviving the loss of every forward-looking criterion.

The traffic light: a rebuild, not a record

The important honesty comes first: the quality traffic light published alongside our analyses is an editorial judgment. It does not exist for the past and cannot be reconstructed retroactively. What we tested instead is a rule-based rebuild: buy only if a stock is not red on the day of purchase, "not red" combining green and amber. Four warning signs feed it, each computed from the figures published as of the purchase date — insofar as those were not restated later:

• Equity has been wiped out (negative).
• Operating income does not cover interest expense (interest coverage below 1).
• Two consecutive years of cash outflow from operations despite a reported profit.
• Cash on hand would not last a year at the current rate of outflow.

What a human reader sees and the rebuild cannot: an auditor's going-concern note, a dispute or change at the top of the company, dependence on one customer or lender, a looming delisting. And the rebuild only filters purchases — a holding already in the portfolio is not sold when it turns red. The table below tests it against 454 published judgments we could recalculate as of their publication date.

Of 105 truly red judgments, the rebuild caught 78 — a recall of 74.3%. But of the 175 red flags it raised in total, only 78 were actually red — a precision of just 44.6%. It finds most of the real warning signs, but more than half its red flags are false alarms. The largest driver is interest coverage, responsible for 89 of them, well ahead of equity (14) and cash runway (12); operating cash flow produced not a single false alarm on its own.

In practice the rebuild mostly turns ambers into false reds — a company with a manageable weakness gets treated like one genuinely in trouble. Whether the filter actually helps returns is a different question, which the table below answers directly.

The filter costs something in every one of the six runs, never more than 0.2 points: Portfolio A drops from 16.95% to 16.89% over the full period, 15.04% to 14.86% from 2010, and 15.82% to 15.74% in the counter-check; Portfolio B drops from 16.36% to 16.2%, 17.08% to 16.98%, and 15.39% to 15.21%. Exactly what the calibration above would predict: a filter that gets less than half its own warnings right mostly excludes healthy companies rather than protecting against failing ones — small in every run, but the same direction in all six.

Limits of the method

1. The score is reconstructed, not recorded. Nobody logged, twenty years ago, how many points a stock carried; this backtest computes it retroactively from figures already published as of each date.
2. Restatements distort the reconstruction. Where a later correction improved a company's figures, the retroactive score comes out more favorably than it would have looked in real time.
3. Expectations cannot be rewound. Criteria 2, 4 and 10 depend on forward-looking estimates that no longer exist historically; the main run substitutes trailing three-year growth, the counter-check drops all three.
4. The early years are thinly populated — exactly why the window-from-2010 section is mandatory, not a footnote.
5. Share count is only roughly traceable and feeds the size filter through market value; where none existed for early years, the oldest known figure was used.
6. The traffic light is a rebuild, not a lookup — see its calibration above against 454 real judgments.
7. Taxes and the bid-ask spread stay out of every run; only the dedicated cost run adds 0.1% per side.
8. Not every stock is scorable. 9,303 of the 22,890 stocks carry no usable financial statements at all, and a further 59 — 0.26% of the universe — remained unscored at the time of the run. No filing means no score, and no score means no purchase.
9. The equal-weighted comparison series only begins in 2003. Its 11.39% a year covers 23.27 years, not the full 33.52 years of the other columns.
10. The window we trust is a single market cycle. The 16.59 years from 2010 were mostly friendly; how the same rule fares in a long sideways or high-rate stretch is not measured here.
11. A backtest is not a promise. It shows what a rule would have returned without a single exception — not what it will return going forward.

What does not follow from this study

This is a historical study, not investment advice. Nothing in it recommends buying or selling any stock, sets a price target, or forecasts any company. A score of 9 or 10 today tells you that a mechanical rule built on that threshold produced a particular historical return across thousands of trades over three decades — it says nothing about what any single stock will do next. Past results, real or simulated, are no reliable indicator of future returns.

Two related studies run the same kind of test on different rules: our Growth Gems backtest retested our own growth-scoring scanner, and our Magic Formula backtest and our study of Benjamin Graham's net-net rule each retested a well-known value formula. The full collection lives under all studies.