Growth Gems Backtested: 13.17% a Year Since 2000 — and Why the Strictest Rule Loses

Our Growth Gems stock scanner scores every US company from zero to ten points, one point for each of ten criteria met, covering how fast a company grows, whether that growth is honestly earned, and how well it is backed by the balance sheet. We wanted to know whether that score rule actually holds up as an investing rule rather than just a screening tool, so we turned it into three separate portfolios and ran them, month by month, across the entire US stock market since the year 2000.

The headline result: the portfolio that buys once a company reaches 9 out of 10 points and sells as soon as the score drops below that mark returned 13.17% a year from January 2000 through July 2026, turning a starting stake of $100,000 into $2,652,359. Over the same span, the S&P 500 returned 6.55% a year. What that comparison does and does not prove is what the rest of this study works through.

The universe behind the test is 22,871 US stocks, and 16,472 of them — 72.02% — are delisted today: bankruptcies, takeovers, and quiet exits from the exchange. All of them stay in the sample for as long as they actually traded and reported numbers, which makes the test survivorship-free. Counting only the stocks that happen to still be listed today would flatter the past. The test runs from January 2000 through July 2026, 319 month-end dates, roughly 26.5 years, starting with $100,000 and charging trading costs of 0.1% of order volume on every buy and every sell. Trades execute at the closing price of the month in which the signal appears, and all prices are dividend-adjusted.

Every number in this study is read directly from fundamental data and SEC filings — annual reports (10-K) and quarterly reports (10-Q) — combined with the historical price series of our own research base.

What the scanner checks: ten criteria, one score

The ten checks behind the Growth Gems score fall into three groups, and every one of them is worth exactly one point.

Speed, and the quality of that speed. A company earns a point if revenue grew by more than 15% a year over the last three years, if more than 10% revenue growth is expected for the coming year, if revenue growth plus free-cash-flow margin add up to at least 40 — a shorthand known as the "Rule of 40" — and if the gross margin sits at 40% or higher without meaningful erosion. These four checks ask a single question from four angles: is the company actually growing fast, and is that growth profitable enough to be worth having?

Is the growth honestly earned? Fast growth can be bought in ways that hurt shareholders later, so three checks look for the usual shortcuts. A point requires that the share count grows by less than 3% a year, so growth is not funded by printing ever more stock; that goodwill from acquisitions does not grow faster than revenue, so the company grew rather than shopped its way to the top line; and that operating cash flow covers the profits of the last three years, so the reported profit is backed by actual cash rather than by accounting choices.

Backing. The last three checks ask what happens if growth slows. A point goes to companies with net debt below twice EBITDA, so a rough patch does not turn into a debt problem; to a return on capital employed of 15% or more, or one that has risen compared with two years ago; and to companies where insiders hold at least 10% of the shares or have been net buyers — a sign that the people who know the business best have their own money in it.

One rule applies to all ten: if the figures needed for a criterion are missing, it counts as neither passed nor failed, and a score is only assigned once at least seven of the ten criteria can actually be judged. A company with too many gaps in its filings simply gets no score at all — and without a score, it is never bought.

The three portfolios

We turned the score into three separate portfolios that differ in nothing but where the buy-and-sell line is drawn.

Portfolio A buys from 9 out of 10 points and sells as soon as the score drops below 9. Portfolio B is the strictest: it buys only when a company reaches the full 10 out of 10 and sells the moment even a single point is lost. Portfolio C buys from 8 out of 10 points — the same threshold our live Growth Gems scanner uses today — and sells as soon as the score drops below 8.

Every other rule is identical across the three: same universe, same starting capital, same order costs, same trading at the month-end close. The only variable is how strict the entry has to be, which lets us ask a clean question — does a stricter score threshold actually buy you a better result? The headline table below gives the first answer.

Portfolio A returned 13.17% a year, turning $100,000 into $2,652,358.74 over the 26.5 years, at a worst drawdown of −52.13%. Portfolio C, one point looser on the buy side, is close behind at 12.58% a year and $2,311,684.10, with a milder drawdown of −46.18%. Portfolio B, the strictest of the three, trails badly at 6.79% a year and $570,464.94 — though its drawdown of −41.10% is also the smallest of the three, since a portfolio that is rarely invested cannot fall as far as one that always is.

Both benchmarks trail Portfolio A and Portfolio C on return: the S&P 500 managed 6.55% a year, the Nasdaq 100 8.12% — though the Nasdaq 100 also carried by far the roughest ride, an −81.07% drawdown during the dot-com bust that dwarfs anything the three portfolios went through. One caveat belongs here, not in a footnote at the end: both benchmarks are price indices without dividends, while the portfolios trade dividend-adjusted prices. That tilts the comparison slightly in the portfolios' favour, and it is worth saying plainly rather than leaving it implied.

The trade statistics round out the picture. Portfolio A turns its capital over roughly once a year (turnover 1.03) with a median holding period of 12 months — the same holding period as Portfolio C, while Portfolio B's rare positions are held a median of 7 months. A win rate of 58.92% for Portfolio A sounds solid until you read it the other way round: four trades in ten still end in the red. That is not a flaw in the presentation, it is what a 13% annual return actually looks like up close.

Portfolio B: why perfection ends up empty

The strictest rule delivers the worst result, and it is worth sitting with that sentence for a moment: more points required, worse outcome. The reason is not that Portfolio B picked bad stocks — it is that it barely picked any. Across the full 319 months it held 1.32 positions on average and sat entirely in cash in 100 of those months, 31.35% of the whole test. In seven calendar years — 2000 through 2005, then again in 2025 — it held no position at all. Its busiest year was 2023, with an average of just 4.8 positions. Across more than 26 years that adds up to 117 orders and 58 completed round trips: not a portfolio in any meaningful sense, but a handful of individual bets.

Perfection is not a portfolio. Demanding a flawless score on all ten criteria filters out almost everything, including plenty of companies that would have scored 9 out of 10 and delivered fine returns inside Portfolio A. With so few trades, Portfolio B's 6.79% a year rests on a small number of individual stocks rather than on a rule tested at any real scale — a cautionary tale about over-tightening a filter, not a verdict on the score itself.

Year by year

A 26.5-year average return flattens a much bumpier ride underneath it. The table below breaks all three portfolios and both indices down year by year, from January 2000 through July 2026 — note that the final row, 2026, covers only January through July and is not a full year's return.

The early years make the case for a survivorship-free test. In 2000, while the dot-com bubble was already bursting, Portfolio A gained 28.9% against a −5.32% S&P 500 and a −34.41% Nasdaq 100. In 2002, the worst year of the bust, Portfolio A lost just 2.4% while the S&P 500 fell 23.37% and the Nasdaq 100 37.58%. 2001 sits in between and is more honest about the limits of a growth score during a growth bust: Portfolio A lost 24.74%, worse than the S&P 500's −13.04%, though still well ahead of the Nasdaq 100's −32.65% — a growth-tilted portfolio does not escape a growth crash unscathed, it just tends to come out of one less battered.

2008 and 2022 are the two years in which every strategy in this study shows some form of drawdown. In 2008, Portfolio A fell 41.38% and Portfolio C 35.4%, close to the S&P 500's −38.49% and a touch better than the Nasdaq 100's −41.89%. 2022 tells a more interesting story: Portfolio A lost only 12.79% against a −19.44% S&P 500 and a −32.97% Nasdaq 100 — one of the clearest years in which the score's demand for balance-sheet backing paid off during a rate-driven sell-off in unprofitable growth stocks.

In several of the more recent years, though, the Nasdaq 100 outran both portfolios by a wide margin — 2023 (53.81% against 20.0% for Portfolio A and 19.54% for Portfolio C), 2024 (24.88% against 17.47% and 17.7%) and 2025 (20.17% against just 5.95% for Portfolio A). A handful of mega-cap technology stocks can carry an index further than a broader, rules-based growth portfolio in years like these. Portfolio B, for its part, shows a flat 0.0% in every year from 2000 through 2005 — not a data gap, but the direct consequence of holding no position at all, itself a result of the insider criterion only becoming measurable from 2006 onward (more on that in Limits of the method).

How many stocks could be judged at all

None of the returns above mean anything without knowing how many stocks were actually eligible to produce them. The table below is the honest counterpart to the 22,871-stock universe: a stock without usable filings gets no score, and without a score it is never bought, no matter how promising its story looks.

In the earliest years, only a fraction of the stocks with a price series carry a score at all: in 2000, just 1,810 of 8,138 stocks with a price — 22% — had a score, because SEC filings were far from universally machine-readable that early and coverage built up only gradually. By the mid-2020s that picture has reversed: coverage climbs past three-quarters and reaches 81.8% in 2025 and 80.3% in 2026. The practical consequence is that the earliest years of this backtest rest on a considerably narrower slice of the market than the later ones, and any comparison across decades should keep that in mind.

One number needs an explanation rather than raising a false alarm: the count of stocks "with filings" drops sharply in 2026, from 6,154 the year before to just 507. That is a reporting-date effect, not a loss of data — the current financial year has simply not been filed everywhere yet by the July 2026 cut-off used in this study. Where a fresh annual report was not yet available, the score for 2026 rests on the most recently published filing rather than on a missing one.

Does the score sort at all?

The three portfolios above each answer a question about a single threshold. They do not answer a more basic one: does the score actually carry information at all, or would any of the ten thresholds have worked more or less the same? To find out, we built ten entirely separate portfolios, one per score level. Each one buys a stock the moment it reaches exactly that score — no more, no less — and sells as soon as the score leaves that level again, whether it moves up or down. Level 10 is simply the ten-point portfolio under another name: "exactly 10" and "at least 10" describe the same rule, since nothing can score higher.

All ten ladder portfolios run over the same 26.5-year reference period with the same $100,000 starting capital as the rest of this study. The table below lines them up side by side.

Read from left to right, the ladder answers the question cleanly: return and win rate rise with the score almost without interruption. A single point buys 8.6% a year with a 47.3% win rate; by seven points that has climbed to 14.4% a year with a win rate of 56.4%. That rising line, not any individual level, is the real finding of this chapter — the score sorts. A company scoring 7 is a demonstrably better bet, on average and over 26 years, than one scoring 3.

Exactly 7 points turns out to be the strongest single level in the entire study at 14.4% a year — stronger than any of the three threshold-based portfolios from the first chapter, including Portfolio A's 13.17%. Exactly 10 points, by contrast, falls out of line at just 6.8%, echoing Portfolio B's result for the same reason: with only 59 buys across 26.5 years, usually a single position and often none at all, there simply are not enough candidates to draw a conclusion from. That is not a mark against the score — it is a sample-size problem, and the ladder makes that easy to see rather than easy to miss.

Two caveats belong directly next to this table, not after it. First, breadth: the middle levels, 3 through 5, hold 700 to 900 stocks at a time. A portfolio that size is not really a portfolio any more — it is closer to an index replica, and its returns say more about where the broad market went than about any selection skill in the score. Second, dividends, worth repeating here because this chapter puts percentages directly against the S&P 500 and Nasdaq 100: the ladder portfolios trade dividend-adjusted prices, the benchmarks do not, so every comparison in this section tilts a little in the portfolios' favour.

What does the exit change?

Every portfolio so far has answered a question about the buy side. This chapter fixes the buy rule and asks what the sell side is worth on its own. All 25 rows tested here buy exactly like the strongest level of the ladder — at exactly 7 points — so that any difference between rows can only come from how and when the position is closed. The reference row is E7: sell as soon as the score leaves exactly 7, the same portfolio from the previous chapter, returning 14.4% a year at a −53.1% drawdown.

The 25 exit rules fall into three groups: stops without a re-entry ban, score-based variants and profit-taking rules, and stops combined with a six-month re-entry ban. The table below lists all three, with the E7 reference set in bold at the top for comparison.

The single hardest finding in this chapter is also the most counterintuitive: a stop-loss without a re-entry ban measures nothing but round-trip costs. The evidence is exact rather than approximate — across all ten stop widths tested, from −5% to −50%, the number of sells triggered by the score alone stays fixed at precisely 7,461. What that number is hiding: a stock that gets stopped out is, in the overwhelming majority of cases, still sitting on exactly seven points. The buy rule fires again in the same month, and the portfolio simply buys it straight back. The stop does not protect the portfolio from anything — it produces an extra round trip and an extra order cost, visible directly in the numbers: the tightest stop, S05, drags the win rate down to 34.5% and pushes the buy count up from E7's 7,942 to 18,118, purely from stopping out and immediately re-buying the same names.

I5, a rule that imposes a minimum holding period of 12 months before the score is even allowed to sell, is by a clear margin the worst row in the table: 13.3% a year at a −59.3% drawdown, both worse than the E7 reference. Delaying an exit that the score has already called for does not protect the position — it keeps money in names the rule itself has flagged as weaker, and the portfolio pays for that delay twice, once in return and once in drawdown.

The three profit-taking rules — I6 (take profit at a double), I7 (take profit at +50%) and I8 (take profit at a triple), all of which keep holding through score upgrades — land within a tenth of a point of each other, at roughly 14.7% to 14.8% a year. That the exact profit target barely matters is itself the encouraging part: the improvement over the E7 reference comes from the underlying principle — hold the upgrades, take the gains when they arrive — rather than from having guessed the one correct multiple.

The clearest result in the whole matrix, though, needs the re-entry ban to show up at all. Add a six-month ban on buying a stock back after a stop, and stops beat the reference for the first time in this study: SB15, a fixed stop at −15% with the ban, reaches 15.6% a year; TB20, a trailing stop at −20% with the same ban, reaches 15.3% at a drawdown of just −47.4% — the mildest of any row in this table, against E7's 14.4% at −53.1%. The ban is what turns an exit into an actual exit instead of a brief pause before the same purchase.

One caveat belongs directly after these numbers: with 25 different sell rules measured against the same reference, some row was always going to come out on top even if none of them made a real difference. What survives that scrutiny is the direction — that a re-entry ban changes the outcome and a stop without one does not — rather than the second decimal place of any single row.

The recipe tested here — buy at exactly 7 of 10 points, sell with a 15% stop-loss or as soon as the score leaves 7, a six-month re-entry lock after a stop-loss sale — now runs as its own, daily-updated scanner: Growth Gems: 7-Point Recipe. The results list only reflects the buy side of the recipe, showing which stocks currently score exactly 7 points. The stop-loss and re-entry lock are rules for your own portfolio that the scanner does not apply automatically.

Data quality: two companies under one ticker

One data problem gets a section of its own rather than a footnote, because left unaddressed it could quietly inflate every result in this study. Under some delisted ticker symbols, two entirely different companies sit inside a single, unbroken price series. The mechanism is simple: an exchange reissues a ticker after the original company delists, and the new company's price history gets glued directly onto the old one's with nothing marking the seam. Looked at in hindsight, a stock like that can appear to have gained a hundredfold in a single month — a move that never actually happened to any real shareholder.

We built a quality gate that blocks any such series from being bought at all, in every portfolio in this study. Of 22,629 stocks checked that have a usable price series, 2,587 — 11.43% — are blocked. Narrowed down to the 2,421 stocks that ever reached Portfolio C's buy threshold of 8 points, 228, or 9.42%, are affected. The gate catches three distinct patterns: 1,185 cases of a hard price jump, a factor of 8 or more within a single month; 1,231 cases of a permanently lifted price level, where after a jump of factor 3 or more the median price of the following six months sits more than 2.5 times above the median of the six months before; and 171 cases where the series itself starts implausibly.

The most important sentence in this section is about direction: the gate blocks upward jumps only. A comparable drop downward is left in the data untouched, because in the overwhelming majority of cases it is a genuine collapse — a real bankruptcy, a real loss of value that actually happened to holders of the stock. Filtering those out as well would not have made this study more accurate. It would have made it look better than it was, by quietly deleting the losses a stock-picking rule is also supposed to be tested against.

Limits of the method

No backtest is free of assumptions, and being clear about them is part of the result, not an afterthought to it. Here are all ten, in full, none of them softened.

• Criterion 2 is a proxy. Analyst estimates from any point in the past do not exist as a usable time series, so this study substitutes the revenue growth a company actually achieved in its most recent financial year — a rear-view mirror standing in for a forecast that cannot be reconstructed after the fact.
• The insider criterion only exists from 2006 onward, because machine-readable SEC insider data does not reach further back. Before 2006, at most 9 of the 10 points are reachable at all: Portfolio B, which requires a perfect 10, is effectively only measurable from 2007, and Portfolio A, which requires 9, measures more strictly before 2006 than after it, since every one of the nine judgeable criteria then has to be met.
• Filing gaps hit failing companies hardest. A company with no usable financial statement gets no score, and a company with no score is never bought — and companies heading toward bankruptcy are disproportionately likely to stop reporting cleanly before the end. That biases the result somewhat too favourably.
• Trades execute at the month-end closing price of the signal month. In reality, days to weeks pass between a filing becoming public and an order actually being filled.
• The benchmarks are price indices without dividends, while every portfolio in this study trades dividend-adjusted prices — a difference that favours the portfolios in every direct comparison.
• Only order costs are modelled — 0.1% of order volume on every buy and every sell. Taxes, the bid-ask spread and the market impact of placing the orders themselves are not included.
• Two companies can share one ticker's price series (see Data quality above), and the quality gate built to catch that problem blocks upward jumps only.
• A rare data error distorts revenue figures. In 584 of 185,030 filing periods examined — 0.32% — an annual revenue figure appears to collapse in the raw data and then spring back the following year, a reporting artifact rather than a real business event. Roughly 1.5% of the scored months are estimated to be affected. The distortion runs in one direction, upward, and hits the K1 sell variant hardest, so its result should be read with some caution.
• Stops and profit-taking rules act on month-end closing prices only — the mildest possible way to measure them, since a real stop in a live account would trigger intra-month, more often and typically at a worse price than the one used here. The re-entry ban, too, is tested in exactly one length: six months. Whether a shorter or longer ban would work better is not something this study measures or claims.
• Twenty-five sell variants invite multiple-testing luck. Test that many rules against the same reference and one of them will come out ahead even without any real underlying effect. What is defensible is the direction of the finding — a re-entry ban helps, a bare stop does not — not the precise ranking or size of any individual row.

What does not follow from this study

What this study shows is how three fixed, mechanical rules would have behaved on a broad, survivorship-free universe of US stocks over roughly 26.5 years. It does not show what to buy today. A score of 9 or higher on a company right now tells you that the same rule, applied historically, produced a 13.17% average annual return across hundreds of different companies over more than two and a half decades — it says nothing about what that one company will do next.

The gap between those two statements matters most in the chapters on the score ladder and the sell-rule matrix. That the score sorts, on average, across thousands of data points does not mean any individual stock sitting at exactly 7 points today is a good investment; that a six-month re-entry ban improved the historical result does not mean it will do so again by the same margin, or at all, the next time markets behave differently than they did between 2000 and 2026. Past results, whether from a real portfolio or from a simulation like this one, are no reliable indicator of future returns.

This is a historical study, not investment advice. It contains no buy or sell recommendation, no price target and no forecast for any company, listed or delisted. Anyone making an investment decision should assess their own situation and risks, with professional advice where that is appropriate.

Two more studies from the same series work the same way: we ran a well-known value-investing formula through an identical kind of test in our Magic Formula backtest study, and rebuilt a classic deep-value screen from scratch in our Graham net-net backtest study. The Growth Gems score itself keeps running as a live, continuously updated screen in our stock scanner, and the full collection of studies like this one lives in Research.