The Graham Enterprising Backtest: The Criteria Don't Beat Their Own Universe

The Result in Four Numbers

• 12.19% a year is what the primary run — Graham's Chapter 15 criteria without the dividend rule, excluding financials, rebalanced annually — delivered from January 2000 through July 2026.
• 8.52% a year is what the S&P 500 total return index (including dividends) delivered over the same period — a 3.67-percentage-point lead.
• 12.27% a year is what an equal-weight basket of EVERY investable stock in the same cleaned universe delivered, with no screening at all — 0.08 points MORE than Graham's selection.
• 0.28% of all monthly checks across the entire universe satisfied Graham's original criteria at all — a far rarer hit rate than the roughly 3% Graham himself found in his own 1970 sample.

Graham's screens added no value in this backtest over the equal-weight basket they were selected from.

Chapter 15: Seven Screens for the Enterprising Investor

In "The Intelligent Investor," Benjamin Graham distinguishes between the defensive and the enterprising investor. Where the defensive investor relies on a few very strict criteria, the enterprising investor is willing to put in more work and take on more risk — in exchange for a broader, mechanically screened set of undervalued stocks. Chapter 15 of the 1973 edition ("Stock Selection for the Enterprising Investor") lays out seven quantitative screens for that purpose, all derived from public balance-sheet and price data, none of them a forecast:

1. Price-to-earnings ratio below 9 to 10 (Graham's Stock Guide exercise used ≤9; this study's primary run uses ≤10, with ≤9 as a sensitivity check).
2. Liquidity: current assets at least 1.5 times current liabilities.
3. Debt: interest-bearing debt no more than 110% of net current assets (current assets minus current liabilities) — written for industrial companies.
4. Earnings stability: no loss year in the past five fiscal years.
5. Earnings growth: the latest annual profit exceeds the profit from roughly four to five fiscal years earlier.
6. Dividend: some current payout.
7. Price: below 120% of tangible book value (equity minus goodwill minus intangibles).

"An investment operation is one which, upon thorough analysis, promises safety of principal and an adequate return. Operations not meeting these requirements are speculative."

That definition is exactly what sits behind the seven Chapter 15 screens: a mechanical, checkable "thorough analysis" in place of a hunch. There is no minimum size — Graham's own 1970 Stock Guide sample took names trading for as little as $1. And Chapter 15 gives no exit rule anywhere: no price target, no holding period, no stop. This study's primary run treats that as annual rebalancing — a portfolio rebuilt from scratch once a year against those criteria.

Setup: Survivorship-Free Over 26 Years

The backtest covers January 2000 through July 2026 and evaluates 319 monthly checkpoints across the entire U.S. stock universe — 1,643,857 individual metric checks. The universe covers 22,871 stocks, of which 16,472 (72.0%) have since disappeared from the exchange through acquisition, bankruptcy, or withdrawal. Counting only the survivors flatters the past; this backtest keeps every company that existed and reported data at the time of a given check, regardless of what later happened to it.

Before the criteria check, some stocks drop out for purely data-side reasons — that says nothing about their valuation, only that a check wasn't technically possible. As of the last checkpoint (2026-07-31), for instance: 4,412 stocks checked as investable, 507 rejected for ticker recycling, 515 for a price series flagged as corrupted, 469 for currency issues, 105 for a stale filing, 4 with no market cap.

The Result: 12.19% a Year

The table below shows the primary run — Graham's original criteria excluding financials and ADRs, rebalanced annually, 0.2% cost per side, delisted positions closed at the last traded price, 12 offset cohorts — against two reference series.

At first glance, a clear win: 12.19% a year over 26 years, a 3.67-point lead over the S&P 500, ahead of it in 56% of individual calendar years. The 62.52% maximum drawdown, though, shows this return doesn't come free — a portfolio holding through the financial crisis or the years 2007 and 2018 lives through losses that can erase two-thirds of its value.

The Most Important Finding: The Criteria Don't Beat Their Own Universe

This is the crux of the study, and it belongs front and center rather than buried in a footnote: a simple, equal-weight basket of EVERY investable stock in the same cleaned universe — bought with no price-to-earnings check, no balance-sheet screen, none of the seven criteria — returned 12.27% a year over the same period. Graham's selection returned 12.19%. The seven screens therefore trail the universe they select from by 0.08 percentage points a year — a gap that, given the noise inherent in a 26-year backtest, is effectively zero, but is not a lead either.

That is the honest answer to what Graham's seven rules actually earned in this test: nothing extra. The lead over the S&P 500 (8.52%, +3.67 points) looks substantial — but it comes mostly from the tilt toward smaller, equal-weighted stocks, a long-documented effect with nothing to do with balance-sheet analysis. The S&P 500 consists of the country's largest, most established companies; this backtest's investable universe spans every size class. Pitting two such different risk classes against each other inflates how impressive the screens look. The honest yardstick is the equal-weight universe — and measured against that, nothing is left of Graham's seven criteria.

One addition that shouldn't get lost: the equal-weight universe is a yardstick, not an investment proposal. It carries the same data guards, the same financials and ADR exclusions, and the same $1 floor as the primary run — the comparison says the screening rule itself adds nothing, not "buy the whole equal-weight market instead".

Why Financials Are Excluded — the SVB Story

Graham's balance-sheet rules — current assets at least 1.5 times current liabilities, interest-bearing debt no more than 110% of net current assets — are explicitly written in Chapter 15 for industrial companies. A bank doesn't carry that same short-term/long-term split; where the data source still reports both line items, the resulting coverage ratios come out absurdly high, and the rule stops checking anything meaningful.

The textbook case: SVB Financial Group — the parent of Silicon Valley Bank, which collapsed in March 2023 — would have shown a calculated coverage ratio of 681 in January 2023, a figure that's simply impossible for an industrial company. The criteria would have bought it — five weeks before its collapse. That's why financials (SIC 6000-6999) are excluded from the primary run; a sensitivity run including them is shown alongside.

That sensitivity run needs a caveat, though: at 12 months holding, no stop, 0.2% cost, and delisting at the last price, the financials-included run returns 14.45% a year (4,126 positions, 66.13% maximum drawdown) — higher than the primary run. But the largest single position in that basket is CVCY (Central Valley Community Bancorp), returning 1,794.66% in six months — a figure traced to a known defect in the underlying price series, not a real business outcome. The financials-included run is therefore an upper bound only, not a reliable result.

Why ADRs Are Excluded — the China Mobile Story

Several American Depositary Receipts report their filings in local currency while carrying a dollar sign in the currency field — the first-stage currency guard can't catch that. China Mobile ends up reporting a "profit" of 119.64 billion supposed dollars (actually renminbi), producing a price-to-earnings ratio of 1.07 instead of roughly 11. Other affected hits include Daqo (0.56), Jupai (0.30), and China Index Holdings (0.33).

Exclusion runs off the data source's country-of-origin flag — a measured attribute, not a guess. This filter costs some solidly reporting names too (Criteo, Celestica, Silicom, for instance) and isn't airtight: 7,326 stocks in the universe carry no country-of-origin flag at all and could let an ADR through unrecognized. A guessed plausibility rule — say, "suspiciously low P/E" — was deliberately not built: it would throw out exactly the genuine deep-value hits this backtest exists to find. Graham's own Stock Guide exercise is also a purely domestic U.S. universe.

Of the 4,584 raw hits under the original recipe (357 stocks), 1,916 stock-months (41.8%) were financials and 458 (10.0%) were ADRs; 78 (1.7%) with no industry code stay in deliberately — excluding for lack of a code would systematically hit the oldest and smallest names, exactly the ones a deep-value recipe targets. What's left for the primary run: 2,275 buys across 179 stocks.

Year by Year

The primary run's annual returns against both reference series show how unevenly the return is distributed. Only complete calendar years are counted — the first and last year of the period are therefore missing.

The primary run led the S&P 500 in 56.00% of the 25 calendar years shown, and led the equal-weight universe in 64.00% of them. Two notably weak years stand out — 2018 (−40.90%) and 2025 (−36.87%) — both well below either reference series, suggesting the criteria disproportionately land in stocks that fall especially hard in certain market phases.

Individual Trades: Hit Rate and Outliers

Across all 2,216 completed trades in the primary run (gross, before costs, so the figures stay comparable across cost variants):

The share below −90% checks whether the backtest actually captures the failures — 1.44% is plausible for a deep-value recipe that deliberately buys battered companies. The share above +1,000% sits at zero — the counter-check against glued price series comes out clean for the primary run (unlike the financials-included sensitivity run above). The best single position in the primary run is GIII (G-III Apparel Group), bought in February 2009 at $1.68 and sold a year later at $10.38 — a hit right in the middle of the post-financial-crisis recovery. Of the 2,216 trades, 2,193 closed via the annual rebalance, and 23 (1.04%) via delisting.

How Rarely the Criteria Trigger Today

Across all 319 checkpoints and the entire universe, 1,643,857 individual metric checks were run. Only 4,584 of them (0.28%) satisfied all six original criteria applied in the primary run at once, spread across 357 distinct stocks. (Graham's dividend rule is left out for data reasons; see "Limits of This Study.") Graham himself reports a hit rate of roughly 3% for his 1970 Stock Guide sample. By this backtest's measure, today's hit rate is roughly an order of magnitude lower than it was more than fifty years ago — Graham's screens together have become a substantially higher bar, whether through higher average valuations, stricter accounting, or simply a different market shape than in 1970.

Recipe Sensitivity: Tighter P/E, Minimum Size, Dividend Requirement

Three variants of the original recipe were tested, each at 12 months holding, no stop, 0.2% cost, delisting at the last price:

A tighter price-to-earnings ratio (≤9 instead of ≤10, as in Graham's Stock Guide exercise) slightly lowers the return — a tighter filter doesn't automatically catch better stocks. A $50 million minimum market cap lifts the return to 14.49% and cuts the drawdown to 59.43% — larger, more tradable names perform better in this test, but cost around 450 of the 2,275 positions. The dividend proxy (a cash-flow line item rather than actual per-share dividends) comes in at 13.60%, above the primary run, but on a weaker data foundation — only about half of all filings carry that figure at all.

Holding Period and Stop Sensitivity

Chapter 15 gives no exit rule, so the primary run uses annual rebalancing. For context, 6- and 24-month holding periods and two fixed stop rules were also tested (original recipe, 0.2% cost, delisting at the last price):

A longer 24-month holding period would have lifted the return to 13.18% while also lowering the drawdown — Graham's criteria apparently need time to pay off, and an annual turnover sells some positions too soon. Both stop rules cut the return meaningfully while also cutting the drawdown substantially: a portfolio that sells after −20% misses part of the recovery that deep-value stocks typically depend on for their return — the same mechanism observed for stop rules in other contrarian strategies across this backtest series.

Costs and the Delisting Assumption

For the primary run (12 months, no stop), three cost levels and two delisting assumptions were tested:

Trading costs of 0.2% per side cost the primary run 0.44 percentage points a year (12.63% versus 12.19%) — moderate, because the delisting rate is low at 1.04% and annual rebalancing trades less often than shorter holding periods. The pessimistic "delisting = total loss" assumption, versus "last traded price," costs a constant 0.99 percentage points across every cost level — the truth is likely somewhere between the two, since not every delisting means a total loss, but not every one means the last exchange price either.

Data Quality: This Backtest's Biggest Gap

The most important caveat of this entire backtest sits in the two bold rows: for the vanished companies, less than half even have an annual filing on record, against 93.9% for stocks still trading. The backtest simply can't form a verdict on more than half of all delistings — and that's exactly the group where the worst failures live. The direction of the gap is known, though: a stock with no filing never enters a buy list, so the worst cases are more likely missing from the test than inflating it. If anything, the result skews too favorable rather than too harsh.

Price-Series Quality and Ticker Recycling

The price data source carries two different companies under the same delisted ticker in one continuous monthly series for some names, with no split evidence. The check runs against the adjusted price; exclusion only runs upward, because a genuine price collapse is real more often than not and should stay in the test — otherwise a value recipe would lose exactly the failures it's supposed to experience in a survivorship-free universe.

The duplicate guard additionally checks whether the filings and the price series actually belong to the same company. One finding from the first real run: the data carries BBBY_OLD with Bed Bath & Beyond's filings and Overstock/Beyond Inc's price series, which also appears a second time under BYON — without the guard, that would have produced a hit in November 2010 with a price-to-earnings ratio of 6.7 instead of 18.8. A second finding: NISN and AIOS share 110 identical months of raw price data but different end dates — a simple full-series match would have missed that, and the backtest would have held the same company twice in the portfolio for seven months of 2018. Both sides of a detected duplicate are dropped, because the data doesn't say which series belongs to which company.

A separate, opposite-direction pitfall concerns tangible book value: in 8.9% of filings, the source reports goodwill and intangible assets at the exact same figure to the cent — meaning goodwill is already baked into intangibles and gets subtracted twice when computing book value. Tangible book value then comes out too small, Graham's seventh criterion becomes too strict, and this error works against the strategy, not for it.

Limits of This Study

Seven caveats belong in any honest account of this result:

• The criteria trail their own universe — the most important finding of this study, discussed above. The lead over the S&P 500 comes mostly from the size and equal-weight tilt, not from the seven rules themselves.
• Delisted stocks are underserved. Only 46.5% of them have annual filings at all, against 93.9% for stocks still trading — the direction of the gap skews the result too favorable rather than too harsh.
• Financials are excluded. The sensitivity run that includes them is an upper bound only, not a reliable result, due to a known price-series defect (CVCY).
• ADRs are excluded, but not completely: 7,326 stocks with no country-of-origin flag could let one through unrecognized.
• The dividend rule is missing from the primary run. The dividend-proxy sensitivity run rests on weaker data (only about half of filings carry the needed figure).
• The delisting assumption is a range, not a point estimate. Between "last traded price" and "total loss" sits a constant 0.99 percentage points a year.
• Annual filings instead of trailing twelve months. The data source only carries annual figures; between two filings the metric can be up to a year stale, while Graham's Stock Guide worked from the most recently reported twelve months.

The strategy tested here runs as its own continuously updated scanner. For the related two-thirds NCAV rule from the same chapter, see The Graham Net-Net Backtest. Further in-house backtest studies are available in the Studies section.

Figures as of August 7, 2026.

This article is a historical analysis of a publicly known investment rule and does not constitute investment advice. It contains no buy or sell recommendation, no forecast, and no statement about any individual, currently listed company. Past results — whether simulated or real — are not a reliable indicator of future returns. Anyone making investment decisions should assess their own situation and risks, seeking professional advice if in doubt.