Home/Learn/Backtests/Graham defensive investor

SledgeKey Backtests

We ran Graham’s defensive investor screen honestly. It was level with the index after seven years and thirteen percent behind after ten.

Published August 15, 2026 · Revised October 1, 2026 · By The SledgeKey Team · Window September 15, 2016 to September 13, 2026

On the rebalance date in September 2017 this screen ran against every operating company listed on the NYSE and NASDAQ and returned ten names, six of them insurers: Assured Guaranty, Unum, Reinsurance Group of America, Old Republic, First American and Allstate, alongside D.R. Horton, Reliance Steel, Avnet and Universal Corporation. The portfolio had room for twenty positions and a ten percent cap on any single one, so ten names filled it exactly, and the screen would not find a full twenty until September 2021, because Benjamin Graham’s rules had nothing else to say yes to.

Graham set out those rules for what he called the defensive investor: buy companies of decent size, conservatively financed, reliably profitable, paying a dividend, and cheap against both earnings and assets. We turned that description into five filters and ran them once a year for ten years against point-in-time data, meaning each rebalance saw only what had actually been filed and published by that date. Companies that later delisted stayed in the universe at every date they traded, so the run owns every exit, the failures and the buyouts alike.

Over the ten years to September 2026 the screen turned $100,000 into $362,352, an annualized 13.75 percent. The S&P 500 turned the same $100,000 into $416,779, an annualized 15.35 percent. Graham’s rules trailed by 1.60 points a year, and the ride was considerably rougher than that gap suggests: a 49.9 percent peak-to-trough fall against the index’s 33.7 percent, on annual volatility of 21.2 percent against 14.3 percent.

The exact screen

The full configuration is below as it ran, so anyone who wants to reproduce the result or argue with the setup has everything needed to do either.

Two rows in that table do more work than they look like they do. The universe line matters because a value screen built on price to book will happily buy a closed-end fund, since a fund trades near the value of the securities it holds and therefore posts a low price to book with a clean balance sheet. Five of them turned up in an early version of this run, along with a $25 par bond listed under its parent insurer’s name and carrying that insurer’s $12.8 billion market cap. All of them are excluded at the universe level now.

The liquidity floor is newer. Every date in the universe now requires a median daily dollar volume of at least $100,000, measured over the previous sixty trading sessions. A median rather than an average matters here, because one block trade in an otherwise dormant listing can lift an average far enough to disguise a security that nobody can actually buy. This screen filled all twenty slots on each of its last five rebalances.

The result against the index

The capture ratios are the plainest summary of what the ten years delivered. An upside capture of 98.72 means the screen rose a little less than the index did in the index’s up months. A downside capture of 108.23 means it fell a little more than the index in the down months. A portfolio that lags slightly going up and loses a little extra going down will finish behind, and that is the whole mechanism behind the 1.60 point annual gap. Whatever else Graham’s criteria bought over this decade, they did not buy protection.

Both drawdown figures above are measured on the daily account value rather than on month-end snapshots, and the difference is large enough to be worth stating. Sampling this same strategy once a month puts its worst fall at 37.1 percent, because a monthly observation can step straight over the trough of a fast crash. The daily path finds 49.9 percent. Comparing a daily strategy drawdown against a monthly index drawdown, or the reverse, would misstate the comparison by several points, so both sides here use the daily path.

Volatility, Sharpe and the other dispersion statistics move by a point or so depending on exactly where a ten-year window starts, because a monthly return series samples a crash somewhat arbitrarily. The window above is pinned to September 15, 2016 through September 13, 2026 for that reason, and every figure on this page comes from that single run.

The years it could not find twenty companies

The most interesting thing in this run is not the return. It is how often the screen refused to buy. Graham’s criteria are absolute rather than relative, so when the market gets expensive the screen does not lower its standards and pick the best of a bad set. It returns fewer names, and with fewer than ten of them the ten percent position cap starts to bind and the remainder sits in cash.

For the first five rebalances the screen never once filled the portfolio, and September 2019 is the low point, with nine companies qualifying out of the entire NYSE and NASDAQ. The portfolio did not hold a full twenty until September 2021, after the crash of early 2020 and the recovery had reset a great many valuations. Cash is a drag in a rising market, but here the drag was small: the position cap only bites once the count drops below ten, which happened once, and a tenth of the portfolio sat in Treasury bills through the year to September 2020.

It is worth being precise about how much cash that actually was, because it is easy to overstate. The engine equal weights across the names it finds and caps any one position at ten percent, so thirteen qualifying names in 2016 still produced a fully invested portfolio at roughly eight percent each, and the ten names of 2017 and 2018 produced one at exactly ten percent each. Cash only appears once the count drops below ten. That happened once in ten years, in 2019.

Whether the empty slots are a flaw depends on what the rules are for. An investor who wanted to be fully invested at all times would find this screen useless in an expensive market. An investor who believed Graham’s point, that paying too much is the reliable way to lose money, would read the empty slots as the strategy working as designed.

Where the ten years actually turned

The screen was ahead for the first year. On the September 2017 rebalance date it stood 1.4 percent above the index, and measured from one rebalance date to the next it beat the index in five of the ten years, which is not what a 1.6 point annual shortfall usually looks like from the inside. What follows is a sequence rather than one turning point, and the chart shows it more plainly than any single statistic does.

The heaviest single episode is the crash of early 2020. Between the February and April observations the screen fell 36.0 percent against 17.2 percent for the index, and on the daily path it fell 49.9 percent. Deep value screens concentrate in cyclicals, financials and commodity producers, and going into that February this portfolio held nine names: three insurers, a retirement and asset management group, an asset manager, two steel producers, a miner and a homebuilder. Those are precisely the businesses that get repriced hardest when the world stops moving. By September 2020 the screen stood at $115,508, up 15.5 percent in four years, while the index sat at $170,842. The cumulative shortfall against the index peaked at 38.4 percent that April.

Then it clawed the whole gap back. The year to September 2021 returned 70.5 percent against 33.6 for the index, and by December 2023 the screen stood at $250,142 against the index’s $248,446, level after seven years. The entire shortfall this page reports came after that. Over the twelve months to December 2024 the screen returned 17.9 percent against 29.9 for the index, and over the twelve months to December 2025 it returned 6.0 percent against 14.0. A 20.5 percent run over the twelve months to September 2026, against 16.9 for the index, closed some of it, and the cumulative shortfall ends the window at 13.1 percent. A decade is long enough to argue from and short enough that the ending point still moves the story.

What this backtest does not prove

A page like this earns its credibility in this section, so here is what the numbers above do not establish.

The strategy lost, and it lost while being considerably more painful to hold. Thirteen and three quarter percent a year is a respectable absolute result, and it is still 1.60 points a year behind a low-cost index fund that required no screening, no rebalancing and no conviction. The drawdown was sixteen points deeper than the index’s, the volatility half again as high, and the downside capture above 100. Nothing here argues that this screen is a better default than owning the index.

These are Graham’s criteria interpreted rather than transcribed. Graham wrote about earnings stability over ten years, uninterrupted dividends over twenty, and moderate price to earnings measured against average earnings rather than trailing. We used five filters that are available point-in-time across the whole universe. They capture the spirit of the defensive investor tests, and a stricter reading of the original would produce a different and probably smaller set of companies.

One window and one configuration. This is a single ten-year run with one set of thresholds, twenty holdings, equal weighting and annual rebalancing. Change the rebalance frequency, the holding count or the market cap floor and the result moves. A ten-year window that happens to contain one violent crash and a long bull market is not a general claim about how Graham’s rules perform.

The largest-cap tiebreak is a real choice. When more than twenty companies qualified, the twenty largest by market cap were selected. That is one rule among several sensible ones, and it tilts the portfolio toward bigger and more liquid names. Ranking by cheapness instead would give a different portfolio and a different answer.

Cash was held at Treasury bill rates. In the one year the screen could not fill ten positions, the unallocated tenth accrued interest at the prevailing three-month bill rate. That is a reasonable assumption and it is still an assumption.

These figures replace two earlier versions of this page, published on August 15 and September 2, 2026. The September version reported 12.99 percent a year and led with a September 2017 rebalance that found only four companies. Later that month we found and repaired a fault in our fundamentals history: for roughly 1,800 companies, later filings had been copied onto earlier dates, so some screens between 2020 and 2024 were reading financials that did not yet exist, and some earlier screens were reading ratios that had been mis-scaled. On the corrected data the 2017 screen finds ten companies, not four, and every number on this page was regenerated. We would rather show the revision than pretend the earlier numbers never existed. Where the versions disagree, this page is the current one. On October 1, 2026 the capture ratios were recomputed with Morningstar's method; the earlier figures, 96.44 and 104.00, divided cumulative returns, a calculation that drifts with the length of the window.

Concepts and metrics used in this backtest: Survivorship Bias · Point-in-Time Data · Max Drawdown · Price to Earnings · Price to Book · Current Ratio · Debt to Equity · Dividend Yield.