Should you trust Meb Faber’s Ivy Portfolio rule? His table, checked on the 5 of its 40 years we could reach, and the 13 years after: the full report

The question this report answers is the viewer’s own: Should I trust my portfolio to a famous published rule, Meb Faber's once-a-month moving-average switch, when the only proof is the author's own table?

The paper we checked: Mebane T. Faber, A Quantitative Approach to Tactical Asset Allocation (free on SSRN; first a May 2006 working paper, updated in February 2013, the version we tested). He is named here as the author of the rule. This report tests his published rule and its published table, not a person, and it is not advice.

The rule, as Faber published it

Split the money into five equal slices, one for each kind of investment: US stocks, foreign stocks, 10-year US Treasury bonds, real estate and commodities. Once a month, at the month-end close, compare each slice’s price with its 10-month moving average, the average of its last ten month-end prices. Above the average, hold that investment; below it, hold Treasury bills in that slice. Each slice stays a fifth of the money; in our run the slices go back to equal each month, for the rule and for holding alike, as our plan set in advance.

His Figure 14 prints the result year by year for 40 years, 1973-2012, the years it trailed included. He reports 10.48% a year for the rule against 9.92% for holding all five, and a worst fall of -9.54% for the rule against -46.00% for holding. (A worst fall, or drawdown, is the drop from a high to the low after it.) His Figure 16 covers 2006-2012, the years after his first version came out, which he treats as out of sample: 6.01% a year for the rule against 3.94% for holding. In his words: “even though it only outperformed in three out of seven years, it beat buy and hold by over two percentage points per year”. He also writes: “On average, the investor is 70% invested”.

The name: in 2009 he co-wrote a book called The Ivy Portfolio, and on his blog he calls “the timing model we published in 2006” the Ivy Portfolio. The book also uses the name for a mix you hold without switching (his 2015 post), so we call the one we tested the Ivy Portfolio’s timing rule.

Our five funds stand in for his five indexes: SPY (US stocks), EFA (foreign stocks), IEF (7-10 year Treasuries, for his 10-year bonds), VNQ (real estate) and GSG (commodities). Funds that trade today follow his indexes closely but not exactly, so small gaps are expected.

The check: his table, on the 5 of his 40 years our data can reach

Before any year beyond his table counted, our code had to reproduce his table on the 5 of his 40 years our data reaches. The pass mark, written down and approved before our first recorded run: within 3 points of his figure in at least 4 of the 5 years; the sign right in every year where his figure is more than 1 point from zero; within 4 points in 2008, the crash year, no matter what; his holding column under the same limits, which tests our data, not just our code; and at most 2 attempts, with no widening of the limits afterward. Like his table, the gate charges no trading costs.

Why only 5 of his 40 years: the youngest of the five funds, GSG, starts in July 2006, and the 10-month average needs ten month-ends, so the rule’s first position is May 2007 and 2008 is the first full year. 2007 is only part of a year for us, and his 34 rows for 1973-2006 need his index data, which we do not hold. A side check with a second commodity fund, DBC, which starts a few months earlier, reaches 2007, one more of his 40 years. The plan made it a side check in advance, so its cells are never counted among the 5 of his 40 years we checked; graded after the results, both of its 2007 cells landed within 3 points, and it missed two cells in the 5 of his 40 years we checked (2008 in the holding column, and the sign of the rule’s 2011).

Download the 10 cells (CSV)

The result, on those 5 of his 40 years: all 5 of the 5 rule cells landed within 3 points, every sign was right, the 2008 gap was -0.89 points against the 4-point limit, and the holding column landed within 3 points in 5 of 5 with every sign right. The gate passed on those 5 of his 40 years, on the one attempt on record. His 2008 was -0.59% for the rule against -30.01% for holding; ours was -1.48% and -29.45%. The widest gap of the ten cells was 1.74 points (the rule, 2011).

The order, from the lab’s own repository (2026-06-09, US Pacific time, UTC-7): the plan at 20:34 (commit bd4a952), its approval at 20:40 (55973ce), the gate’s result at 20:43 (e43737b), the 2013-2025 run at 21:50 (629b80b). One commit holds the gate’s result, and no later commit edits it: that is the one attempt on record (self-reported; the repository is consistent with it). These are repository timestamps, not an independent timestamp.

What a five-year match can and cannot tell apart (measured after the results): random monthly switching almost never passes the gate’s first three clauses (0 of 1,000 random switchers that flip independently each month; 8 of 1,000 that hold a position longer), so the match separates his rule’s timing from chance. It cannot tell his fixed fifths from our own mis-built earlier bot, which makes the same monthly calls but spreads the money differently and also passes them (4 of 5 cells within 3 points, 2008 inside 4). So the match supports "our copy times the market the way his rule does, on five of his forty years", never "faithful" or "exact"; the fixed fifths rest on an independent re-implementation written from the plan’s text (see Can you trust our check?).

After the results we rebuilt the five funds from a second price source, our own survivorship-free store (dividends reinvested on the ex-date): the gate gave the same grades on the same 5 of his 40 years, 5 of 5 in both columns with every clause passed, and every cell within 0.4 points of the first run.

The 13 years after he published

We count from the end of his table, December 2012 (his last update came out in February 2013), and ran January 2013 to December 2025: 156 months, the same five funds, the trade at the month-end close, 5 basis points charged on every switch into or out of a fund. Holding means the same five funds in equal slices, rebalanced monthly and always invested. Before our first recorded run, the plan set three conditions: (1) beat Treasury bills by at least 1 percentage point a year; (2) a worst fall within two thirds of holding’s, or within 10%; (3) at least match holding on the Sharpe ratio, the return above Treasury bills per unit of ups and downs. None of the three asked it to beat holding on return. Those years had already happened when the plan was written: writing it first cannot make anyone unaware of what markets did; it stops us tuning the test. The plan named a loss on risk-adjusted return as “the plausible 2013-2025 bull-decade outcome”; that did not happen.

Condition 1: 5.05% a year against the bar of 1.67% for Treasury bills plus 1 point: met. Condition 2: a worst fall of 5.63% against two thirds of holding’s 20.66% (the plan writes 0.67, so 13.84%): met, well inside. Condition 3: a Sharpe ratio of 0.58 against 0.44: met. All 3 of 3 conditions were met, so by the plan’s own rule the test passed.

The three falling periods the plan named, each measured from its opening value (the corrected figures, found after the results):

It fell less in all 3 of the 3 falling periods the plan named, and both worst falls of the whole run came in the early-2020 crash. Each period is measured from its opening value. The lab’s own printout skipped each window’s first month, which changed the figures for two of the three periods, 4 of its 6 printed figures (2022 came out the same); these are the corrected figures, found after the results, and no pass or fail changed.

The 13 years, year by year (computed after the results, from the same registered run):

Download the 13 years (CSV)

The rule led holding in 5 of the 13 years (counted after the results: 2014, 2015, 2018, 2020, 2022) and trailed in 8; holding rose in all 8 of those years. Its biggest shortfall came in 2019, an extreme, not a typical year: 5.29% against 21.27%. In his 40 years, by our count of his table, it led in 17; in 2006-2012, in 3 of the 7. These yearly rows come from the same registered run, with its costs, and were computed after the results.

The trade, over the 13 years: it gave up 0.85 percentage points of return a year (5.90 - 5.05), before taxes, which we did not test, for a worst fall 15.03 percentage points shallower (20.66 - 5.63).

Why it trailed is an inference, not a measurement: on average about two thirds of the money sat in the funds (66.4% in 2013-2025; his paper puts it at 70% on average for his years) and the rest in Treasury bills, which earn less in a long rise. No registered arm isolates the cause.

The plan’s check for one fund carrying the result: take out the largest contributor to the rule’s return, US stocks, and re-run. The lab’s code ran the other four funds at a quarter each, for the rule and for holding, a choice the plan does not make. The rule then earned 3.76% a year against 3.70% for holding the four, with a worst fall of -6.06% against -20.98%, and all three conditions were still met. So the gap in worst falls does not rest on US stocks, and the return shortfall leans on them. It is a check on what drove the cost, not a better rule; the verdict stands on all five funds.

One receipt of how the protection works, from early 2020 (measured after the results; one episode, not a law): the rule was 100% invested in January (the rule -1.77% that month against -1.76% for holding), 80% in February (-3.93% against -5.51%) and 20% in March (+0.73% against -14.53%). It takes the first part of a fall and is mostly out for the deepest part.

His paper’s Sharpe ratios use a fixed 5.41% rate, and ours use the window’s own Treasury bills, so we do not set his and ours side by side.

What this does not tell you

• Two limits we say in the video. First, his test years and ours are different years on different data (his indexes; our funds), so they cannot tell us whether the rule got weaker after he published: in his window it beat holding, and in the 13 years after, on our run, it trailed. Second, this is one rule, on five funds, over one stretch of time, so it tells you nothing about timing rules or trend following in general.
• Nothing here says it will fall less in the next crash. Three falls are few, and if the protection was luck, they say little.
• Not tested: taxes; other windows; other investments, or more of them; our mis-built earlier bot as a verdict (it is only a side measurement).
• Granted without testing: his 35 earlier rows and the index data behind them; that our five funds stand in for his indexes with some tracking gap (the widest gate gap was 1.74 points); his paper’s other variants of the rule.

Can you trust our check?

Our fund prices are Yahoo Finance’s adjusted closes (dividends included), pulled on 2026-06-09, not our survivorship-free store, which keeps companies that died. That costs nothing here: five funds that all still trade leave no dead fund to miss. The dividend adjustment is real: ten years before that pull, SPY’s adjusted close sat 15.0% below its raw close (178.64 against 210.27, a figure restated after the results: the lab’s prose gave the ratio the other way round). Treasury bills come from one source, the Federal Reserve Bank of St. Louis series TB3MS.

Rebuilt from our store after the results, the 2013-2025 return and worst fall matched to two decimal places (5.05% and 5.90%; -5.63% and -20.66%), and every grade it re-ran held (the gate’s four clauses, the three conditions and the three falling periods; the check without US stocks and the DBC side check were not re-run on the store). Both runs share the one Treasury bill source.

After the results, the lab’s work was re-checked rather than copied. An implementation written from the plan’s text, without the lab’s own engine, re-derived every printed value it checks in the lab’s two result files: 82 of 82 line checks, covering 81 distinct figures (the 2008 gap is printed twice, in the table and on its condition line, and checked on both), each on its printed line at its printed precision. 8 of the 82 matched only when it copied two of the lab’s own departures from the plan (the 4 falling-period figures and the 4 next-open figures): problems (1) and (2) below. Not among them: his figures and the plan’s thresholds, quoted there for comparison (inputs); the side check’s 12 printed gaps (that fund’s cells minus his); the pass counts; and the other repeats on the condition lines. The lab’s own scripts re-ran byte-identical on 2 of 2 result files. For this report we ran that re-implementation again: its printout came back byte-identical.

That re-check, after the results, found four problems in the lab’s printouts, all corrected here, none changing a pass or fail: (1) the falling-period figures skipped each window’s first month, which changed 4 of the 6 printed figures, in two of the three periods; (2) the cost of trading at the next month’s first open instead of the month-end close was printed without trading costs; with them, the rule earns 4.74% a year (0.31 points less) with a worst fall of -7.16% (1.53 points deeper); (3) the side check with the second commodity fund was never graded; graded, it misses two cells and stays outside the 5 of his 40 years we count; (4) a line saying no ten-year stretch passed without a switch was printed, not computed; computed, the longest stretch without a switch in any one fund is 43 months over the whole signal history (April 2007 to June 2026), 32 inside 2013-2025. It also completed one note (the monthly return to equal slices is uncharged for the rule as well as for holding; charging it moves the yearly return by under 0.02 points and no grade) and corrected one ratio in the lab’s prose: the adjusted close is 15.0% below raw, not 17.7% (17.7% is raw over adjusted).

What we can say, and how strongly: 14 statements, graded

Each row is one statement about this test, with how strongly the record supports it and what it rests on. Measured means a run, a count or the repository’s order shows it; self-reported means it rests on our own record, which is consistent with it; inferred means it fits the numbers but no registered arm isolates it.

Download all 14 graded statements (CSV)

Every figure said in our video, with its source

Here are all 68 figures our narrator says, one row per figure, with the words around each saying, the parts of the video where it is said, what it is, how we got it and its source. A figure here is a cardinal number, a decimal, a signed figure, a year, a fraction (fifth, quarter, two thirds) or the count word once; ordinals (first, second, third) in any use are not figures, and neither are both, none and single. 7 uses of number words in the narration are not figures and have no row: “checks each one against” (a pronoun (each slice)); “the lesser half of his claim” (one of the claim’s two parts, not a quantity); “The bigger half is” (one of the claim’s two parts, not a quantity); “we call the one we tested” (a pronoun (the rule)); “not a typical one” (a pronoun (a year)); “the bad ones included” (a pronoun (the years)); “the bad ones included” (a pronoun (the years)). Every number shown on screen, in digits or in words, is one of these figures, matched by value and by the part of the video where it is said, then read by eye, with these exceptions: two numbers are shown in the part after the one where they are said (two thirds, shown once in Part 6 and said in Part 5; four, shown three times in Part 7 and said in Part 6); two numbers the narrator does not say, ten-year (TEN-YEAR TREASURY BONDS, his 10-year bonds) and one (OURS: ONE FUND EACH, one fund per index); and four uses of words that are not figures, as in the narration (SMALLER HALF; BIGGER HALF; BAD ONES INCLUDED; FUTURE ONES).

The parts of the video: Cold open · Part 1: What does the rule promise? · Part 2: Where does it come from? · Part 3: Can you check Faber's table? · Part 4: Did Faber's table check out? · Part 5: What happened after he published? · Part 6: Did it hold up when prices fell? · Part 7: What did the protection cost? · Part 8: Can you trust our check? · Close: How do you check the next table?

Download all 68 spoken figures (CSV)

The plan, written before our first recorded run

Written, committed and approved before our first recorded run (the order is above), and not changed after the first recorded result (its only two commits come before it). Its main terms, in plain words (the plan also carries alarms for a broken run, which did not fire, and two information-only checks of the gate window):

• The rule exactly as published: the 10-month moving average on month-end total-return closes; fixed 20% slices; Treasury bills in a slice whose investment is below its average; the trade at the month-end close, on the same bar as the signal.
• The universe: SPY, EFA, IEF, VNQ and GSG; a second commodity fund, DBC, as a side check only. Signals use closed months only.
• The benchmarks: holding the same five funds (equal slices, rebalanced monthly, always invested), and Treasury bills.
• The pass mark for his table (2008-2012, the 5 of his 40 years our data reaches; no trading costs, as his paper): within 3 points in at least 4 of the 5 years; the sign right wherever his figure is more than 1 point from zero; 2008 within 4 points; his holding column under the same limits; at most 2 attempts; no widening afterward.
• The three conditions for 2013-2025, with 5 basis points on every switch into or out of a fund: beat Treasury bills by at least 1 point a year; a worst fall within 0.67 times holding’s, or within 10%; a Sharpe ratio at least holding’s. All three must hold.
• The middle case, ruled in advance: if the rule beat cash and kept its worst fall inside the limit but missed the Sharpe condition, that is a partial pass on risk only, counted as a pass only if the gate on his table had passed and the rule fell less than holding in at least 2 of the 3 named falling periods. It did not arise: all three conditions were met.
• Its checks: the three named falling periods (August 2015 to February 2016, January to April 2020, 2022); each investment’s share of the return, and a re-run without the largest contributor; trading at the next month’s first open instead of the month-end close.
• Four choices fixed in advance where the plan judged his paper silent or different: holding is rebalanced monthly, like the rule’s slices (a choice the plan made in advance, calling his paper silent on holding’s rebalance); the Sharpe ratio is over the window’s own Treasury bills (his paper uses a fixed rate); IEF (7-10 year) stands in for his 10-year bonds; 5 basis points on every switch into or out of a fund, a cost his paper leaves out.

Deliberately not tested: taxes; other windows; other investments; the bot we built from the rule earlier, as a verdict. The design’s main terms, the commit ids, the registered results and what was added after them are in the locked-design file (JSON).

What would change our mind

Five things would reopen this test; the strongest is his earlier years on index data:

1. His earlier years, with index data: total-return series for his five indexes, running the same gate on 1973-2007. Material misses would shrink "his record is checkable" to "the tail of his record is checkable". This is a data purchase, not a re-run.
2. A crash where the rule falls further than holding: a new observation on the claim that held, and the claim to watch, because the return claim already trailed.
3. New rules or results from Faber: we would re-check our copy against them, then re-run.
4. A mistake in our work: if a re-run of our re-implementation gave a different grade, this test would reopen (its code and the 2026-06-09 price pull are not published with this report).
5. A fault in the single-source Treasury bill data: an error there would move the rule’s cash slices, the Treasury bill arm and every Sharpe ratio at once.

Sources

Every source behind a figure in this report:

• Mebane T. Faber, A Quantitative Approach to Tactical Asset Allocation (2013 update; first a May 2006 working paper), free on SSRN: the rule, his Figure 14 and Figure 16 (read off the page images), his words, and the paper’s dates: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=962461
• Faber’s blog, 2011-11-03, Measuring the Performance of the Ivy Portfolio: the name: https://mebfaber.com/2011/11/03/measuring-the-performance-of-the-ivy-portfolio/
• Faber’s blog, 2015-06-02, Chapter 9, The Endowment Portfolio: his 2009 book and its never-switching mix: https://mebfaber.com/2015/06/02/chapter-9-the-endowment-portfolio-swensen-el-erian-and-ivy/
• Yahoo Finance adjusted daily closes for SPY, EFA, IEF, VNQ, GSG and DBC, pulled on 2026-06-09: the fund prices behind our runs (this pull is not published with this report)
• Federal Reserve Bank of St. Louis (FRED), 3-month Treasury bill rate, TB3MS: the Treasury bill leg: https://fred.stlouisfed.org/series/TB3MS
• Our own survivorship-free store (a market-data vendor’s daily bars and dividend records): the second price source, after the results (not published with this report)
• The lab’s own record and our port audit of it (after the results), neither published with this report: the plan, the gate’s result, the 2013-2025 run, and the repository’s commit ids and times

Downloads: the 14 graded statements · the 68 spoken figures · the 10 gate cells · the 13 years · the locked design.