EDIT: as per the comments from sajornet below, there is a major error in the article, and some of the conclusions drawn here don't have the support I thought they had, At the same time, there is one more aspect to check on as mentioned in my response to sajornet. Readers seeing this post for the first time may want to wait until that's done, I will update here again when that has happened. I apologize for my error.

*****
Hi all, while working on my "DIY hedge fund" portfolio policy, I investigated whether there is a safe way to use leverage. (This post is a substantially updated version of a previous post I made in /LETF. Although I use U.S. ETFs in this, the same logic applies to European investors using European leveraged ETFs/ETNs.)
 
According to various versions of the adventurous hedgefundie's strategy, a balanced portfolio of two or three 3x leveraged ETFs "only" drew down in the vicinity of max 70% over the available simulated history. But what has not happened yet might just happen in future. A couple of months where everything - equities, bonds, gold - drop hard together could push the drawdown into 80% territory. Maybe even worse. I wanted to see if I can add something of value here. Something that goes beyond relying on diversfication alone to keep things on the rails.

Headline result, equities-only: in a test running from May 1977 through June 2026, a macro-economic rule applied to simulated all-world equities produced a CAGR of 14% annual return, compared with 13% for continuous 2x exposure. Maximum drawdown fell from 87% to 60%. The rule changed allocations just 28 times in almost half a century. (I'm personally more interested in all-world equities than US-only but I did also test on SPY.)

TL;DR:

• The rule holds 2x equities in a “Goldilocks” macro regime and ordinary, unlevered equities otherwise. This rule never moves equities to cash!
• Across the tests described below, it produced higher Sharpe ratios than both permanent 2x exposure and a 10-month moving-average rule evaluated monthly using the same 1x/2x choices.
• Returns were broadly competitive with both alternatives. Goldilocks had smaller maximum drawdowns and required significantly fewer allocation changes than the moving-average rule.
• The signal was developed for another purpose. None of its parameters were chosen to fit leveraged-equity returns.

My portfolio policy has additional risk filters that directly address recessionary bear markets and other shocks, separate from the macro signal modulating leverage. Those other risk filters do move equities to cash and other defensive assets on rare occasions.

Where this came from

The Macro Regime Overlay ("MRO") originally had a much bigger job in my project. I wanted it to adjust the weights of the different sleeves inside the portfolio according to the prevailing macro conditions. So, shift weights among value, momentum, ballast, bitcoin, and other allocations. A sleeve’s percentage of total NAV would rise or fall depending on what the overlay signalled was happening in the economy.

The whole mechanism became far too involved for my taste. Lots of moving parts, lots of trading, not sufficiently impressive gains. So, I abandoned that approach.

Then, later, I wondered if maybe I could give the overlay a smaller job. Might it signal when to use leverage on an equity allocation that I already wanted to hold?

As mentioned, the overlay was never designed with leveraged investing in mind. No observations of leveraged returns went into its design. I didn’t optimize it for its original purpose, and I haven’t optimized it for this one. The thresholds, smoothing and confirmation rules have never changed since inception.

The backtests aren't out-of-sample but there is no curve-fitting either.

What does the overlay actually measure?

The MRO looks at economic activity, inflation and the yield curve. It sorts conditions into Goldilocks, Reflation, Stagflation and Slow Growth, with a further distinction for yield curve stress.

For the leverage decision, only Goldilocks matters. It requires two things:

Economic activity is at or above its recent historical average.

Inflation is neither too cold nor too hot.

Activity is measured using the Chicago Fed National Activity Index, smoothed over three months and compared with its own trailing ten-year history. I use a rolling z-score, with zero as the cutoff.

Inflation is the year-on-year change in core US PCE, also smoothed over three months. Between 1% and 3%, inclusive, is the acceptable band. I use headline CPI as fallback if core PCE is unavailable, with a slightly higher band of 1.4% to 3.4%.

The US yield curve helps distinguish the other regimes, but doesn't veto Goldilocks. So despite the broader overlay’s name and ambitions, the allocation decision here comes down to economic activity and inflation.

The signal is evaluated monthly, using prior-month (or older) observations. Two consecutive monthly signals must agree before the allocation changes, in either direction. So yes, I used US macro signals and applied them to world equities! I also backtested SPY/SSO the same way, for another reference point.

If anyone wants the details for the MRO, just ask, I'll post them in the comments.

What the backtests show

I used simulated VT history from testfol.io. For SPY, I used Yahoo adjusted prices after inception and an S&P 500 history reconstructed with Shiller dividend data beforehand. For each, there are four portfolios: ordinary buy-and-hold, continuous 2x, Goldilocks switching between 1x and 2x, and a 10-month SMA evaluated monthly, using the base asset’s total-return index and also switching between 1x and 2x leverage.

All returns are nominal, in USD, with dividends reinvested. The 2x series are daily-reset simulations throughout, with borrowing costs linked to the federal funds rate and an additional 1% annual implementation cost while levered. (Note that 1% TER is higher than what SSO and WLDU charge today.) The results include those financing and implementation costs, but exclude trading costs and taxes. Signals trade at the following session’s close. There are no contributions or withdrawals.

First, the full period. May 1977 is where the growth signal has accumulated enough history for its ten-year calculation. Maximum drawdowns below are measured daily.

May 1977–June 2026

Asset	Strategy	CAGR	Sharpe	Max DD
SPY	Buy & hold 1x	12.02%	0.522	-55.19%
SPY	Buy & hold 2x	14.86%	0.461	-88.23%
SPY	Goldilocks 1x/2x	15.86%	0.598	-57.27%
SPY	SMA 10-month 1x/2x	15.45%	0.517	-59.20%
VT	Buy & hold 1x	10.86%	0.450	-58.35%
VT	Buy & hold 2x	12.87%	0.403	-86.80%
VT	Goldilocks 1x/2x	14.01%	0.510	-59.61%
VT	SMA 10-month 1x/2x	14.19%	0.473	-67.63%

The permanent 2x portfolios beat ordinary equities on return, but the drawdowns are brutal. Goldilocks produces a higher return than either permanent 2x portfolio, with a worst loss close to ordinary equities. (Remember, we don't shift from equities to cash. We shift from levered equities to unlevered.)

Against the SMA, it's a toss-up for returns alone. But Goldilocks has the higher Sharpe and smaller drawdown in both cases.

And importantly, Goldilocks has less than half the number of trades of the SMA approach .

Full-period allocation changes	Goldilocks	SMA 10-month
SPY	28	73
VT	28	66

(Each change means selling one holding and buying the other. For VT, that’s 56 individual orders for Goldilocks versus 132 for the SMA, excluding opening and closing the test portfolio.)

Goldilocks held 2x for about 39% of trading days over the full period. The VT SMA held it for about 78%. Getting broadly comparable returns with much less time at 2x is something I find particularly interesting.

What about revised macro data?

There is an obvious objection to a macro backtest. Economic data arrives late and gets revised. Using the final series can give a historical strategy information it couldn’t have had at the time.

The long test above uses revised history. To check this, I also ran a 2001 onward test using archived Chicago Fed releases and ALFRED inflation vintages. For the more recent period, we can reconstruct the signal using archived releases rather than retrospectively revised history. That gives us greater confidence that the inputs reflect what an investor could actually have known at the time. The signal is reconstructed from what was available at each decision date, rather than the final numbers we see now. The first three months default to 1x because the usable CFNAI archive starts later; April is the first month with the required inputs.

Here is that higher-confidence period.

January 2001–June 2026

Asset	Strategy	CAGR	Sharpe	Max DD
SPY	Buy & hold 1x	8.99%	0.523	-55.19%
SPY	Buy & hold 2x	11.27%	0.446	-84.38%
SPY	Goldilocks 1x/2x	15.45%	0.696	-55.19%
SPY	SMA 10-month 1x/2x	13.15%	0.571	-59.20%
VT	Buy & hold 1x	7.97%	0.445	-58.35%
VT	Buy & hold 2x	9.10%	0.381	-86.80%
VT	Goldilocks 1x/2x	13.43%	0.592	-58.35%
VT	SMA 10-month 1x/2x	11.44%	0.494	-67.63%

 The result survives. Goldilocks beats both alternatives on returns and Sharpe for both markets. Its worst drawdown equals ordinary buy-and-hold because it was unlevered throughout the decisive peak-to-trough decline.

Here is the earlier segment on its own, using revised macro history. This helps show whether the result is merely a feature of the more recent market.

May 1977–December 2000

Asset	Strategy	CAGR	Sharpe	Max DD
SPY	Buy & hold 1x	15.38%	0.521	-32.96%
SPY	Buy & hold 2x	18.85%	0.477	-60.56%
SPY	Goldilocks 1x/2x	18.56%	0.588	-32.96%
SPY	SMA 10-month 1x/2x	17.99%	0.470	-58.95%
VT	Buy & hold 1x	14.06%	0.455	-28.62%
VT	Buy & hold 2x	17.08%	0.431	-57.37%
VT	Goldilocks 1x/2x	16.83%	0.517	-29.04%
VT	SMA 10-month 1x/2x	17.24%	0.452	-44.03%

 It's essentially all a toss-up for returns. Goldilocks has substantially smaller drawdowns than the two other strategies using leverage. It also beats the SMA on Sharpe.

Across all six asset/period comparisons, Goldilocks has the highest Sharpe of the four portfolios. It beats the SMA’s return in four of six and has the (significantly) smaller maximum drawdown in all six. (The periods overlap, of course; these aren’t six independent discoveries.)

Other Questions and Objections

The headline comparison raises three reasonable questions: whether the benefit comes simply from using less leverage, whether avoiding the 1987 crash was fortunate timing, and whether a few major bear markets explain most of the result. My thoughts on each of these below.

Does the strategy merely use less leverage?

Goldilocks was levered for about 39% of trading days over the full sample, giving it mean exposure of about 1.391x. Some reduction in drawdown is therefore to be expected simply because its average leverage was lower. A fairer test compares Goldilocks with constant leverage matched separately to its average exposure in each period.

In that comparison, Goldilocks had higher CAGR and Sharpe and smaller maximum drawdown for both assets in all three periods. The results below report Goldilocks first and matched constant exposure second in each pair.

Period	Asset	Mean exposure	CAGR	Sharpe	Max DD
1977–2000	SPY	1.267x	18.56% / 16.61%	0.588 / 0.504	−32.96% / −41.05%
1977–2000	VT	1.267x	16.83% / 15.07%	0.517 / 0.444	−29.04% / −36.29%
1977–2026	SPY	1.391x	15.86% / 13.56%	0.598 / 0.492	−57.27% / −69.58%
1977–2026	VT	1.391x	14.01% / 12.03%	0.510 / 0.425	−59.61% / −72.67%
2001–2026	SPY	1.472x	15.45% / 10.57%	0.696 / 0.480	−55.19% / −72.02%
2001–2026	VT	1.472x	13.43% / 9.01%	0.592 / 0.409	−58.35% / −75.04%

 These results suggest that the timing of leverage added value beyond simply lowering average exposure.

Was avoiding the 1987 crash just luck?

Maximum drawdown is particularly sensitive to a single extreme event. A strategy that happens to be unlevered during that event can look much safer than one that remains levered. This makes the treatment of Black Monday a meaningful robustness question. One way to stress-test it is to force the strategy to remain at 2x through the crash, even though its actual signal called for 1x, and compare the result with the original backtest.

Forcing Goldilocks to remain levered during the 1987 crash reduced full-period VT CAGR from about 14.0% to 13.5%. Quite the dent over so many years. But this also shows that the reported return advantage was not entirely dependent on avoiding that one crash.

More importantly, Goldilocks was unlevered in 1987 because inflation had risen outside its acceptable band. In other words, arguably, being out of leverage then was not luck at all. The MRO had a reason to be out of leverage that may have been a necessary condition for Black Monday. Inflation was a big part of the narrative that led up to Black Monday. Without it, Black Monday may never have happened. I don't know that for a fact but it is plausible.

The broader objection remains valid: an event resembling 1987 could occur while the macro signal still says Goldilocks. Monthly data, smoothing and a two-month confirmation rule cannot provide reliable warning of every sudden shock. The MRO is a rule for modulating leverage, not a comprehensive crash-protection mechanism. In my wider portfolio policy, diversification into assets with low or no correlation and a standing tail hedge address risks that can arise outside the macro signal’s scope.

Is the advantage concentrated in the dotcom and 2008 bears?

A rule can appear broadly successful while deriving most of its advantage from just a few episodes. To investigate this, I examined the full-period cumulative log-return advantage over constant leverage matched to Goldilocks’ average exposure. Approximately 57% of SPY’s advantage and 52% of VT’s advantage came from outside the dotcom and 2008 bear markets. After excluding October 1987 as well, approximately 44% for SPY and 43% for VT came from all other periods combined.

Those percentages concern the contribution to cumulative log-return advantage. They are not the share of drawdown reduction, and they're not the fraction of years or trades that were successful. They indicate that the return benefit extended beyond the three crash episodes, while also showing that those episodes contributed materially to the overall result.

What happens when both rules are forced to bear the same crash risk?

A further stress test forced both Goldilocks and the moving-average strategy to remain levered through 1987, the dotcom bear market and the 2008 financial crisis. This deliberately overrides the signals in their most consequential periods. It asks how the two approaches compare when their ability to reduce exposure during those episodes is removed.

My follow-up reported full-period VT figures of 19.08% CAGR, 0.728 Sharpe and −44.58% maximum drawdown for MRO, against 20.99%, 0.718 and −44.03% for SMA. On those figures, SMA had the higher return and slightly smaller drawdown, while MRO retained a slightly higher Sharpe.

My preference for MRO rests partly on its economic rationale: it adjusts leverage in response to activity and inflation, whereas the SMA responds to prices. Critically, the MRO can act where price trends have not yet weakened. I can imagine scenarios where that would happen. Because I have other risk filters in my portfolio policy as well that react, in part, to price action, the MRO adds specific value.  

Would I use it?

My conclusion is: the MRO’s Goldilocks signal is the most useful signal I have found so far for deciding when to add leverage. It produces competitive returns, a better Sharpe and less than half the allocation changes required by the SMA.

Over these tested histories, it would have avoided the most catastrophic losses associated with continuously holding 2x exposure. In the full-period tests, the worst drawdowns fell from roughly 87–88% to roughly 57–60%. I would rather face the latter.

So, yeah, I would and will use it.

How I put this together

I used LLMs for the legwork: writing the backtest code, pulling and processing the data, producing the comparisons, etc.. I supplied the original macro rule, the research questions and the portfolio definitions. The final comparisons were standardized to use the same return data, costs and execution assumptions. I read and examined everything several times over and found a few major errors that I had corrected. I also had alternate LLMs double-check. Still, it's possible that there are mistakes in the backtests. I wrote the article myself! This is not "AI-generated". If you see em dashes, that's because Word converted them from regular dashes!

If there is interest, I can upload the data sets and code to GitLab so that anyone can examine and replicate them.

*****

I will be publishing and tracking my DIY hedge fund experiment at www.hedgefol.io.