The point of running a strategy you expect to lose
There’s a quiet rule in quantitative investing that doesn’t get said often enough: the fastest way to find alpha is to ruthlessly catalog what isn’t alpha.
Alpha — return above what a passive benchmark would have handed you for free — is rare almost by definition. Most strategies that look clever are, after fees and after honest benchmarking, just beta in a costume. So before you commit capital, the question worth answering isn’t “did this make money?” It’s “did this beat the dumbest possible alternative — owning the index and doing nothing?”
We ran exactly that test on a classic, well-respected tactical strategy: Global Equities Momentum (GEM), the dual-momentum system popularized by Gary Antonacci. The benchmark was the most boring thing imaginable — buy SPY, hold it, go to sleep.
The verdict, stated plainly up front: GEM did not produce alpha in this window. It lost to buy-and-hold. And that result is more useful than a win would have been.
The whole case in one glance — we’ll walk through each panel below. (A dark-theme version is also available.)
What GEM is supposed to do
Dual momentum combines two ideas that each have decades of academic support:
Relative momentum — among risky assets, hold whichever has been strongest. GEM picks between US equities (SPY) and international ex-US equities (VEU).
Absolute momentum — only hold any equity if it’s also beating the risk-free rate (T-bills, BIL). Otherwise, hide in bonds (BND).
Every month, GEM looks back 12 months, ranks the candidates, and rotates into the winner. The promise is seductive: ride strength when markets trend, and step aside into bonds before the worst of a crash. Capture the upside, dodge the drawdown.
That’s the theory. Here’s what the tape said.
The scorecard
Universe: SPY (US equity), VEU (intl ex-US), BND (US aggregate bonds), BIL (T-bills as the risk-free floor) Window: 2023-01-31 → 2025-12-31 Signal: canonical single-lookback Antonacci GEM, 12 months, monthly rebalance Data: yfinance Adjusted Close (dividend- and split-adjusted)
Metric | GEM | SPY (buy & hold) | Δ (GEM − SPY) |
|---|---|---|---|
CAGR | 15.28% | 21.35% | −6.07pp |
Annual volatility | 11.07% | 11.79% | −0.72pp |
Sharpe (rf=0) | 1.345 | 1.710 | −0.365 |
Max drawdown | −8.33% | −8.33% | +0.00pp |
MAR (CAGR / |MDD|) | 1.836 | 2.565 | −0.729 |
Total return: GEM grew $1 → $1.51 (+51.4%). SPY grew $1 → $1.76 (+75.8%).
Two dollars, three years. Both lines climb — this was a good market for being long equities — but the grey line (do-nothing SPY) pulls steadily ahead of the blue (GEM). The gap doesn’t come from one bad month; it opens early and never closes. By the end, the “smart” tactical system left a quarter of the index’s gain on the table.
Read that table slowly, because every row is telling a piece of the same story.
GEM gave up 6 percentage points of CAGR per year and a third of a point of Sharpe. It shaved a sliver off volatility (11.07% vs 11.79%) — but that’s a rounding-error of “smoother” in exchange for a fortune in foregone return. And the line that should haunt any tactical-allocation believer: the max drawdowns were identical, to the basis point. −8.33% for both.
The entire pitch of dual momentum is downside protection. In this window, it protected nothing the index didn’t already avoid on its own — and charged 6 points of CAGR for the non-service.
The same story, drawn as a deficit. This plots GEM’s value relative to SPY — how far ahead (above zero) or behind (below) the tactical strategy ran at each point. It drops below zero in the first few months and stays there for the entire window, bottoming at roughly −16% in late 2025 before closing the period −13.9% behind. There is no stretch where GEM is winning. A strategy with genuine alpha would spend meaningful time above the line; this one never gets there.
Why it lost: the strategy was right about the trend and still couldn’t win
Here’s the twist that makes this a useful failure rather than a boring one. GEM didn’t lose because it made dumb calls. Look at where it actually spent its time:
Holding | Months held | % of window |
|---|---|---|
SPY | 23 | 63.9% |
VEU | 8 | 22.2% |
BND | 5 | 13.9% |
For nearly two-thirds of the window, GEM was holding SPY — the winning asset. It correctly identified that US equities were the place to be. So why the 6-point gap?
GEM’s black line tracks the grey SPY reference, with the background shaded by what the signal was pointing at each month: blue = US equities, amber = international (VEU), green = bonds (BND). Notice the green band at the very start — GEM hid in bonds through early 2023 while SPY was already recovering, which is where the first chunk of the deficit was minted. Then watch the amber stripes in 2025: every time the signal flips to VEU, GEM is stepping out of the asset (SPY) that kept grinding
Two structural drags, both baked into the strategy’s DNA:
1. The one-month lag. GEM reads the signal at month-end and establishes the position at the next month-end. In a market that’s grinding higher, that lag means you’re perpetually one step behind the move. You buy strength after some of it has already happened, and you exit after the turn. In a relentless bull, lag is pure tax.
2. The detours into bonds and international. GEM spent 14% of the window in BND and 22% in VEU. Every month parked in bonds while SPY climbed, and every month in international equities while US led, was a month of deliberately under-owning the best asset. The strategy did this on purpose — that’s the whole mechanism — but in a window where US equities simply never broke down, the insurance never paid out. You bought flood coverage during a drought.
The regime-transition log shows GEM flip-flopping between SPY and VEU repeatedly in 2025 (SPY → VEU → SPY → VEU across consecutive months as their 12-month returns ran neck-and-neck). Each flip is a transaction, a tax event, and a fresh dose of lag — churning between two assets that were barely distinguishable. That’s not signal. That’s noise being dignified with a trade ticket.
This is the engine room. Each line is one candidate’s trailing 12-month return — the actual number GEM ranks on. Bonds (green) and T-bills (dotted) sit low and rarely threaten, so the real contest is blue (US) vs. amber (intl). For most of the window US leads cleanly, but look at the pink bands: those mark the months where US and international were within 2 percentage points of each other — a statistical coin-flip. GEM is forced to “decide” a winner anyway, and that’s exactly when it churns. The strategy is making confident-looking rotations on differences too small to be real edge.
“No alpha” ≠ “no value”
It would be easy — and wrong — to file GEM under “debunked” and move on. Three reasons this result is a map toward alpha rather than a dead end:
The test was deliberately hostile to GEM. 2023–2025 was a near-uninterrupted equity bull market with one shallow drawdown. That’s the worst possible regime for a strategy whose edge is crash avoidance. GEM’s value proposition is asymmetric: it’s supposed to lose a little during calm bull runs and win big when the market falls apart. This window contained no apart-falling. We tested a life raft in a swimming pool and noted it didn’t help you swim faster. The honest conclusion isn’t “GEM is bad” — it’s “GEM has no edge in trending bull markets, and this window can’t speak to its behavior in a real crash.”
The drag sources are addressable. The two things that killed it — execution lag and over-trading between near-tied assets — aren’t laws of physics. A faster signal cadence, a momentum threshold (only rotate when the leader wins by a meaningful margin, not by a hair), or trading at the open after the signal instead of a full month later, each directly attacks a named loss. Those are testable hypotheses, not hand-waving.
It clarifies what alpha would have to look like. If a tactical overlay can’t beat buy-and-hold even while holding the right asset 64% of the time, then the alpha isn’t in asset selection — it’s in timing and friction. That’s a sharper question than we started with. We now know which knob to turn.
The honest caveats (read these before quoting any number)
No backtest is the truth; it’s a hypothesis dressed in hindsight. This one carries the usual load:
Monthly-only signal. Intra-month drawdowns are invisible to a month-end signal. The −8.33% figure is a month-end snapshot, not the worst tick.
No transaction costs. Real rotations pay spread and commission. Small for liquid ETFs, but nonzero — and GEM trades more than buy-and-hold by construction.
No taxes. Every regime change held under a year is a short-term capital gain. In a taxable account, GEM’s churn is expensive in a way SPY’s do-nothing approach simply isn’t. The after-tax gap would be wider than the 6-point pre-tax gap.
Fill assumption. Signal at month-end close, fill at the next month-end close — a one-month-lagged approximation of a real implementation that would trade at the next open.
Sharpe with rf=0. Standard for backtest comparison; the real risk-free-adjusted Sharpe would be modestly lower for both.
Every one of those caveats makes GEM look worse, not better. That’s the right direction for a skeptic’s bias.
The takeaway
We set out to answer one question: is dual momentum alpha? Against a buy-and-hold S&P 500, over 2023–2025, the answer is a clean no. Lower CAGR, lower Sharpe, identical drawdown, more complexity, more tax drag. By every metric that matters, the boring index won.
But “not alpha” is not a wasted result — it’s the normal result, and it’s the one that does the work. Each strategy you can confidently cross off the list narrows the search space for the one that isn’t beta in disguise. GEM earned its place on that list, and in doing so it pointed at exactly where the next experiment should look: not at what to hold, but at when to move and how much it costs to move.
The most expensive mistake in this field isn’t running a strategy that loses. It’s not testing it — and quietly believing it would have won.
Methodology note: single-lookback canonical Antonacci GEM, 12-month momentum, monthly rebalance, yfinance adjusted-close data. Backtest run 019e713f. Charts generated from curves.csv and regime_log.csv via make_charts.py. Past performance is not indicative of future results; this is research, not investment advice.