Does Tactical Asset Allocation (Sector Rotation) Actually Work? I Tested 100 Years of Data to Find Out

I came across a tactical asset allocation strategy on Substack claiming a 16% compound annual growth rate — beating the MSCI World Index in 25 of the last 26 years.

I was intrigued, so I rebuilt it and tested it myself. Here's what I found…

The mechanics were disclosed: rotate monthly into the best factor and sector ETFs by momentum, with a trend filter for bear markets. Fully explainable means fully testable. So I rebuilt it against the Ken French data library — the same hundred-year dataset academics use.

Real result: +0.75% a year better than the market, not +12% as claimed! It beat the index in 11 years out of 27, not 25 out of 26 — because a momentum strategy that wins almost every year doesn't exist.

They win big occasionally and lose often. A strategy beating the market 96% of the time would be the most consistent long-only track record ever produced. It wouldn't sell for €10 a month.

That sent me looking for what actually works: since 1926, a sector has hit its ten-year low relative to the market 283 times, through every crisis this century produced. Then, after that, what happened next?

I spent this month finding out, with the most systematic test I could construct. The results changed how I invest. Three of them will probably annoy you (I’m not above a bit of clickbait).

My starting hypothesis was the classic contrarian’s: buy sectors at generational lows relative to the market, wait for mean reversion. Templeton’s “point of maximum pessimism,” implemented with sector ETFs. I’ve run positions on exactly this logic myself.

To test it properly you need more than ETF history — most sector ETFs are younger than a single market cycle. So the testbed was the Ken French industry portfolios: 30 US industry groups, monthly and daily total returns, July 1926 to May 2026. Every test that follows uses out-of-favour defined relatively: an industry’s cumulative return ratio against the total market, hitting its lowest level in ten years. Call that a washout — the sector has never been more out of favour in a decade. A century of data supplies 283 of them, across every industry and every regime: depression, war, inflation, disinflation, bubbles and their funerals.

Then you ask the only question that matters: what happened next?

Buying the washout the moment it prints — maximum pessimism, the purest form of the contrarian trade — lost 1.4–3% a year against the market over the following twelve months, with a 38–41% hit rate. Not occasionally. On average, across 283 episodes and a hundred years.

The contrarian instinct isn’t wrong about the destination; it’s wrong about the schedule. The median washout kept underperforming for roughly ten further months before its true bottom. And depth was no defence: the most crushed tercile of washouts performed no better than the mildly crushed. Cheapness alone carries no information about when the bleeding stops.

Extend the horizon and the picture merely flattens: by five years the raw washout-buyer roughly matches the market. All that pain, no premium. If you’ve ever bought a hated sector and watched it get more hated, you weren’t unlucky — you were on schedule.

Now change one thing. Take the identical 283 washouts, but don’t buy the low. Wait — sometimes many months — until two conditions hold: the sector’s own 12-month return has turned positive, and it has beaten the market over the trailing six months. In other words, the recovery has visibly begun on both an absolute and a relative basis.

Same sectors. Same crises. Different entry. The results invert: +3.7% a year over the market in the first year after entry, a 57–60% hit rate, and the edge persists at +1.7% annualized over five years. It survives de-clustering (it isn’t one lucky crisis), and it worked in every 20-year era except the 1980s–90s — the one period when structurally dying industries (coal, textiles) stayed dead through a disinflation mega-bull.

The two filters are worthless in isolation — the absolute filter alone earns roughly nothing, the relative filter alone a fraction of the combination. This is the retail-scale echo of what AQR found for market timing (“Sin a Little,” for those keeping score at home): value and momentum signals are individually weak and negatively correlated, so the blend is worth more than the sum. Valuation tells you what to buy. Momentum tells you when.

Assembled into a portfolio — six equal positions maximum, spare capital in the index, entries only on confirmation — the full system beat the market by 1.4–2% a year across ninety years, with 59–74% of trades profitable. That is, I believe, close to the honest ceiling for this entire category. It’s also, compounded over twenty years, roughly 50% more terminal wealth. Modest and meaningful aren’t opposites.

Here’s the one that stung. If you enter because a sector is cheap, symmetry suggests you exit when it’s no longer cheap — sell when the ratio recovers to its long-run fair level. I tested exactly that.

The fair-value exit produced a 92% win rate and negative alpha.

Sit with that pair of numbers, because the seduction is the win rate and the truth is the alpha. Selling at “fair value” banks a small, certain profit on every recovery — and hands back the entire right tail, because recoveries don’t stop at fair. The same behavioural under-reaction that created the washout drives the recovery straight through fair value and beyond; roughly half the strategy’s total profit was earned after the ratio passed its long-term median. Meanwhile the positions that never recover — the value traps — sit in the book indefinitely, waiting for a reversion that isn’t coming. High win rate, cut winners, ride losers: the retail investor’s P&L in one exit rule.

The cleanest evidence was surgical: I took the best-performing exit (a trailing stop on the relative ratio — sell only when the recovery retraces 15% from its peak) and merely added a fair-value take-profit on top. That single addition collapsed the excess return from +2.0% to +0.2% a year. Nothing else changed.

So the system ends up asymmetric in a way that offends the value investor’s aesthetic: enter on cheapness-plus-turn, exit on momentum-break — never at a price target, and never on a calendar. (I tested calendar exits too. Fixed 36-month holds were the worst rule of all, negative since 1970. And for completeness: on daily data, the strategy performs identically whether you trade at month-ends or any day the signal fires — but binary regime switches need slow evaluation, because at daily frequency threshold rules whipsaw themselves to death, tripling turnover and halving the edge.)

All of this collides with the most crowded debate in markets. The S&P 500 is historically top-heavy; equal-weight has rarely been this beaten-down against cap-weight (the ratio printed a fresh ten-year low within the past year). A lot of smart money is pre-positioned for the great broadening — and it keeps not coming.

The century of data has two precedents for today’s concentration: the Nifty Fifty and the dot-com bubble. Both times, “the index is too top-heavy” was correct. And both times, the investor who waited for momentum confirmation captured more of the unwind than the investor who pre-positioned on conviction.

Dot-com, concretely. The equal-weight/cap-weight ratio bottomed in February 2000; a simple 12-month relative momentum signal confirmed the turn that October. The conviction investor who rotated to equal-weight in January 1998 — right thesis, two years early — earned +42 points of cumulative outperformance through 2005, after bleeding 19 points waiting to be right. The investor who waited for the October 2000 confirmation earned +50 points. Waiting beat conviction, because concentration unwinds are slow: 2000–2005 paid out over six years. An event that unfolds over years does not need to be pre-positioned; you can board it after it visibly starts and keep nearly all of the payoff. What can’t be recovered is the bleed of being early — and “too top-heavy” was true in 1995, then cost 8 to 14 points a year for five years before paying.

That’s the discipline the data suggests for the AI-concentration trade: have the thesis, define the trigger, and let the trigger — not the conviction — move the money. As I write, that particular signal sits within a rounding error of confirming. Which is precisely why one shouldn’t jump it.

I went looking for a system that would deliver the alpha the marketing promises. What I found instead was a ceiling — about 1–2% a year at the index level, and a century of evidence that most of what’s sold above that number is backtest archaeology — plus something more useful than the mythical system: a sequencing discipline that converts a good instinct from a cost into an edge. The contrarian eye still does the bulk of the work: it finds what’s washed out. But cheapness is a watchlist, not a trade. The turn is the trade. And the market, not a price target, ends it.

This is research, not investment advice; I run these rules with my own money and my own risk tolerance. Data: Kenneth French Data Library, 30 industry portfolios, monthly and daily total returns 1926–2026; costs modelled at 0.2% per side. Full methodology available on request.