MACD across market regimes: was the benchmark the bias?

Last time we tested the MACD crossover on the S&P 500 and it lost. But the S&P is a market that essentially only goes up — the one place a defensive rule can never win. So we ran the identical rule across regimes: the S&P split by trend, the Nikkei's lost decades, and a driftless currency. The verdict flips with the market — and testing only on markets that went up turns out to be a survivorship bias of its own.

Start here. This builds directly on our MACD teardown, which found the 12/26/9 crossover loses to buy-and-hold on the S&P 500. This piece asks the harder question that result raised: was that MACD's fault, or the market we tested it in?

1 / You can't out-defend a market that only goes up

Here's the objection I kept coming back to about my own MACD test. In a market that rises over time, being always 100% invested is the ceiling for any long-or-cash strategy — every day you spend in cash is a day of guaranteed underperformance. A signal like MACD-timing, which sits in cash roughly half the time, is structurally doomed to lag there. Judging a defensive tool in a relentless bull market is like judging an umbrella on a sunny day.

And it goes deeper. Reaching for the S&P 500 as "the benchmark" is itself a choice — of the single most successful equity market in modern history. That's survivorship bias one level up from the usual kind: other markets stagnated for decades, and some (Russia 1917, China 1949) went to zero for investors. The classic study of this is Dimson, Marsh & Staunton's Triumph of the Optimists. "Stocks always recover" is a statement about the markets that did. So let's stop testing MACD only where the answer is rigged, and separate the regimes.

2 / Split the same S&P test by regime

First, decompose the exact SPY test (1993–2026) by trend regime: tag every day by whether the S&P was above or below its own 200-day moving average (a standard, causal bull/bear filter), then measure both strategies within each regime.

Two things. First, the deficit is concentrated in the bull market: MACD gives up nine points a year there, but only four when the S&P is below its average — exactly as the "umbrella on a sunny day" logic predicts. But second, the surprise: buy-and-hold is still positive and still ahead even in the "bear" bucket (+9% vs +5%). How? Because "below the 200-day average" is not the same as "falling." That bucket is where the violent first legs of the 2009 and 2020 recoveries happen — huge up-days that a lagging, defensive rule sits out. On US stocks the upward pull is so strong it beats defense even inside its own downturns. Which tells you the only way to see a defensive rule shine is to leave the market that keeps rescuing it.

3 / The same rule on a market that stayed down

So take the identical MACD-timing rule — no re-tuning — to the one major market that spent a generation going nowhere: Japan's Nikkei 225, from its 1989 bubble peak onward.

The rule that lost on the S&P now looks completely different. Through the great decline of 1990–2012, buy-and-hold bled −5.6% a year; MACD-timing, sitting out the persistent downtrends, lost only −1.1% — a positive Sharpe against a negative one — and it cut the maximum drawdown from a portfolio-ending −82% to −60%. Over the full window buy-and-hold edges back ahead on raw return (1.2% vs 0.8%), but only because it caught the post-2013 recovery. The defensive tool did exactly what defense is for — in the market where defense was actually needed. Nothing about MACD changed. The regime did.

4 / Strip out the drift entirely

One more test isolates the point. A currency pair like EURUSD has no structural upward drift — buy-and-hold returned roughly 0% a year over 2003–2026, just oscillating. And because there's no up-trend to be long-biased toward, we can let MACD trade the way trend-followers actually do: long and short.

With the drift gone, so is the performance: MACD long/short lost −5.4% a year (Sharpe −0.43) after a trivial 2 bps in costs — death by a thousand whipsaws. This is the quiet truth behind the S&P result. MACD wasn't extracting a trend edge there; it was borrowing the market's upward drift and handing some of it back in lag and costs. Take away the drift and there's nothing underneath but the friction.

5 / Every indicator is a bet on a regime

If MACD is a bet on trends, its mirror image is a bet on ranges. Run our RSI(2) mean-reversion signal on the same S&P, split by the same regimes:

It's a clean inversion. When the market trends up, buying dips works because the dips get bought back. When it falls or chops, dip-buying catches knives and trend-following — staying with the move, or out of the way — pulls ahead. There is no context-free question "does MACD work?" or "does RSI(2) work?" There is only: which regime rewards this bet, and are you in that regime now? — a question a backtest of the past can never answer about the future.

Test your own strategy like the market has more than one mood