Can you actually avoid crashes?

Every investor who has lived through a −50% year has had the same thought: surely there is a signal that would have gotten me out. The market doesn't crash in a day; 2008 took over a year from peak to trough. Why ride the whole thing down?

Trend following is the respectable, systematic version of that thought, with a century of academic literature behind it. We tested it the way we test everything — out-of-sample, across eras, with costs — because the dream deserves a real answer, not a vibe.

The classic implementation is almost embarrassingly simple: compare the index to a slow signal such as its 10- or 12-month moving average. Price above the line — stay invested. Price below — move to cash. Check once a month, no discretion, no headlines.

The slowness is deliberate. A slow signal ignores ordinary turbulence and only reacts to sustained declines — which is exactly what the big crashes are. By the time a 2008 is well underway, price is far below any 12-month average, and the rule has long since stepped aside.

On roughly a century of data, the result is consistent: the rule reliably softens the very worst drawdowns. The catastrophic peak-to-trough collapses — the kind that take portfolios down 50% or more — get meaningfully truncated, because the rule exits partway down and waits out the rest.

This part of the promise is real, and it is worth being precise about, because almost everything else sold under 'crash protection' is not. The drawdown chart on our what-works page shows it plainly.

Here is the part the backtest screenshots never show. Most signals that drop the index below its average are not 2008 — they are ordinary corrections that recover within months. The rule exits on those too, sells low, watches the rebound from cash, and buys back in higher. Each false alarm costs a slice of return.

Summed over a century, the whipsaw costs roughly cancel the crash savings. Trend following converts the journey — shallower valleys, more missed rallies — without improving the destination. Our tests found no robust after-cost return advantage over simply holding, and that matches the academic consensus.

So the honest framing is: trend following is insurance, not alpha. You pay premiums (whipsaws) for a smaller maximum loss. Insurance can be rational. It is not free money.

NISA adds a quota mechanic that backtests on US data never see: when you sell, your tax-free quota does not come back until the following year. A trend exit in March cannot fully re-enter tax-free in October when the signal turns back up.

That friction lands on exactly the rule's weakest spot — the whipsaw, where it sells and needs to re-buy quickly. Some of your re-entry gets pushed into taxable space or delayed, and either way the cost of every false alarm goes up.

This is why our verdict keeps trend following out of the recommended NISA default: the structure of the account itself penalises the strategy's most frequent move.

If a −50% drawdown would make you sell everything at the bottom, a rule that caps your worst case nearer −25% — at the cost of some long-run return — might genuinely keep you invested. The best strategy on paper is worthless if you abandon it at the worst moment; we wrote about the same trade-off in the lump-sum article.

That is a statement about your psychology, not about the market. In this project the overlay runs as a paper-trading crash hedge — monitored, published, and explicitly not a recommendation.

Notice what even the working version of the dream is not: prediction. The rule never knows a crash is coming — it reacts after declines have already begun and pays for every false start. Nobody reliably sells the top; the data is unkind to everyone who claims otherwise.

Preparation, unlike prediction, is boring and works: hold an index broad enough to survive anything, keep cash you'll need within a few years out of equities entirely, and size your exposure so the historical worst case is survivable for you. Then the crash becomes something you hold through — which is, on every century-long dataset we have, the move that actually got paid.