Everything below is a backtest on survivorship-inclusive data, net of trading costs, over one fixed window. A backtest measures the past under a set of assumptions; it is not a forecast and not a promise. Live results are lower than backtests, always. Nothing here is investment advice. Decisions are yours.
We did the unglamorous thing. Before running a single test we fixed the judge in public, then fed it every idea we and our readers could think of. Most of them died. This is the record of what died, and the one plain thing that lived, because a research process you only hear about when it produces a winner is not a research process. It is a highlight reel.
The judge came first
The order matters. We wrote down the validation battery before we had a candidate to run through it: held-out walk-forward windows so nothing is judged on the data it was tuned on, a probability-of-backtest-overfitting test to catch a lucky variant with hidden siblings, and a deflated Sharpe that raises the bar for every extra thing you try. The point of fixing the goalposts first is that no later version of us can move them. Every idea below faced the same fixed judge on the same graveyard-inclusive universe, so the comparisons are honest even when the results are unflattering.
What died
The list is longer than the survivor list, which is the normal and healthy ratio. Social attention, measured as growth in coins' mention volume, looked predictive in one window and dissolved out of sample; it moves with price, not ahead of it. Rotating into the broad altcoin universe instead of the majors was worse than useless on honest data: chasing the hottest liquid alts lost money once the coins that later died were left in the sample, a textbook survivorship trap. Fine-tuning the momentum lookback, the basket size, and the rebalance clock moved the result around inside the noise and never survived deflation. Swapping the trend gate for cleverer variants, breadth dials, longer moving averages, changed nothing an honest test could keep. Volatility-targeted and inverse-volatility position sizing cut drawdown but sacrificed more return than it saved. An MA trailing exit did the same, worse: it whipsawed the book out of the exact runs that make trend-following pay. And buying weakness, entering coins below their long moving averages, the classic mean-reversion move, was comprehensively beaten by simply buying strength. Each of those is a real edge in some market. None of them is an edge in trend-gated crypto majors.
The one that lived
Exactly one addition improved a trend-gated momentum book and held up out of sample, and it is not a return signal. It is a risk filter built from funding rates. The rule is almost embarrassingly simple: among the momentum leaders you would otherwise hold, skip the ones whose perpetual funding is the most extremely crowded on the long side, and hold the rest. Crucially, funding alone predicts nothing directional; coins with high funding are not systematically about to fall. But the coins carrying the most crowded long leverage are the ones that, when they do fall, fall in a liquidation cascade. Sidestep them and the strategy's fat left tail collapses. In the test the ordinary return numbers barely moved, but the kurtosis of the daily returns dropped roughly fivefold and the worst drawdowns shrank. It is a seatbelt, not an engine, and it survived precisely because it never touches the momentum logic that does the earning.
Two details made us trust it rather than merely like it. First, it is not a knife-edge: every crowding threshold we tried improved the book, so we are not balanced on one lucky number, which is the failure mode that killed the attention idea. Second, the funding data is current, not a historical curiosity, so unlike some of the other candidates this one can actually run forward and be measured in public rather than only admired in a backtest.
The numbers, with the asterisks attached
For scale: over the validation window the trend-gated majors momentum strategy, with the funding seatbelt on, compounded far faster than holding the largest coins would have, at roughly half the peak-to-trough drawdown, and it beat a simple bitcoin buy-and-hold too. The gate is what halves the drawdown, because it sits in cash through bitcoin downtrends rather than riding them down; the momentum ranking is what earns; the funding filter is what trims the tail. We are deliberately not printing a headline multiple as a selling line, because a backtest multiple is the number most often abused. What we will commit to is the shape: a real but ordinary momentum edge, made less dangerous by one risk filter, and certain to look tamer live than on paper. The honest audit of why even our best number is an upper bound lives in the backtest audit, and it applies here in full.
What a real search looks like
The reason to publish the losers is that they are the evidence the judge is real. Anyone can show you the one idea that worked. A process shows you the dozen that did not, run through the same fixed gate, so you can see that the survivor earned its place rather than being the one variant that got a press release. The result is undramatic on purpose: after all that searching, the market handed back a plain trend-follow of the majors plus a funding seatbelt, and told us in a dozen different ways to stop adding cleverness. That is usually what an honest search returns. The edges that survive scrutiny are boring, and the exciting ones are the ones that do not. Decisions are yours.