Funding Momentum, Tested Both Ways: Why We Didn't Just Flip the Sign

In the first article of this series, a coin list invented a strategy. This one is about a quieter trap: deciding which way a signal works only after you have seen the data. Few mistakes in backtesting are easier to make, because the data seems to hand you the answer.

The idea: watch funding move, not its level

Most funding-based signals look at the level: is funding high or low right now? Funding momentum looks at the change. For each coin, we took the average funding rate over the last 30 days and subtracted the same 30-day average as it stood one week earlier. A positive number means funding is getting more expensive for longs; a negative number means it is getting cheaper.

Which way should that predict prices? There are two stories, and both sound right:

That is the first warning sign. When both directions come with a convincing story, the story is not evidence. Only a test can decide, and only if the direction is fixed before the test runs.

Three tests, in the order they happened

1. Exploratory check · not written down in advance
   Contrarian, 10 large coins
   A quick look on ten large perpetuals, three coins long and three short. It looked promising: 13.2% a year, Sharpe 0.57, placebo p = 0.024. But the whole result came from the second half: 2022–2023 lost money (Sharpe −0.80), 2024–2026 made it back and more (Sharpe 1.43).
   Sharpe 0.57 · p 0.024 · halves −0.80 / 1.43
2. Registered test 1 · written down before the run
   Continuation, 41 liquid coins
   A formal test on a broader universe with fixed pass criteria. The hypothesis we wrote down was continuation, the opposite of the exploratory hint: with both stories sounding plausible, we picked one and committed to it. It failed clearly.
   Sharpe −0.92 · p 0.79
3. Check · no hypothesis, no verdict
   Is the change just the level in disguise?
   Before testing the mirror image, we measured how much it would overlap with plain funding-level signals. More on that below.
4. Registered test 2 · written down before the run
   Contrarian, same 41 coins
   The mirror image of test 1, registered as a new hypothesis with its own pass criteria, and counted as one more test. It also failed.
   Sharpe 0.21 · p 0.056

Test 1: continuation fails clearly

41 liquid coins, Jan 2022 – Aug 2026, after costs and funding.

Negative in both halves (−0.22 and −1.68), negative or flat in all four walk-forward windows, worse at double costs (−1.26). In the placebo test, where we assigned coins to the long and short side at random 10,000 times, 79% of the random portfolios did at least as well. The registered direction wasn't just unhelpful; it pointed the wrong way.

The tempting move: just flip the sign

A Sharpe of −0.92 looks like an invitation. If going long rising funding loses this reliably, surely going short it wins? Two things stand in the way.

1. Choosing the direction after the fact is a second test

If you test a signal, look at the result and then keep whichever direction worked, you have run two tests and kept the better one. The reported p-value then pretends only one test happened. Statisticians call this the garden of forking paths: every choice made after seeing the data (direction, lookback, coin list, threshold) quietly multiplies the number of tests, and with enough of them something always looks significant.

So we did not flip the sign. We wrote the contrarian version down as a new hypothesis, fixed its pass criteria before running it, and counted it as one more test against our overall error budget.

2. Costs break the symmetry

The two directions are exact mirror images only before trading costs. Both directions pay costs on every rebalance, so flipping a losing strategy does not turn its loss into a gain of the same size:

Test 2: the contrarian mirror fails too

Against the criteria written down before the run, it fails every single one:

"Almost significant" isn't

A p-value of 0.056 looks close to the textbook 0.05. It isn't close to anything that matters here. By the time this test ran, we had logged 33 hypothesis tests in this research programme, not counting about 120 earlier exploratory ideas. If you run 33 tests on pure noise, the chance that at least one of them clears p = 0.05 by luck is over 80%. To keep the overall false-alarm rate at 5%, each single test has to clear roughly 0.05 / 33 ≈ 0.0015. The contrarian result missed that bar by a factor of about 37. Counting the earlier exploratory ideas too, the bar would be closer to 0.0003.

The same pattern, three times

The exploratory check lost money in 2022–2023 and made it in 2024–2026. The registered contrarian test did the same: −0.51 in the first half, 1.00 in the second. Split into four windows, most of its return comes from a single stretch:

Here are all three tests side by side:

* Exploratory check: 1,000 placebo runs instead of 10,000; halves split at January 2024 instead of July 2024.

Is the change just the level in disguise?

Before running the second test, we checked how much "funding is falling fastest" overlaps with "funding is lowest", the classic level-based signal. On average, the two measures are only loosely related (correlation 0.24). But the overlap is lopsided: the long side of the contrarian test shared about 40% of its coins with a plain "lowest funding" basket, twice what chance would give; the short side shared 21%, about chance level.

So a part of the contrarian long leg was buying the same coins as a funding-level strategy. That matters, because in the survivorship article exactly that leg, the coins with the most negative funding, turned out to be the weak spot once the coin universe was chosen honestly.

What we take away

1. When both directions have a good story, the story is not evidence. Fix the direction before the test runs, or the data will pick it for you.
2. Don't flip the sign of a failed signal. If the mirror image is worth testing, register it as a new hypothesis, fix its criteria in advance and count it as one more test.
3. Costs break the symmetry. Reversing a strategy with a Sharpe of −0.92 does not give +0.92. Both directions pay to trade.
4. Judge a p-value against the number of tests. After 33 logged tests, 0.056 is not "almost significant"; the bar is closer to 0.0015.
5. A sign change between halves is a red flag. A full-period number that averages a losing regime and a winning one describes neither of them.

What this means when you evaluate any signal

When someone shows you a backtest of "long X, short Y", ask whether the direction was fixed before the data was looked at, whether the opposite direction was also tried, and how many variants were run before this one was shown. A single good-looking result is not informative without that context.

FundingArb doesn't try to predict where funding goes next or which way prices move. It collects funding with a hedged position, and only enters when projected funding clears fees by a margin.

Methodology