Since late September 2026 we have put roughly 420 trading hypotheses on crypto perpetuals through the same testing pipeline. Almost all of them fail, and most failures are boring. This one is not. It wasn't a weak idea that failed but our strongest one at the time, and the cause sits in a step that nearly every backtest takes without a second thought: choosing the coins.
Why this is on a funding-arbitrage site. The strategy below bets on price direction using funding rates as a signal. That is a different mechanism from collecting funding with a delta-neutral hedge, which is what FundingArb does. We publish failed research because the lesson applies to any backtest you are shown, including ones for funding strategies.
The strategy
The idea comes from research on crowding in perpetual futures. The funding rate shows which side of a market is crowded. When it is strongly positive, longs pay a lot to hold their bet; when it is strongly negative, shorts pay.
- Every week, rank all coins in the universe by their average funding rate over the past few weeks.
- Long the fifth with the lowest funding (the crowd is short, so a rebound is the bet).
- Short the fifth with the highest funding (the crowd is long and paying for it).
- Market-neutral, volatility-scaled, trading costs and funding payments included.
Tested on 50 liquid Binance perpetuals from January 2022 to August 2026:
| Universe | APR | Max drawdown | Sharpe |
|---|---|---|---|
| Fixed list of 50 coins | 27.0% | 13.9% | 1.55 |
It held up for four and a half years, through the Luna collapse, the FTX bankruptcy and several flash crashes. It was positive in both halves of the period and robust to doubled costs. It looked like something real.
The problem: where did the list come from?
The 50 coins were the highest-volume perps that are still trading today and were already listed before October 2021. That sounds harmless and is common practice. But it contains two pieces of information that nobody had in January 2022:
- Which coins would survive until 2026. Delisted coins are missing by construction.
- Which coins would be big in 2026. Today's top-volume coins are the winners of the last altcoin cycle.
So the backtest trades a group of coins that can only be identified in hindsight.
Hindsight list
- Start from today's exchange listing
- Keep coins still trading in 2026 and listed before Oct 2021
- Rank by 2026 volume, keep the top 50
- Run the strategy on them from 2022 onwards
Point-in-time universe
- Start from all 895 perps ever listed, delisted ones included
- Each week, keep coins at least 220 days old
- Rank by trailing 30-day volume, keep the top 50
- Trade only that week's universe
The test, written down before the run
Before seeing any result, we wrote the test down and committed it with a timestamp:
- Universe: the point-in-time top 50 described above.
- Engine, costs, period: identical to the original backtest.
- Pass criteria: placebo test significant after correction for the number of tests run so far, positive Sharpe in both halves and in at least three of four walk-forward windows, positive with any single coin removed, positive at double costs, and at least 12% APR.
Our expectation, also written down beforehand: "The survivorship bias is large enough to measurably reduce APR and Sharpe, but not large enough to destroy them."
The result
| Universe | APR | Max drawdown | Sharpe |
|---|---|---|---|
| Hindsight list | 27.0% | 13.9% | 1.55 |
| Point-in-time, incl. delisted | −12.5% | 61.8% | −0.42 |
Not a haircut but a change of sign. Our expectation was wrong.
The dashed line removes every delisted coin from the point-in-time universe. It barely moves: dead coins are not what breaks the strategy.
Every check agrees
Placebo. On every rebalancing day we shuffled the funding values randomly among the coins in the universe, 10,000 times. Half of these random portfolios did at least as well as the real rule (p = 0.50). The signal is indistinguishable from a coin toss.
For scale: the hindsight list's 1.55 would sit far off the right edge of this chart. It is not drawn on it because this null distribution belongs to the point-in-time universe only.
Stability over time. The hindsight list is positive in every window; the point-in-time universe is negative in three of four.
| Window | Hindsight list | Point-in-time |
|---|---|---|
| First half (Jan 2022 – Jun 2024) | 1.64 | −0.23 |
| Second half (Jul 2024 – Aug 2026) | 1.46 | −0.59 |
| Walk-forward 1 (Jan 2022 – Mar 2023) | 1.49 | −0.82 |
| Walk-forward 2 (Mar 2023 – May 2024) | 1.82 | 0.48 |
| Walk-forward 3 (May 2024 – Jul 2025) | 1.93 | −0.78 |
| Walk-forward 4 (Jul 2025 – Aug 2026) | 0.93 | −0.58 |
| Full period, double costs | 1.38 | −0.55 |
Single coins. Whichever of the 332 coins involved we remove, the Sharpe stays between −0.59 and −0.32. No single outlier explains it.
Universe size. With the top 30 instead of the top 50, the Sharpe is 0.16; with the top 80, it is 0.38. Either way, it stays far below the hindsight list.
Crashes. The point-in-time portfolio was hit harder in four of five stress episodes:
| Episode | Hindsight list | Point-in-time |
|---|---|---|
| Luna, May 2022 | 2.8% | 15.4% |
| FTX, Nov 2022 | 2.7% | 7.7% |
| SVB, Mar 2023 | 1.5% | 1.8% |
| Yen carry unwind, Aug 2024 | 0.8% | 2.9% |
| Flash crash, Oct 2025 | 5.8% | 3.1% |
Maximum drawdown within each episode window.
In the Luna crash, LUNA itself sat in the long leg: before the collapse, the crowd was short and funding was deeply negative. The rule bought exactly that.
Where did two Sharpe points go?
The obvious suspect is the dead coins. The breakdown says otherwise:
Even without a single delisted coin, the strategy collapses (−0.38). Delistings in the narrow sense explain only a small part. The real effect is the list itself: having volume today means having won the cycle.
A follow-up (chosen after seeing this result, so it only helps interpretation) limited the point-in-time universe to coins listed before October 2021 and landed at a Sharpe of 0.28. Roughly: coin age adds about 0.7 Sharpe, excluding delisted coins (almost entirely LUNA) about 0.3, and selecting survivors by today's volume about 1.0. The largest item is pure hindsight. That split is rough, though: the parts overlap and don't add up cleanly, and the 0.28 itself depends heavily on universe size (top 30: 0.10, top 80: 1.18). Only the ranking holds, not the individual numbers.
What actually lands in the long leg
The mechanism becomes visible when you look at which coins the rule buys:
| Long leg, average per week | Hindsight list | Point-in-time |
|---|---|---|
| Coins from outside the hindsight list | 0% | 66% |
| Price change over the previous 30 days | +2.8% | +27.5% |
| Weeks in which the leg had risen over those 30 days | 45% | 72% |
| Price change in the following week | +0.05% | −1.65% |
| Negative funding, as a multiple of the list's level | 1× | 4.6× |
| Time since listing (lower bound) | 3.8 years | 2.4 years |
Price only, before funding and costs; equal-weighted. Coins listed before 2020 count with their age since January 2020, so the age row is a lower bound.
- On the hindsight list, the coins with the most negative funding are mostly established projects that too many traders are temporarily betting against. They had moved sideways before, and on average they held their price afterwards. With the funding income on top, the strategy earns.
- In the point-in-time universe, two thirds of the long leg are coins that never made the list: younger coins right after sharp rallies, with funding several times more negative. In the following week, they fell on average. Negative funding there doesn't mean "crowded short"; it means "justified short".
The same signal measures two different things in two universes. The list had already filtered out the dangerous cases.
A side note on honesty: an earlier case study of ours concluded that funding had been "unremarkable" before the Luna collapse. It had checked the wrong leg. In the point-in-time test, LUNA sits in the long leg.
What remains: the short leg was positive in both universes. Taken on its own as an unhedged directional short, exposed to price moves as well as funding, a later test could not distinguish it from a random basket of altcoin shorts. That is not a tradable strategy.
What we take away
- A coin list is a parameter. Any list built with today's knowledge (trading today, liquid today, top 50 today) contains look-ahead, even if no line of code ever looks into the future.
- Adding delisted coins is not enough. Most of the error comes from how the survivors were selected, not from the missing dead coins.
- Point-in-time is the only honest basis. On every day, the universe must consist of data that was available on that day.
- Robustness checks on the wrong universe are worthless. The original backtest passed halves, cost stress and crash replays, all on the same hindsight list.
- Write your expectation down first. Otherwise we would later have considered "the bias is just a haircut" to be obvious.
What this means when you evaluate any bot
If someone shows you a backtest on "the top 50 coins", ask: top 50 as of when? Are delisted coins in the data? Was the universe chosen before or after the results were known? A backtest that can't answer these questions tells you more about the coin list than about the strategy.
This is also why FundingArb doesn't sell on backtest returns: the live dashboard and the funding payments in your own exchange history are things you can check yourself.
Methodology
- Data
- Binance USDⓈ-M perpetuals: daily candles, quote volume and funding; 895 symbols including delisted ones.
- Period
- 2022-01-15 to 2026-08-31 (tested window).
- Universe
- Weekly top 50 by trailing 30-day quote volume; minimum age 220 days; sufficient recent funding history.
- Portfolio
- Weekly rebalancing; quintiles (at least 3 coins per side); 50% gross per side; volatility-targeted.
- Costs
- Per side and per change in position, plus daily funding payments.
- Placebo
- 10,000 permutations of the signal values within the universe on each rebalancing day.
- Multiple testing
- Threshold 0.05 divided by the number of logged tests run so far (41 at the time). About 120 earlier exploratory hypotheses are not in this count; including them makes the threshold stricter and changes no verdict here.
- Known limits
- Binance data only; execution on other exchanges is not modelled, and many point-in-time coins would not have been tradable there. Price artefacts on delisting days are deliberately included.