Does the ETH-BTC Funding Differential Actually Predict Anything? A Backtest, Adversarially Reviewed

1. The Question

A prior piece in this series (ETH/BTC relative-value trade note) used the fact that ETH's perpetual funding rate was running mildly above BTC's as one leg of support for a long ETH/BTC position: the logic being that a real, non-crowded funding premium for ETH was consistent with a genuine, if quiet, rotation into ETH.

That's a plausible-sounding argument. It's also exactly the kind of claim that should be tested rather than asserted. So: does an elevated ETH-minus-BTC funding differential historically predict higher forward ETH/BTC returns, as the trade note's logic implies, or does it predict something else?

2. Methodology and Real Results

Data: 500 aligned 8-hourly observations of BTCUSDT and ETHUSDT perpetual funding rates (from Binance futures) and spot close prices (from Binance spot, confirmed below), spanning 2026-03-05 to 2026-08-18: about 167 continuous days, one regime, no gaps.

Signal: funding_diff(t) = ETH funding rate(t) − BTC funding rate(t), observed at each 8h settlement: public information at time t, no look-ahead in the signal itself.

Target: forward ETH/BTC price-ratio return from t to t+horizon, tested at three horizons: 1-day, 3-day, and 7-day.

Results, correlation and both naive-OLS and Newey-West (HAC) corrected t-stats:

Horizon	n	Correlation	Naive OLS t-stat	Newey-West t-stat
1-day	497	−0.119	−2.66	−2.30
3-day	491	−0.135	−3.00	−2.21
7-day	479	−0.223	−4.99	−2.08

The sign is consistently negative across all three horizons: the opposite of the trade note's momentum-confirmation logic. A higher ETH funding premium over BTC predicted lower, not higher, forward ETH/BTC returns in this sample. A naive median-split backtest (long when funding_diff is above the sample median, flat otherwise, no costs) showed hit rates of 40.6%–46.0% and small negative mean returns when the signal fired at every horizon.

For context: the sample's mean funding differential annualizes to about −0.69% (this wasn't even an ETH-bid regime on average), with an annualized standard deviation of about 4.9%. The most recently settled print annualizes to about +0.87%, sitting at roughly the 64th percentile of the sample's own distribution: meaningfully less extreme than the ~+4% annualized figure quoted in the trade note, which came from live estimated ("next funding") rates rather than the last settled print. Both are legitimate numbers; they're just measuring different instants, since estimated funding accrues continuously toward the next settlement.

3. What the Adversarial Review Found

Both reviewers were told explicitly to find flaws, not validate the finding, and both did real, quantitative critique rather than generic hedging.

Where they agreed, and it mattered:

• The overlapping-window problem is real and the Newey-West fix likely still understates it. With 8h steps and a 7-day (21-period) horizon, the forward-return windows overlap heavily: the residual is mechanically autocorrelated up to lag 20. Both reviewers flagged that a 20-lag Bartlett kernel only fully corrects this if the underlying 8h returns are themselves serially uncorrelated, which is an assumption, not a fact, given known volatility clustering in crypto. Kimi went further and named the specific literature this resembles: the Stambaugh/Valkanov long-horizon-regression problem, where a persistent predictor (funding is sticky, with plausible lag-1 autocorrelation above 0.8) combined with an overlapping, correlated dependent variable produces t-stats that are upward-biased in magnitude even after a standard HAC correction, because the asymptotic theory the correction relies on doesn't hold well with this few effective observations.
• Multiple-horizon testing without correction inflates apparent significance. Three horizons were tested; none of the individual t-stats survives even a conservative Bonferroni adjustment at the 5% level. DeepSeek noted something sharper: because the horizons are nested (the 7-day return literally contains the 1-day and 3-day returns as sub-components), the fact that the sign is consistent across all three "is not three findings; it's one marginal finding restated three times." Treating horizon-consistency as corroborating evidence was a mistake in how the original results were framed.
• The threshold backtest is a diagnostic toy, not a strategy. No transaction costs, no sizing, and, as both reviewers specifically flagged as more serious than it sounds, no funding carry in the P&L, despite funding being simultaneously the signal and the actual cost or revenue of holding the position. DeepSeek quantified it: at the 7-day horizon, the -21bp mean return when the signal fires is roughly one round-trip trading cost before even accounting for carry; folding in the estimated funding carry on the losing side pushes the loss to roughly -27bp. This is a small, fragile number being treated as if it were a clean result.
• Sample size is the core problem, and both reviewers quantified how far short it falls. Effective independent observations (not raw row count) are roughly n/horizon: about 165 at the 1-day horizon, about 55 at 3-day, and only about 23 at 7-day. DeepSeek's power calculation: detecting a true correlation of 0.22 (the 7-day point estimate) with 80% power at 5% significance requires roughly 150 independent weekly observations: about three years of data. Kimi's version, using Fisher-z confidence intervals, put a number on how uncertain the headline correlation really is: the 95% CI on the 7-day correlation is roughly [−0.59, +0.22]. It includes zero and includes values of either sign. With 23 independent weekly windows, this dataset is roughly an order of magnitude short of what would be needed to trust this specific correlation.

Where the reviewers added distinct value beyond each other:

• Kimi caught the sharpest potential design flaw: whether the "close prices" behind the ratio were spot prices or perpetual mark prices. If the ratio had been built on perp prices, a positive funding differential would mechanically correlate with the perp trading at a premium that later converges: an accounting identity, not a behavioral finding, of roughly the same order of magnitude as the estimated effect. This was worth checking directly rather than trusting either model's assumption. Confirmed against the actual script: the ratio was built on api.binance.com/api/v3/klines, the spot endpoint, not fapi.binance.com futures mark prices. The mechanical-convergence concern doesn't apply here, but it was the right thing to demand proof of rather than take on faith, and neither reviewer could have known the answer without it being checked.
• Kimi also flagged an unresolved confound neither the original design nor DeepSeek's review addressed: funding differential is itself partly a function of recent ETH/BTC outperformance (aggressive buying is what generates elevated perp premium in the first place). Without a control for lagged 7d/30d ratio returns, this backtest cannot cleanly distinguish "funding differential predicts reversal" from "short-term mean reversion exists and funding differential is a noisy proxy for the very recent-return that's reverting." That control was not run here and is a real gap, not a nitpick.
• DeepSeek did the more useful economic-magnitude translation: converting the raw regression beta into an actual expected move at the current funding reading. At the last settled differential (+0.87% annualized, about 1.56 points above the sample mean), the 7-day beta implies roughly a −0.20% expected ratio move before costs: a real number, but a small one, on a variable that explains roughly 5% of variance (R² ≈ 0.05 at the 7-day horizon). Even taken completely at face value, this was never a large effect.
• Kimi also raised, and this note cannot fully resolve, a legitimate researcher-degrees-of-freedom question: why does the sample start 2026-03-05 when Binance serves funding history back to 2019? If that start date wasn't chosen for a documented, exogenous reason (data availability at the time this piece was built, in this case), an arbitrary window start is itself a way results can be quietly cherry-picked, even unintentionally.

4. An Original Robustness Check: How Fragile Is the Sign, Really?

DeepSeek's review computed non-overlapping ("independent-block") re-estimates of the regression as a cross-check and reported t-stats of roughly −1.53, −0.98, and −1.02 at the three horizons, all well below conventional significance, reinforcing that the headline HAC-corrected t-stats overstate confidence. Reproducing that check directly from the same data did not replicate those exact numbers, which, rather than being a discrepancy to gloss over, turned out to be the most important finding in this entire piece.

The reason is that "independent blocks" from an 8h-step series isn't a single well-defined dataset: it depends on which of the horizon_periods possible starting offsets you choose to build non-overlapping windows from. Running the 7-day independent-block regression at every one of the 21 possible starting offsets gives:

• t-stats ranging from +0.48 to −2.93 depending purely on the arbitrary choice of where to start counting
• a mean t-stat across all 21 offsets of −1.14
• only 4 of 21 offsets reaching |t| ≥ 2
• 2 of 21 offsets flipping to a positive sign entirely

This is the concrete, numerical version of exactly what Kimi's review warned about without fully quantifying: with this few independent observations, the result is highly sensitive to arbitrary methodological choices, and "you don't know whether you have a coefficient or an anecdote" until you check. Now it's checked. The average outcome across every legitimate way of slicing this sample into independent blocks is a t-stat of about −1.14: nowhere near the −2.66 to −4.99 the naive overlapping-window regression originally reported, and not distinguishable from noise by any conventional standard.

5. Reconciling with the Trade Note

Kimi's review made a distinction worth adopting exactly as stated, because the trade note actually made two separable claims, and this backtest treats them very differently:

Claim A: "funding isn't crowded, so this isn't a squeeze setup." This is a statement about tail risk, and it is not contradicted by this backtest. The current settled reading (64th percentile) and even the higher estimated-rate reading (~83rd percentile) both sit well short of the sample's tail. Nothing in this analysis says anything about tail behavior: only about the conditional average, which is a different question. On this narrower claim, the trade note's framing holds up.

Claim B: "the differential is supportive, a genuine bid confirming continuation." This is a claim about the sign of expected forward returns, and here the backtest's point estimate runs the other way, though, per Section 4, not reliably enough to call it established. The honest position is: there was never real evidence for Claim B to begin with, and there still isn't strong evidence against a truly null relationship either. What's actually been learned is that the specific mechanism the trade note leaned on ("positive funding differential = supportive tailwind") should not be treated as a load-bearing part of that thesis. It wasn't well-supported before this backtest, and after a genuinely adversarial test, it still isn't.

6. Bottom Line

This is suggestive at most, not tradeable, and after independently verifying the numbers rather than taking either model's arithmetic on faith, "suggestive" is a generous read. The sign is consistently negative across three testing methods (naive OLS, Newey-West correction, and independent-block re-estimation), which is worth something: a pure noise process wouldn't reliably point one direction across that many angles of attack. But the magnitude of confidence any single number in this piece would suggest (a t-stat near −5, or even the corrected −2.1 to −2.3) substantially overstates what a properly de-overlapped test actually supports, which is closer to a mean t-stat of −1.1 with enormous variance depending on arbitrary choices.

What would be needed before this graduates from "interesting direction" to "tradeable signal," combining both reviewers' lists:

1. Multi-year history (3-7 years, ideally back to perp inception): this sample is roughly an order of magnitude too short at the horizon where the effect looks strongest.
2. Replication on at least one other venue (funding is venue-specific) and at least one other asset pair (SOL/BTC, or a cross-sectional panel): if this is a real crowding-unwind mechanism, it shouldn't be unique to one pair on one exchange.
3. A regression that controls for lagged ETH/BTC returns, to separate "funding differential predicts reversal" from "this is short-term mean reversion wearing a funding-rate costume."
4. A real backtest: realistic costs, funding carry included in P&L (it cuts both ways depending on which side of the trade you're on), walk-forward rather than full-sample thresholds.

None of that was done here, on purpose: the point of this piece was not to manufacture a clean-looking signal, but to show what happens when a plausible-sounding discretionary argument actually gets tested, reviewed adversarially by more than one independent reviewer, and then checked by hand rather than trusted. The honest result is: an interesting, economically sensible direction (crowded positioning tends to unwind) sitting on a sample too small and too fragile to act on, and a trade thesis that loses one of its supporting legs without losing its core risk-management logic.