ASF

Resolution-Dependent Direction Reversal in BTC/ETH Lead-Lag on Binance Spot

March 9, 2026|
A
A

At lags of 10 ms and below, ETH leads BTC on Binance spot. At 100 ms resolution the direction inverts and BTC leads ETH. Both effects are statistically unambiguous across 500,000+ windows and hold in January 2025 and the full calendar year alike.

The conventional claim in crypto market microstructure is the opposite: BTC leads ETH. At multi-second resolution this is broadly true but the direction reverses as the clock ticks finer. This post establishes that reversal across three datasets: January 2025 (31 days, analysed at four resolutions), the complete calendar year 2025 at , and the complete calendar year 2025 at .

1. Mathematical Preliminaries

Let and let denote the -th non-overlapping 60-second window of Binance spot trade data. Within each window construct the per-millisecond signed-volume process

where is trade size in base-asset units and with . The discrete cross-correlation between BTC and ETH in window is

Notation. denotes the bin width: the resolution at which is discretised before the argmax is taken. The peak-lag estimator for window is

rounded to the nearest multiple of . A window is classified as BTC-leads if , ETH-leads if , and no-lag if .

Among the directional windows, the BTC-leads fraction is

To test whether any observed directional bias is statistically significant, we use the large-sample two-sided proportion z-test

To quantify tail asymmetry at lag magnitude , define the pointwise asymmetry ratio

means BTC-leads windows outnumber their ETH-leads mirror images at that lag; means the opposite.

2. January 2025

2.1. Data and Method

The January 2025 dataset covers 30 trading days on Binance spot (the file for 19 January is empty), processed at four bin resolutions . Each day contributes 1,440 non-overlapping 60-second windows, for a total of qualifying windows at (8,640 at ). The search radius narrows as the bin width shrinks. The January runs left it unrecorded. The histogram has mass out to and the run has quartiles at .

Per-window cross-correlations are computed via the FFT using the Cooley-Tukey radix-2 algorithm as implemented in rustfft. At samples the complexity per window is and all 31 days are processed in parallel across CPU cores via rayon.

2.2. Multi-Resolution Histograms

Lag histograms at 100ms, 10ms, and 1ms bin resolution
Lead-lag histograms at three resolutions. Orange: BTC leads. Blue: ETH leads. Grey: no lag detected. The dominant lag sharpens from a wide zero spike at to a clear mode as .

At the distribution collapses almost entirely to : most lags are sub-100 ms. Among directional windows BTC leads 29.4% and ETH leads 27.4% of total windows with no lag in 43.3%. The run reported these fractions rounded to 0.1%. The counts below are reconstructed from them ( and , truncated). The BTC-leads fraction is

The rounding of the two fractions moves within . The signal is statistically present at 100 ms but obscured by the zero spike. At a secondary bump emerges at . That bin has 3,091 windows against 1,816 in the bin: 1.7 times as many. At the distribution resolves into a left-skewed unimodal shape with mode .

1ms resolution lag histogram close-up over ±20ms
Close-up at resolution over . The mode is (ETH leads BTC). The sample mean is , pulled rightward by a fat BTC-leads tail from windows in which BTC leads strongly.

The interquartile range at 1 ms resolution is . The per-window lag estimate is noisy.

2.3. Direction Breakdown by Resolution

Direction breakdown by bin resolution
Fraction of windows in each direction by bin resolution. The ETH-leads fraction sits near 27–29% at and and rises to 34.2% at and 52.7% at .

The ETH-leads fraction is flat between 1,000 ms and 100 ms and rises steeply below 100 ms:

BTC leadsETH leadsNo lag
1000 ms30.2%29.3%40.5%
100 ms29.4%27.4%43.3%
10 ms31.8%34.2%34.1%
1 ms39.9%52.7%7.4%

The direction of dominance reverses as resolution improves. At BTC leads marginally with a p-value of . At BTC is ahead by 0.9 points on 8,640 windows (, ). That gap falls inside the noise. At millisecond resolution ETH leads with a 12.8 percentage-point gap. Section 4.4 extends this comparison to the full year and adds the full-year row for a direct cross-dataset test of the reversal.

2.4. Interpretation

Millisecond-scale ETH lead. On Binance spot in January 2025 ETH leads BTC by approximately 1 ms. This inverts the conventional multi-second narrative. One candidate mechanism (untested here) is that ETH's lower per-trade notional value enables faster order-book clearing. ETH is then the preferred signalling leg for cross-asset arbitrageurs who exploit the BTC/ETH cointegration relationship.

The resolution-dependent reversal is informative: at coarse timescales BTC-driven macro moves (larger in notional magnitude) dominate the cross-correlation sign while the ETH-leads signal at fine scales is consistent with high-frequency arbitrage. The aggregate bias of 12.8 percentage points across 43,200 windows is robust but too small to exploit reliably on a per-window basis.

3. Full Year 2025 at 100 ms Resolution

3.1. Data and Scale

Extending the analysis to the complete calendar year 2025 yields qualifying 60-second windows (99.5% data completeness over 365 days). This is more than January alone. All full-year analysis in this section uses a single resolution with .

The larger sample reduces the standard error of :

compared to . The full year therefore detects directional biases approximately four times smaller in absolute terms.

3.2. Distributional Structure

Full-year 2025 lag histogram at 100ms resolution
Lead-lag histogram at , full year 2025 ( windows). Dashed: sample mean (). Dotted: median (). The distribution is a zero spike at 35.9% flanked by approximately exponential tails.

The full-year distribution at is a mixture of a point mass and a continuous part. A spike at contains 35.9% of all windows, surrounded by tails that decay approximately exponentially. The zero spike is smaller than in January (43.3%). The full year contains proportionally more windows with detectable directional structure.

StatisticJanuary 2025Full Year 2025
Total windows43,200522,719
No-lag fraction43.3%35.9%
BTC-leads fraction29.4%34.2%
ETH-leads fraction27.4%29.9%
Sample mean+19.6 ms+25.9 ms
Median0 ms0 ms
−100 ms−100 ms
+100 ms+200 ms
Normalized histograms: January 2025 vs Full Year 2025
Normalised lead-lag histograms at : January 2025 (left) and Full Year 2025 (right). Both panels share the same -axis. Dashed lines mark sample means. The zero spike shrinks and the BTC-leads tails fatten outside the January bull-run period.

The side-by-side comparison reveals two structural shifts. First the no-lag fraction decreases from 43.3% to 35.9%. The full year contains proportionally more windows with a directional signal. Second the BTC-leads tails are heavier in the full year. shifts from to and the sample mean from to .

3.3. Tail Asymmetry

Classical skewness is uninformative about directional asymmetry because the zero spike dominates first- and second-order moments. The pointwise asymmetry ratio from [eq:ratio] directly measures the imbalance at each lag magnitude without conflation from the zero spike.

Asymmetry ratio R(tau) for full year and January 2025
Pointwise asymmetry ratio for . Full year 2025 (orange circles) and January 2025 (blue squares). marks the symmetric baseline. The full-year ratio exceeds 1 at every lag magnitude and peaks at .

For the full year at every tested lag magnitude:

The ratio peaks near . This suggests that moderately delayed BTC-leads episodes (those reflecting large cross-asset price moves) drive the directional asymmetry. January 2025 also has at every lag from 100 to 1,000 ms, with values between 1.00 and 1.21. The January asymmetry is weak: a chi-squared symmetry test over the ten lag magnitudes gives (, ) against 1,646 for the full year.

3.4. Empirical CDF and Stochastic Dominance

Empirical CDFs of peak lags, in-range windows only
Empirical CDFs of peak lags restricted to the in-range portion. The full-year CDF (orange) lies to the right of the January CDF (blue) for , consistent with first-order stochastic dominance of BTC-leads outcomes in the full year.

Let and denote the empirical CDFs of in-range peak lags for the full year and January, respectively. The observed ordering

constitutes approximate first-order stochastic dominance: conditional on a window having the full-year distribution assigns more probability mass to larger BTC-leads lags than January does. The third quartile shifts from (January) to (full year). This quantifies the extent of this dominance.

3.5. Directional Dominance Test

We test whether BTC-leads and ETH-leads windows occur with equal probability among directional windows:

With and for the full year the directional count is and the z-statistic is

Dataset-value
January 20250.517524,5395.52
Full Year 20250.5330335,01238.17
z-test visualisation: point estimates and test statistics
Left: point estimates with 95% confidence intervals. The full-year CI is , invisible at this scale. Right: -statistics on the standard normal density. Both statistics fall far beyond the visible axis range and reject .

The full-year z-statistic of 38.17 rejects at any conventional significance level. Both January and the full year show BTC dominance at . This is consistent with BTC as the primary locus of price discovery at multi-millisecond timescales and ETH adjusting after a delay from arbitrage execution latency.

4. Full Year 2025 at 1 ms Resolution

4.1. Data and Method

The 1 ms full-year analysis uses the same 522,719 qualifying windows as the 100 ms run but with and . At this resolution the FFT length expands to 131,072 bins per window and the entire year is processed in approximately 16 minutes on 32 CPU cores (I/O across approx. 1.3 TB of Parquet data is the binding constraint).

Across 522,719 windows the directional counts are:

The no-lag fraction at 1 ms is 2.4% against 35.9% at 100 ms. At this resolution virtually every 60-second window has a detectable cross-correlation peak.

4.2. Distributional Structure

Full-year 2025 lag histogram at 1ms resolution, ±10ms shown
Lead-lag histogram at , full year 2025. Only the central is shown (18.1% of all windows); the full distribution extends to . Mode: (ETH leads). The sample mean of is pulled rightward by heavy BTC-leads tails beyond .

The mode at matches January 2025 at the same resolution. The millisecond-scale ETH-leads signal extends beyond the January bull run. The directional z-test gives

ETH leads BTC over the full year at 1 ms resolution. The effect size (Cohen's ) is substantially diluted relative to January (), where the BTC bull-run concentrated a strong ETH-leads regime.

The mean of appears inconsistent with the mode of . The explanation lies in the long positive tail of the full distribution: among windows outside the shown range (81.9% of all windows), BTC-leads windows at large positive lags outnumber ETH-leads windows at large negative lags. This pulls the mean rightward.

4.3. January vs Full Year at 1 ms

January 2025 vs Full Year 2025 at 1ms resolution, side-by-side
Normalised lead-lag histograms at : January 2025 (left) and Full Year 2025 (right). ETH leads in both panels. The January distribution is more left-skewed: the ETH-leads regime was stronger during the January bull run. Cohen's : January , Full Year .

Both datasets share the same mode but the magnitude of the ETH-leads bias differs substantially. In January the BTC-leads fraction among directional windows is 43.1% (), versus 49.0% over the full year (). The effect size dilutes by a factor of approximately 7 as the dataset expands from 31 days to 365 days:

DatasetCohen's
January 2025, 0.430940,003−27.65−0.139
Full Year 2025, 0.4904509,963−13.73−0.019

The January ETH-leads signal () is seven times the full-year one but still below Cohen's 0.2 threshold for a small effect. The full-year signal is negligible in absolute magnitude and statistically significant owing to the large sample.

4.4. Scale-Dependent Direction Reversal

Direction fractions at 1ms vs 100ms for January and Full Year 2025
Direction fractions at each resolution, for January 2025 (top) and Full Year 2025 (bottom). At , ETH leads in both datasets. At , BTC leads in both. The direction of dominance reverses between these two resolutions.

The analysis finds a resolution-dependent direction reversal:

ResolutionDatasetBTC leadsETH leadsNo lagDominant
January 202539.9%52.7%7.4%ETH−27.65
Full Year 202547.8%49.7%2.4%ETH−13.73
January 202529.4%27.4%43.3%BTC+5.52
Full Year 202534.2%29.9%35.9%BTC+38.17

The sign of flips from negative (ETH leads) to positive (BTC leads) as increases from 1 ms to 100 ms, in both January and the full year independently. For January the 10 ms run already favours ETH (, ). The crossover therefore lies between 10 and 100 ms.

4.5. Pointwise Asymmetry at 1 ms

Asymmetry ratio R(tau) at 1ms resolution for both datasets
Pointwise asymmetry ratio at , for . Full Year 2025 (orange circles) and January 2025 (blue dashed squares). Both curves lie strictly below : ETH-leads windows dominate at every lag magnitude in the shown range.

At 1 ms resolution for every lag magnitude in both datasets. This is the opposite of the pattern seen at 100 ms. The minimum asymmetry ratio for the full year occurs at :

This confirms that at sub-10-millisecond timescales ETH-leads windows far outnumber BTC-leads windows at the same lag magnitude.

A joint chi-squared symmetry test of across all 10 lag magnitudes rejects symmetry in both datasets:

The full-year chi-squared is approximately 4.6× larger than January's despite the 7× dilution in Cohen's because the chi-squared statistic grows with . Both reject distributional symmetry at 1 ms.

4.6. Significance Across All Resolution and Dataset Pairs

Forest plot: z-statistics and Cohen's h effect sizes for all 6 resolution/dataset combinations
Left: with 95% CI for all six resolution/dataset pairs. Right: Cohen's effect sizes. Blue bars: ETH leads (). Orange bars: BTC leads (). The sign of flips with resolution, and the largest effect sizes occur in January at (ETH, ) and full year at (BTC, ).

Across all tested combinations the sign of both and Cohen's depends on alone. At ETH leads. At BTC leads. At (January only) is positive but the test does not reject (). The January 1 ms scenario has the largest absolute effect size (), consistent with a concentrated ETH-leads regime during the BTC bull run.

4.7. Distributional Distance between January and Full Year at 1 ms

Empirical CDFs at 1ms: January 2025 vs Full Year 2025, with KS test
Empirical CDFs of peak lags (±10ms range shown) at : Full Year 2025 (orange) and January 2025 (blue dashed). The full-year CDF lies to the right of January for . The full year has heavier BTC-leads mass. KS test (inset): , .

A two-sample Kolmogorov–Smirnov test on the central histogram bins rejects the null hypothesis that January and full-year 1 ms lag distributions are drawn from the same population:

The January distribution is shifted further into the ETH-leads (negative) side relative to the full year, consistent with the stronger bull-run ETH-leads bias in January. The full-year distribution has heavier BTC-leads mass at . This pulls the cumulative distribution rightward. For comparison, the same KS test at 100 ms gives , , with the full year shifted rightward (BTC-leads direction) relative to January in that regime as well.

4.8. ETH-leads Advantage vs Lag Threshold

Cumulative ETH-leads fraction as a function of max lag threshold
ETH-leads fraction among directional windows with detected lag , as grows from 1 ms to 2000 ms (log scale). Solid: Full Year 2025. Dashed: January 2025. Solid and dashed markers are exact for . Triangles are aggregates from other runs: the ETH-leads fraction of the whole -resolution run (plotted at 100 ms) and of the whole -resolution run (plotted at 2,000 ms). The dotted segments only join points from different runs.

The solid curves are exact. At ETH-leads windows account for 72.7% of directional windows in the full year and 70.7% in January. At the fractions are 65.4% and 67.2%. The curves stop at 10 ms because the 1 ms histograms were recorded only to .

The full-year 1 ms run still separates the two regimes within a single estimator. Its totals give 259,883 ETH-leads and 250,080 BTC-leads windows, of which 53,530 and 28,382 lie within . Among the directional windows with , ETH leads in only 48.2% (206,353 against 221,698). At 1 ms resolution ETH therefore leads at small lag magnitudes and BTC leads at large ones.

The triangles place the 100 ms-resolution run at 46.7% ETH (full year) and 48.2% (January). The crossover in lag magnitude therefore lies between 10 and 100 ms. A straight line through the dotted segment crosses 50% near 15–20 ms. That segment joins two different runs and does not narrow the interval.

4.9. Interpretation

The full-year 1 ms analysis has three findings:

  1. Direction reversal between 10 and 100 ms. Below 10 ms the high-frequency ETH-leads channel dominates. At 100 ms the BTC macro price-discovery channel has taken over. Within the full-year 1 ms run ETH leads in 65.4% of windows with and in 48.2% of those beyond.

  2. ETH lead at 1 ms resolution. The mode is in both datasets. The directional bias is statistically significant in both ( in January, in the full year). The effect size dilutes substantially outside the January bull run (Cohen's : vs ).

  3. Difference between the full-year and January 1 ms distributions. The KS statistic (, ) confirms that the January bull run produced a distinctly left-skewed 1 ms distribution. The full year is more symmetric in the central ±10 ms region, consistent with the ETH-leads regime being diluted by non-trending market conditions.

5. Conclusion

At coarse timescales (100 ms and beyond), BTC leads ETH, consistently across January 2025 alone and across the full calendar year. The full-year signal is strong: , and an asymmetry ratio peaking at .

At millisecond resolution the direction reverses. Both January and the full year show ETH leading BTC with mode . The effect is larger in January (Cohen's , below the 0.2 threshold for a small effect) than over the full year (, negligible in magnitude but statistically unambiguous given ).

The crossover lies between 10 and 100 ms. Every exact cumulative point up to 10 ms has , the 100 ms run has in both datasets, and the full-year 1 ms run has among windows with lags beyond 10 ms. Over all lags to 2,000 ms the full-year 1 ms fraction is 50.96% ETH. The HFT arbitrage channel (ETH leads, 10 ms and below) and the macro price-discovery channel (BTC leads, 100 ms+) are statistically separable and operate in opposite directions.

is a function of , the market regime and the prevailing level of cross-asset arbitrage activity. Treating any single lag estimate as a universal constant discards the most informative feature of the data: its resolution dependence.