The crypto liquidity revolution: BTC spreads compressed 99% in six years

1. The dataset — what 13TB of order books actually contains

We pulled L2 order-book snapshots from a historical crypto-market archive — every venue's full bid/ask ladder, sampled monthly at first-of-day across 4 venues × 6 USDT pairs. Coverage is uneven: Binance and OkEX dominate the panel; Bybit only entered with meaningful breadth in 2021. The asymmetry matters — any cross-venue ranking has to weight observation counts; within-venue time-series comparisons are clean.

Figure 1. Venue coverage by year. Only 2021 has all four exchanges trading all six pairs simultaneously — the only "fair" cross-venue comparison year. Pre-2021 rankings exclude Bybit; 2022-2023 exclude Bybit again because Bybit's data feed exited the panel after their 2021 BTC spike. We report headline numbers on the full panel and provide the 2021-only restricted estimates as the robustness anchor.

The full panel is 3,082 venue-day observations. Spread distribution: median 2.07 bps, mean 5.89 bps, max 312 bps — a log-normal right tail that the next section dissects.

2. Spread compression — a 99% collapse

BTC/USDT on Binance traded at 21.77 bps average spread in 2017. By 2023 that number was 0.19 bps — a 99.0% reduction. ETH/USDT compressed from 32.42 to 0.06 bps, a 99.8% reduction. This is the empirical signature of professional market-making arriving: the Jane Street / Jump / Wintermute build-out of crypto desks tracks the same calendar.

Figure 2. 60-day rolling spread on Binance for the six pairs (log scale). BTC and ETH cross the 1 bp line in mid-2020 and stay there. Alt-coins (SOL, ADA, DOGE, XRP) settle in the 2–5 bp range — tight by any standard but reflecting realised volatility differences (see Section 5).

Figure 3. Spread compression sorted by magnitude. All six pairs lost between 78% and 100% of their initial spread. ETH at 0.06 bps in 2023 is tighter than SPY at most retail venues — a result that surprises every traditional-finance reader.

3. Depth explosion — capital arrived alongside tight spreads

Spread compression with shallow books is brittle. The 2017→2023 series shows the opposite. ETH/USDT on Binance grew from $51,747 (2017) to $269,857 (2023) of resting depth in the top-10 levels — a 4.72× expansion. The $10K-trade impact for ETH fell from 42.72 bps to 0.07 bps — a 610× reduction in execution cost.

Figure 4. Top-10 USD notional depth, 90-day rolling on Binance. SOL emerges as the deepest book by 2023 ($1.56M average) — post-FTX rotation into Solana made it the “high-beta L1 trade” and market makers responded by committing serious capital.

Figure 5. Execution cost over time. Left: $10K market order impact, 60-day rolling. Right: $100K market order impact (10× larger). The $100K curve sits only 0.03 bps above the $10K curve for BTC by 2023 — the books are so deep that 10× the size barely moves the price.

4. Exchange head-to-head — one BTC, four liquidity regimes

Across 245 metric-venue-pair rankings on the full panel, Binance takes #1 in 100 of them (40.8%). But this number is contaminated by coverage asymmetry — Bybit only enters in 2021, so pre-2021 rankings exclude it by construction.

Figure 6. BTC/USDT across the four venues. Top: spread (log scale). Bottom: top-10 depth (log scale). Bybit's 2021 BTC book is the outlier — deeper than Binance for a brief window. By 2023, Binance reasserts dominance on depth while spread converges across venues.

What this tells us. OkEX leads on quote frequency (re-quotes more aggressively, shallower books). Bybit's 25 #1 rankings are almost entirely from the 2021 BTC perp window — real but unrepresentative. The post-2022 picture bifurcates: Binance captures 93–96% of cross-venue depth for the long-tail pairs (ADA, SOL, XRP), but BTC and ETH remain competitively fragmented because the absolute market is large enough to support multiple deep books.

5. The volatility-spread nexus — what causes what

Spreads are not constant — they widen with realised volatility. On the raw daily panel BTC spread-vol Pearson r = +0.36 (p < 0.001). But raw-level correlation is biased by shared downward trend; we report both raw and log-difference correlations to separate the two effects.

Figure 7. Spread vs realised volatility on log-log axes (Binance, daily). The fitted power-law slopes are ~3.0 bps/vol-unit for BTC and ~2.0 for ETH — every additional unit of annualised volatility costs traders 2–3 bps of spread. SOL's slope of ~15 is the steepest in the panel, the classic illiquidity penalty for lower-cap assets.

Figure 8. Stress days flagged on the BTC time series (top 5% by spread, threshold ~7.7 bps). Point estimate: stress days carry 1.88× normal volatility; block-bootstrap 95% CI [1.57, 2.35] (B=5,000, block size = 5 to handle serial correlation). Welch t = +4.89, p = 2.1e-06. The relationship is statistically robust after correcting for both unequal variance (Welch) and autocorrelation (block bootstrap).

6. Cross-asset correlation — trend bias vs real co-movement

The six Binance pairs co-move at the spread level. Mean off-diagonal correlation: 0.578 on raw levels, 0.512 after log-difference. BTC-ETH spread correlation is the standout — 0.978 on levels, 0.704 after detrending. The detrended number is what matters — it isolates day-to-day co-variation from shared 6-year compression trend.

Figure 9. Cross-asset spread correlation matrices. Left: raw level correlation (trend-biased — both spreads fell over time, inflating apparent co-movement). Right: log-difference correlation, the right answer. BTC-ETH at 0.70 is the still-strong signature of a shared global market-maker cohort quoting both books. SOL decouples on raw (−0.00) but couples back on log-diff (0.49) — the level-correlation hid the daily mechanics.

7. Market maturation — book symmetry and quote intensity

Two orthogonal indicators of market maturity. Depth imbalance measures bid-side vs ask-side asymmetry — mature books are symmetric; immature books have one-sided pressure. Quote frequency measures how often the L1 quote updates — mature books saturate at a venue-dependent ceiling.

Figure 10. Depth imbalance (bid_depth − ask_depth) / total, 90-day rolling, Binance. Most pairs converge toward zero (symmetric) by 2021–2022. DOGE remains visibly bid-heavy through the meme-coin era — the retail buy-the-meme flow persists in the order book.

Figure 11. Quote frequency (L1 updates per second in the first-day snapshot), 90-day rolling. Saturation at ~0.033 reflects the sampling convention, not the true tick rate — underlying tick streams sustain thousands of updates per second. The flat ceiling tells us the L1 quote is being maintained continuously by 2021+.

8. What this means for execution today

The punchline: crypto execution is, in 2023, a solved problem for major pairs. Residual alpha is in when to execute, not where. For tail assets it's still a venue-selection problem with a clear answer (Binance) and a measurable cost (2–5 bps + impact). The whole hand-wringing literature about “crypto being too illiquid for institutions” died in this dataset somewhere around 2021.

9. Methodology & limitations

• Source: historical crypto-market L2 archive, 10-level snapshots. Raw tick data, reduced to daily / monthly / yearly aggregates on Cornell research3 (112 cores, 7 days of compute by the upstream pipeline).
• Sampling: first trading day of each month, full-day L2 stream. Not the full 24×7 panel — a stratified monthly sample. Annual / yearly statistics are means across all available snapshots.
• Spread metric: time-weighted average of best-bid/best-ask spread in basis points.
• Depth: sum of resting USD-notional across the top 10 ladder levels, both sides.
• VWAP impact: counterfactual sweep cost of a $10K (and $100K) market order against the snapshot ladder, expressed in bps from mid.
• Volatility: annualised standard deviation of 1-second mid-price log returns within the trading day, scaled by √(365·86400).
• Statistical tests: Welch's t-test for stress vs normal comparison (allowing unequal variance); moving-block bootstrap (B=5,000, block size=5) for ratio confidence intervals to handle serial correlation in daily spreads. Cross-asset correlations reported on both raw levels and log-differences to separate trend co-movement from real co-variation.
• Headline figures re-verified: All headline numbers in this digest were re-computed from the 3,082-row daily panel by an independent script and cross-checked against the upstream report. Deltas from the prior version: ETH depth growth updated 5.2× → 4.72× (year-mean rather than peak/trough); ranking interpretation corrected to flag coverage bias.